Handheld NIR Spectroscopy for Mango Maturity Testing: A Comprehensive Guide for Researchers and Scientists

Abstract

This article provides a systematic review of handheld Near-Infrared (NIR) spectroscopy for non-destructive mango maturity assessment. It explores the foundational principles of how NIR light interacts with mango constituents like sugars, acids, and dry matter. The methodological section details hardware configurations, from commercial devices like the F-750 and NeoSpectra to custom prototypes using Raspberry Pi, and examines key data preprocessing and machine learning models, including PLSR, SVM, and novel direct classification approaches. The guide addresses critical troubleshooting and optimization challenges, such as selecting preprocessing techniques and managing model robustness. Finally, it presents a comparative validation of different methodologies, highlighting performance metrics and the superior accuracy of direct classification and hybrid models like LDA-SVM and fuzzy logic, which have achieved up to 97.44% and 95.7% accuracy, respectively. This resource is tailored for researchers, scientists, and professionals developing rapid, non-destructive quality control systems for fruit and pharmaceutical applications.

The Science Behind NIR: Understanding Light-Matter Interactions in Mango Maturity

Core Principles of Near-Infrared Spectroscopy and Molecular Bond Interactions

Near-Infrared (NIR) spectroscopy has emerged as a powerful, non-destructive analytical technique with significant applications in agricultural product quality assessment, particularly for determining mango maturity. This technology operates on the fundamental principle of molecular bond interactions with NIR light, enabling rapid, chemical-free analysis of fruit internal quality parameters. The core value of NIR spectroscopy lies in its ability to penetrate fruit tissue and provide quantitative data on critical maturity indicators without destroying the sample, making it ideal for supply chain quality control and optimal harvest timing decisions [1] [2]. For mango quality assessment, handheld NIR spectrometers have revolutionized in-field testing by bringing laboratory-grade analytical capabilities to orchards and packing houses, allowing growers to make data-driven decisions that maximize fruit quality and marketability.

Core Principles of NIR Spectroscopy

Molecular Bond Interactions and Spectral Absorption

The fundamental mechanism of NIR spectroscopy involves the interaction between NIR electromagnetic radiation (typically in the 780-2500 nm wavelength range) and molecular bonds in organic compounds [3]. When NIR light irradiates a material, chemical bonds undergo vibrational transitions that correspond to specific energy absorption patterns. The NIR region primarily captures overtone and combination vibrations of fundamental molecular bonds, including C-H, O-H, N-H, S-H, and C=O functional groups [1] [3].

These vibrational transitions occur because bonds behave like mechanical oscillators with characteristic resonant frequencies. When the frequency of incident NIR radiation matches the natural vibrational frequency of a molecular bond, energy is absorbed, creating detectable absorption patterns that serve as molecular fingerprints. The specific wavelengths at which absorption occurs provide qualitative information about chemical composition, while the intensity of absorption correlates with concentration, enabling quantitative analysis [4].

Table 1: Primary Molecular Bond Interactions in NIR Spectroscopy for Fruit Analysis

Molecular Bond	Wavelength Range (nm)	Quality Parameter	Vibration Type
O-H	1400-1450, 1900-1950	Water Content, Dry Matter	Combination, 1st Overtone
C-H	1100-1250, 1600-1800	Sugars, Soluble Solids	2nd & 3rd Overtones
C-H-O	2000-2200	Carbohydrates	Combination Bands
N-H	1500-1550, 1900-2000	Proteins	1st Overtone, Combination

Measurement Geometries in NIR Spectroscopy

The configuration of light source, sample, and detector—known as optical geometry—critically influences the type and depth of information obtained from fruit quality analysis. Three primary geometries are employed in NIR spectroscopy for mango testing:

• Reflectance Mode: The detector captures light reflected from the external or near-surface layers of the fruit, making it suitable for assessing external characteristics such as skin properties or superficial defects [1].
• Transmittance Mode: Light passes completely through the fruit and is detected on the opposite side, providing information about internal traits like sugar content, dry matter, or internal disorders [1].
• Interactance Mode: A compromise geometry that captures light that has entered the fruit and traveled through a portion of the tissue before exiting, offering a balance between surface and internal property assessment [1].

For handheld NIR devices used in mango maturity testing, interactance and reflectance modes are most commonly implemented due to their practical implementation advantages for whole fruit analysis.

NIR Spectroscopy Measurement Geometries

Application to Mango Maturity Testing

Critical Maturity Parameters in Mangoes

For mango quality assessment, NIR spectroscopy has proven particularly effective for measuring several key maturity indicators that correlate with eating quality and consumer acceptance. The most significant parameters include:

• Dry Matter Content (DMC): A reliable maturity index that increases gradually during fruit development and correlates strongly with final eating quality. DMC represents the non-water components of the fruit, including sugars, acids, and structural carbohydrates [2].
• Total Soluble Solids (TSS): Primarily representing sugar content, TSS increases sharply during the ripening phase as starch hydrolyzes to sugars. This parameter is critical for determining sweetness and flavor development [5] [6].
• Titratable Acidity (TA): The concentration of organic acids decreases during ripening, affecting the sugar-to-acid balance and overall flavor profile [6] [4].
• Firmness: While more challenging to predict via NIR spectroscopy, firmness decreases during ripening and can be correlated with spectral data in some varieties [2].

Research on Palmer mangoes has demonstrated that dry matter content serves as an excellent maturity index, with fruits reaching the industry standard of 150 g/kg at approximately 105 days after bloom, before the sharp rise in soluble solids content that occurs between 112-126 days after bloom [2].

NIR Spectral Ranges for Mango Quality Parameters

Different chemical components in mangoes absorb NIR radiation at characteristic wavelengths, enabling simultaneous quantification of multiple quality parameters.

Table 2: NIR Spectral Ranges for Key Mango Quality Parameters

Quality Parameter	Spectral Range (nm)	Molecular Basis	Prediction Performance (R²)
Dry Matter (DM)	1100-1300, 1400-1500	O-H, C-H bonds	0.84-0.87 [2]
Soluble Solids (TSS)	1100-1250, 1600-1750	C-H, O-H bonds	0.81-0.87 [5] [2]
Titratable Acidity (TA)	1400-1500, 1900-2000	O-H, C-O bonds	0.63-0.81 [5] [6]
Vitamin C	1400-1550, 1900-2100	C-H, C-O bonds	R² = 0.81 [4]

Experimental Protocols for Handheld NIR Spectroscopy in Mango Research

Protocol 1: Development of Maturity Prediction Models

This protocol outlines the comprehensive procedure for developing robust chemometric models for predicting mango maturity using handheld NIR devices.

Materials and Equipment:

• Handheld NIR spectrometer (e.g., Felix Instruments F-750/F-751)
• Reference laboratory equipment (refractometer, pH meter, oven for dry matter)
• Statistical software for chemometric analysis
• 150-200 mango samples representing different maturity stages

Procedure:

• Sample Selection and Preparation: Harvest mangoes at progressive maturity stages (e.g., 91, 98, 105, 112, 119, and 126 days after bloom for Palmer mangoes). Include fruits from different positions on trees and various orchards to capture natural variability [2].

• Spectral Data Acquisition:

  • Standardize measurement conditions (temperature, fruit orientation, light exposure)
  • Clean the fruit surface and position the handheld NIR spectrometer firmly against the fruit
  • Take multiple readings from opposite sides of each fruit and average the spectra
  • Use wavelength range of 740-1070 nm for standard handheld devices [5]

• Reference Measurements:

  • TSS: Extract juice from scanned areas and measure with a digital refractometer (°Brix)
  • Dry Matter: Collect flesh samples from scanned areas, weigh, dry at 70°C for 24-48 hours, and reweigh
  • Acidity: Titrate juice with 0.1N NaOH to pH 8.1, calculate as percent malic acid [6]

• Data Preprocessing:

  • Apply Standard Normal Variate (SNV) to minimize scattering effects
  • Use Savitzky-Golay derivatives (1st or 2nd) to enhance spectral features
  • Consider Multiplicative Scatter Correction (MSC) or Baseline Linear Correction (BLC) [4] [2]

• Model Development:

  • Split data into calibration (70%) and validation (30%) sets
  • Develop Partial Least Squares Regression (PLSR) models with full cross-validation
  • Evaluate model performance using R², RMSEP, RPD, and MAPE statistics

Protocol 2: Routine Mango Maturity Assessment in Orchards

This streamlined protocol is designed for practical, in-field maturity assessment by growers and technicians.

Materials and Equipment:

• Calibrated handheld NIR device with mango-specific models
• Sample collection bags and markers
• Data recording forms or mobile device

Procedure:

• Orchard Sampling Strategy:
  • Select 20-30 fruits per hectare, representing different canopy positions
  • Sample from multiple trees throughout the orchard
  • Include fruits from various sides of the tree (north, south, east, west)

• Standardized Measurement:

  • Wipe the fruit surface clean of dust and moisture
  • Hold the NIR device firmly against the fruit's cheek area
  • Ensure consistent pressure and orientation for all measurements
  • Take duplicate readings from opposite sides of each fruit

• Data Interpretation:

  • Record Dry Matter and TSS values directly from device display
  • Compare results with variety-specific maturity thresholds
  • Calculate average values and variability across the sample set

• Harvest Decision:

  • Determine optimal harvest window based on established maturity standards
  • For Palmer mangoes, harvest when DMC exceeds 150 g/kg [2]
  • For Kent mangoes, commercial standards may use different thresholds

Mango NIR Testing Workflow

Data Analysis and Chemometric Modeling

Performance Comparison of Modeling Techniques

Multiple chemometric approaches have been applied to NIR spectral data for mango quality prediction, each with distinct advantages and limitations.

Table 3: Comparison of Chemometric Techniques for Mango Quality Prediction

Modeling Technique	Principle	Best For	Performance (MAPE)	Advantages
Partial Least Squares Regression (PLSR)	Linear regression on latent variables	General purpose maturity models	R²: 0.81-0.87 [5]	Robust, interpretable
Support Vector Machine (SVM)	Nonlinear classification and regression	Variety classification	97-100% accuracy [5]	Handles nonlinearities
Multi-Predictor Local Polynomial Regression (MLPR)	Nonparametric local fitting	Nonlinear maturity patterns	<10% MAPE [6]	Flexible, data-driven
Artificial Neural Networks (ANN)	Multilayer nonlinear transformation	Complex spectral patterns	Research stage [1]	Powerful pattern recognition

Advanced Spectral Data Enhancement

Modern NIR spectroscopy applications employ various preprocessing techniques to enhance prediction performance by reducing noise and emphasizing meaningful spectral features:

• Multiplicative Scatter Correction (MSC): Compensates for additive and multiplicative scattering effects in reflectance measurements [4]
• Standard Normal Variate (SNV): Normalizes individual spectra to reduce the impact of path length variations and particle size effects [2]
• Savitzky-Golay Derivatives: Enhances resolution of overlapping peaks and eliminates baseline drift through polynomial smoothing [6] [2]
• Baseline Linear Correction (BLC): Removes linear baseline shifts caused by instrumental drift or sample matrix effects [4]

Research has demonstrated that proper spectral enhancement can significantly improve prediction accuracy, with some studies reporting improvement in R² values from 0.63 to 0.81 for acidity prediction after applying appropriate preprocessing techniques [4].

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful implementation of handheld NIR spectroscopy for mango maturity testing requires specific reagents, materials, and instrumentation.

Table 4: Essential Research Toolkit for NIR-Based Mango Maturity Assessment

Item	Specifications	Function/Application
Handheld NIR Spectrometer	740-1070 nm range, mango-specific models	Field-based spectral data acquisition
Digital Refractometer	0-32° Brix range, ±0.1° accuracy	TSS reference measurements
Laboratory Oven	70°C, forced air circulation	Dry matter content determination
pH Meter	±0.01 accuracy, temperature compensation	Acidity reference measurements
Savitzky-Golay Algorithm	2nd polynomial, 11-15 point window	Spectral preprocessing and derivative analysis
PLS Regression Software	Cross-validation, latent variable optimization	Chemometric model development
Standard Reference Materials	Ceramic reflectance standards	Instrument calibration and validation
Paniculoside II	Paniculoside II, MF:C26H40O9, MW:496.6 g/mol	Chemical Reagent
MAP855	MAP855, MF:C28H23ClF2N6O3, MW:565.0 g/mol	Chemical Reagent

Near-infrared spectroscopy represents a paradigm shift in mango maturity assessment, replacing subjective visual inspection with quantitative, data-driven decision making. The core principle of molecular bond interactions with NIR radiation enables simultaneous prediction of multiple internal quality parameters critical for determining optimal harvest timing. As handheld NIR technology continues to evolve with improved spectrometer miniaturization, enhanced computational power, and more robust chemometric models, its adoption throughout the mango supply chain promises to reduce postharvest losses, improve fruit quality consistency, and enhance consumer satisfaction. Future advancements in deep learning algorithms and multi-spectral data fusion will further strengthen the accuracy and applicability of this non-destructive technology for mango quality assurance.

The accurate determination of mango maturity is critical for ensuring fruit quality, optimizing harvest timing, and minimizing postharvest losses. Maturity indices provide objective criteria for predicting ripening potential and final eating quality. Traditional methods of assessing maturity often rely on destructive sampling, which is impractical for large-scale commercial operations. This document details the key physiological and biochemical indicators of mango maturity—Dry Matter (DM), Total Soluble Solids (TSS), Acidity (TA), and Starch—and outlines standardized protocols for their measurement using modern, non-destructive technologies, with a specific focus on handheld Near-Infrared (NIR) spectroscopy.

Core Maturity Indicators and Their Significance

The following indicators are well-established predictors of mango maturity and final quality.

Dry Matter (DM)

• Definition & Significance: Dry Matter represents the solid portion of the fruit remaining after water removal, comprising structural (e.g., cellulose) and non-structural (e.g., starch, sugars) compounds [7]. It is a reliable maturity index as it increases during fruit development at a rate of approximately 0.72% DM per week [8] [7]. Since DM is stable after harvest and highly correlated with TSS in ripe fruit (correlation >80%), it serves as an excellent predictor of final sweetness and eating quality [8] [7].
• Typical Values: The ideal harvest DM varies by cultivar and region. A minimum of 14% DM is often used, with common values ranging from 16.5% for Calypso and KP varieties in Australia to 14.0% for Brazilian 'Tommy Atkins' [7].

Total Soluble Solids (TSS)

• Definition & Significance: TSS, primarily composed of sugars, is a direct measure of fruit sweetness and a key consumer preference parameter [9]. During ripening, starch is converted into soluble sugars, leading to a marked increase in TSS [10] [7].
• Typical Values: TSS can be monitored non-destructively using NIR spectroscopy. For 'Tommy Atkins' mango, a handheld NIR spectrometer achieved a high-precision prediction of TSS with a coefficient of determination (R²) of 0.92 and a root mean square error of prediction (RMSEP) of 0.55 °Brix [11].

Titratable Acidity (TA)

• Definition & Significance: TA measures the concentration of organic acids, predominantly citric acid in mangoes, which contributes to the fruit's flavor profile [10] [12]. Acidity decreases as the fruit matures and ripens, and its measurement is crucial for assessing the sugar-to-acid balance, which defines taste perception [10] [13].
• Typical Values: NIR-based models have been successfully developed for TA prediction. For 'Tommy Atkins' mango, a model yielded an R² of 0.50 and an RMSEP of 0.17% citric acid [11]. Advanced regression approaches like Artificial Neural Networks (ANN) have shown improved performance, achieving a correlation coefficient (r) of 0.985 in calibration for intact mango [12].

Starch

• Definition & Significance: Starch is the primary carbohydrate reserve accumulated during mango fruit development [8] [7]. Its conversion to sugars during ripening is a critical metabolic process driving sweetness development. Starch content is, therefore, a fundamental indicator of physiological maturity at harvest [13].
• Typical Values: While quantitative ranges are cultivar-specific, starch content is used alongside Firmness, TA, and TSS in advanced classification models, such as fuzzy logic systems, to determine mango maturity indices with high accuracy [13].

Table 1: Summary of Key Maturity Indicators in Mango

Indicator	Chemical Basis	Relationship with Maturity	Typical Measurement Range (Varies by Cultivar)	Prediction Performance (NIR Example)
Dry Matter (DM)	Structural & non-structural solids (starch, sugars)	Increases during maturation	Harvest: ~14-17% [7]	R² = 0.67, RMSEP = 0.51% [11]
Total Soluble Solids (TSS)	Sugars (sucrose, glucose, fructose)	Increases during ripening	Ripened fruit: Varies widely	R² = 0.92, RMSEP = 0.55 °Brix [11]
Titratable Acidity (TA)	Organic acids (e.g., citric acid)	Decreases during maturation & ripening	Varies with maturity stage	R² = 0.50, RMSEP = 0.17% [11]; ANN: r = 0.985 [12]
Starch	Carbohydrate polymer	Decreases during ripening (converted to sugars)	High at harvest, low when ripe	Used in fuzzy logic models for maturity classification [13]
Firmness	Physical integrity of cell walls	Decreases during ripening	High at harvest, soft when ripe	iPLSR model: R²p = 0.75 [9]

Experimental Protocols for Handheld NIR-Based Assessment

This section provides a standardized workflow for developing and deploying NIR-based calibration models for mango maturity assessment.

Protocol 1: Development of a Firmness Calibration Model for 'Kent' Mango

This protocol is adapted from research on building a robust firmness model using interval Partial Least Squares Regression (iPLSR) [9].

• Sample Preparation:

  • Select a representative set of mango fruits (e.g., n=50) from the target cultivar at various maturity stages.
  • Acclimatize fruits to a constant ripening temperature (e.g., 20°C) and relative humidity (e.g., 85%).
  • Label each fruit and mark specific measurement spots on the fruit's lateral side.

• Spectral Data Acquisition:

  • Use a calibrated handheld NIR spectrometer (e.g., F-750 Produce Quality Meter).
  • Collect spectra in the NIR range (e.g., 700-1130 nm), excluding the visible region to minimize interference from skin color [9].
  • Take multiple scans (e.g., 6 scans) at each marked spot on the fruit and average them to improve signal-to-noise ratio.
  • Conduct spectral acquisition repeatedly over the ripening period (e.g., on alternating days for 10 days).

• Reference Data Collection:

  • Measure firmness at the same spots immediately after NIR scanning using a validated destructive (e.g., penetrometer) or non-destructive (e.g., acoustic) method [9].

• Chemometric Modeling:

  • Pre-processing: Apply spectral pre-processing techniques such as Savitzky-Golay smoothing to reduce high-frequency noise [9].
  • Data Splitting: Randomly split the dataset into a calibration set (e.g., 80% of samples) for model building and a prediction set (e.g., 20%) for validation.
  • Variable Selection: Employ iPLSR to identify key wavelength intervals most relevant to firmness. For 'Kent' mango, critical intervals were identified at 743–770 nm (associated with C-H and CHâ‚‚ bonds in sugars and cell wall components) and 870–905 nm (associated with CHâ‚‚ and CH₃ bonds) [9].
  • Model Building & Validation: Develop a PLSR model using the selected intervals. Validate model performance using the independent prediction set, reporting metrics like R² of prediction (R²p) and Root Mean Square Error of Prediction (RMSEP). The iPLSR model for 'Kent' mango showed a 12% improvement in R² and a 14% reduction in error compared to a full-spectrum PLSR model [9].

Protocol 2: Multi-Parameter Maturity Index Classification Using Fuzzy Logic

This protocol outlines an advanced approach for classifying mangoes into discrete maturity classes by integrating multiple parameters [13].

• Hardware and Software Setup:

  • Spectrometer: Utilize a portable NIR micro-spectrometer (e.g., NeoSpectra Micro, 1350–2500 nm).
  • Computing Unit: Integrate a compact computer like a Raspberry Pi.
  • Software: Develop a custom application (e.g., in Python) for data acquisition, model execution, and result display [13].

• Comprehensive Data Collection:

  • Spectral Acquisition: Collect spectra from multiple positions on each fruit. Perform extensive spectral pre-processing, testing up to 12 different spectral transformation operators (e.g., clipping, scatter correction, smoothing) to find the optimal input [13].
  • Reference Analytical Data: For each fruit, perform destructive tests to measure the four key parameters: TA, SSC (TSS), Firmness, and Starch [13].
  • Maturity Index Labeling: Classify each fruit into a maturity index (e.g., 80%, 85%, 90%, 95%, 100%) based on standard guidelines that consider Days After Full Bloom (DAF), flesh color, and taste [13].

• Model Development and Deployment:

  • Predictive Modeling: Build PLSR or ANN models to predict the four quantitative parameters (TA, SSC, Firmness, Starch) from the pre-processed NIR spectra.
  • Fuzzy Logic Integration: Implement a fuzzy logic system that uses the predicted values of the four parameters as inputs. Define fuzzy sets and rules to map the combinations of these inputs to the five discrete maturity indices [13].
  • Validation: This indirect classification approach has been shown to achieve high accuracy (95.7%), outperforming direct classification models [13].

The workflow below visualizes the key steps in a handheld NIR-based maturity assessment program.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Key Research Reagents and Equipment for NIR-based Mango Maturity Analysis

Item	Function/Description	Example Products/Models
Handheld NIR Spectrometer	Core device for non-destructive spectral acquisition in field or lab.	Felix Instruments F-751 Mango Quality Meter [7], F-750 Produce Quality Meter [9], NeoSpectra Micro [13], Scio [13]
Calibration Standards	Reference materials for spectrometer calibration to ensure measurement accuracy.	Barium Sulfate (BaSOâ‚„) pellets or discs [13]
Reference Analytical Instruments	For destructive measurement of reference values to build calibration models.	Penetrometer (Firmness), Refractometer (TSS/Titratable Soluble Solids), Titration Kit (TA), Laboratory Oven (DM) [9] [13]
Computing & Control Unit	For device control, data processing, and model execution in portable systems.	Raspberry Pi, Intel Compute Stick [8] [13]
Data Analysis Software	For spectral pre-processing, chemometric modeling, and algorithm development.	Python (with scikit-learn, PyPLS), MATLAB, R, Proprietary SDKs [13]
Quality Metrics	Statistical parameters to validate the performance and reliability of calibration models.	Coefficient of Determination (R²), Root Mean Square Error (RMSE/RMSEC/RMSEP), Ratio of Prediction to Deviation (RPD) [12] [9] [11]
Purpurogallin	Purpurogallin, MF:C11H8O5, MW:220.18 g/mol	Chemical Reagent
TMPyP4 tosylate	TMPyP4 tosylate, MF:C72H70N8O12S4+4, MW:1367.6 g/mol	Chemical Reagent

Near-infrared (NIR) spectroscopy has emerged as a cornerstone technology for the non-destructive assessment of fruit internal quality attributes, with its efficacy being profoundly influenced by the selected optical geometry. The configuration of the light source, fruit sample, and detector—collectively termed optical geometry—determines the type and extent of light-tissue interaction, thereby dictating the quality and nature of the spectral data acquired [1]. For researchers focused on handheld NIR method development for mango maturity testing, the choice between reflectance, interactance, and transmittance modes represents a critical methodological decision that directly impacts prediction accuracy for key maturity indices such as dry matter content (DMC), soluble solids content (SSC), and flesh color [14] [15].

This application note provides a structured comparison of these fundamental optical geometries, detailing their underlying principles, relative performance characteristics, and implementation protocols specifically contextualized within mango maturity research. The guidance presented herein aims to equip researchers with the necessary knowledge to select and optimize optical configurations for robust, field-deployable mango maturity assessment systems.

Comparative Analysis of Optical Geometries

The performance of reflectance, interactance, and transmittance modes varies significantly based on the target attribute and fruit characteristics. The table below summarizes their key operational and performance characteristics.

Table 1: Comparison of Optical Geometries for Fruit NIR Spectroscopy

Feature	Reflectance	Interactance	Transmittance
Basic Principle	Measures light reflected from the fruit surface and immediate subsurface layers [16].	Measures light that has penetrated the fruit and scattered back out, with the detector field of view separated from the illuminated area by a light seal [14].	Measures light that has passed entirely through the fruit, with the detector positioned diametrically opposite the light source [14].
Typical Setup Diagram	Light Source → Fruit → Detector (same side)	Light Source \| Fruit \| Detector (same side, with light barrier)	Light Source → Fruit → Detector (opposite sides)
Penetration Depth	Shallow; primarily probes surface and near-surface properties [16].	Intermediate; captures information from both surface and partial internal layers [16].	Deep; probes the entire flesh volume between the source and detector [14].
Key Advantage	Easy to implement, no contact required, high signal intensity [14].	A good compromise, less susceptible to surface properties than reflectance [14].	Potentially better for detecting deep internal disorders and properties [14].
Key Limitation	Susceptible to variations in superficial properties (e.g., skin color, roughness) [14].	Requires a physical light seal, which can be challenging on high-speed conveyor belts [14].	Very low light signal, requiring sensitive detectors and potentially longer acquisition times [14].
Suitability for Thick/Rind Fruit	Limited for internal quality of thick-skinned fruits like mango [17].	Well-suited, as it can probe beyond the thick skin of a mango.	Highly suitable for internal quality assessment, though signal strength can be very low [17].
Reported Performance (Example)	In kiwifruit, provided good SSC calibrations but was less accurate than interactance [14].	In kiwifruit, provided the most accurate results for SSC, density, and flesh hue [14].	In kiwifruit, spectral range was limited to 700–950 nm; less accurate than interactance [14].

Optical Geometry Configurations

The following diagrams illustrate the fundamental configurations and data processing workflows for the three primary optical geometries used in fruit NIR spectroscopy.

Figure 1: Optical geometry configurations for (A) Reflectance, (B) Interactance, and (C) Transmittance modes. Note the critical light seal in interactance mode that prevents surface-reflected light from reaching the detector.

Experimental Protocol for Geometry Comparison in Maturity Testing

This protocol provides a standardized methodology for evaluating the performance of different optical geometries for assessing mango maturity parameters, specifically DMC and SSC.

Research Reagent Solutions and Essential Materials

Table 2: Essential Materials for NIR-based Maturity Assessment Experiments

Item Category	Specific Examples & Models	Critical Function
NIR Spectrometer	Portable devices (e.g., Felix Instruments F-750), USB2000+, miniature spectrometers (e.g., Hamamatsu C11708MA) [16] [18].	Acquires spectral data in the Vis/NIR range (e.g., 640–1050 nm or 300–1100 nm).
Light Source	Halogen lamp (e.g., Welch Allyn 997418, 1.5W) [16].	Provides stable, broad-spectrum illumination in the NIR region.
Optical Setup	Light seal (for interactance), probe holder, integrating sphere (for diffuse reflectance) [14] [19].	Defines and maintains the specific optical geometry during measurement.
Reference Analytics	Digital refractometer (for SSC), oven (for DMC), texture analyzer (for firmness) [15] [1].	Provides destructive reference measurements for model calibration and validation.
Calibration Standards	Polytetrafluoroethylene (PTFE) white reference board, dark current standard [16].	Calibrates the spectrometer before sample measurement to ensure data consistency.
Data Analysis Software	Python with scikit-learn, MATLAB, or proprietary chemometrics software [15] [20].	Performs spectral preprocessing, feature selection, and regression/classification modeling.

Sample Preparation and Spectral Acquisition

• Sample Selection: Obtain a minimum of 120 mango fruits (e.g., 'Keitt' or 'Kent') harvested across different maturity stages—from one week before to one week after the optimal commercial harvest date—to ensure a representative range of DMC and SSC values [15].
• Sample Conditioning: Transport fruits to the lab and store at 4°C in sealed polyethylene bags. Before measurement, wipe samples clean, number them sequentially, and equilibrate to room temperature (e.g., 22 ± 1°C) for 24 hours to minimize spectral interference from temperature variation [16].
• Instrument Calibration: Power on the NIR system and allow a 5-minute warm-up. Collect reference spectra using a white PTFE calibration tile and dark current to establish baseline and background signals [16].
• Spectral Collection:
  • For each fruit, collect spectra at three equidistant points along the equatorial region.
  • For each geometry mode, ensure consistent and firm placement of the fruit against the sensor or measurement aperture. In interactance mode, verify the light seal is flush with the fruit surface.
  • Rotate the fruit to collect the triplicate measurements and compute the average spectrum for each fruit to mitigate spatial heterogeneity [16].
• Reference Measurement: Immediately after spectral acquisition, destructively measure SSC via a digital refractometer and DMC using standard oven-drying methods on tissue plugs extracted from the scanned locations [15] [1].

Data Processing and Model Building Workflow

Figure 2: Workflow for developing chemometric models from spectral data to predict mango maturity attributes.

• Spectral Preprocessing: Apply preprocessing algorithms to mitigate noise and light scattering effects. Common techniques include:
  • Savitzky-Golay (SG) Smoothing: Reduces high-frequency noise [16].
  • Standard Normal Variate (SNV): Corrects for multiplicative scatter effects [14].
  • Detrending: Removes baseline shifts by fitting and subtracting a polynomial trend [14].
• Feature Wavelength Selection: Employ variable selection methods to identify the most informative wavelengths and reduce model complexity.
  • Competitive Adaptive Reweighted Sampling (CARS): Effectively selects key wavelengths related to SSC and DMC, often leading to robust models [15] [16].
  • Successive Projections Algorithm (SPA): Another common method for selecting feature wavelengths with minimal collinearity [15].
• Model Development: Construct regression models to predict DMC and SSC from the processed spectra.
  • Partial Least Squares (PLS) Regression: The most widely used linear method for relating spectral data to reference values [15] [16].
  • Least Squares-Support Vector Machine (LS-SVM): A non-linear algorithm that can improve prediction accuracy for complex attributes [15].
  • Deep Learning Models (e.g., 1D-CNN): Can automatically extract features from raw or preprocessed spectra and may enhance performance [20].
• Model Validation: Assess model performance using an independent validation set of fruits not used in model calibration. Key performance metrics include:
  • Coefficient of Determination (R²): The proportion of variance in the reference method explained by the model.
  • Root Mean Square Error (RMSE): The average magnitude of prediction error.
  • Residual Predictive Deviation (RPD): The ratio of the standard deviation of the reference data to the RMSE; values above 1.4 indicate models with some predictive utility [21].

Application to Handheld NIR Devices for Mangoes

For handheld NIR spectrometer development targeting mango maturity, the selection of optical geometry involves critical trade-offs. While research indicates interactance mode often provides superior accuracy for internal properties in fruits like kiwifruit [14], its implementation on a handheld device is challenging due to the requirement for a physical light seal.

Diffuse Reflectance offers a practical compromise for handheld design, being easier to implement without contact. Studies on kiwifruit have shown that with optimal preprocessing (e.g., SG smoothing combined with CARS feature selection), diffuse reflectance can achieve high prediction accuracy for SSC (R² = 0.98) [16]. However, researchers must be aware that calibrations can be susceptible to variations in superficial properties like skin color and roughness [14].

Ultimately, the choice must align with the core research objectives: whether to prioritize maximum potential accuracy (favoring interactance) or practical design simplicity and cost-effectiveness (favoring reflectance), while acknowledging that thick rinds can limit the effectiveness of reflectance for internal quality assessment [17].

Advantages of Non-Destructive Testing for Supply Chain Management and Quality Control

Non-destructive testing (NDT) represents a critical methodology for evaluating materials, components, and structures without causing damage. Within supply chain management and quality control, NDT enables continuous verification of product integrity from manufacturing through distribution. This application note examines the specific advantages of NDT implementation, with particular focus on handheld Near-Infrared (NIR) spectroscopy for mango maturity testing as a case study. We detail protocols, data interpretation methods, and practical implementation frameworks to guide researchers and quality assurance professionals in adopting these methodologies to enhance product quality, reduce waste, and optimize supply chain efficiency.

Non-destructive testing (NDT), also referred to as non-destructive evaluation (NDE) or inspection (NDI), encompasses a range of analysis techniques used to evaluate material properties, component integrity, and product quality without causing damage to the original specimen [22]. Unlike destructive testing methods that require samples to be pushed to failure, NDT allows products that pass inspection to remain in service or continue through the supply chain, creating significant efficiencies [22].

The fundamental principle of NDT involves using scientific processes to examine materials through techniques such as electromagnetic testing, visual inspection, and radiographic analysis. These methods can detect surface and subsurface defects, measure material properties, and verify quality parameters while preserving the utility of the tested item [22]. This capability makes NDT particularly valuable for quality control processes where preserving product integrity is essential.

Core Advantages of NDT in Supply Chain and Quality Control

Enhanced Quality Assurance and Control

NDT plays a vital role in comprehensive quality assurance programs by detecting defects and irregularities that could compromise product performance [23] [24]. Through techniques that identify surface cracks, internal flaws, and material inconsistencies, NDT enables timely corrections before products advance through the supply chain [23]. This proactive quality management ensures consistent product reliability and compliance with client specifications and industry standards [24]. The ability to test products without sacrifice allows for more frequent quality checks throughout manufacturing processes, leading to better overall quality control [25].

Cost Efficiency and Waste Reduction

The non-destructive nature of these testing methods delivers significant cost advantages across multiple dimensions. By eliminating the need to destroy products for testing, NDT substantially reduces material waste and associated costs [23]. Early defect identification prevents costly repairs, recalls, or product failures later in the supply chain, generating long-term savings [23] [24]. Additionally, NDT can be performed without disassembling components or shutting down production lines, minimizing operational downtime [24]. The preservation of tested products represents a fundamental economic advantage over destructive methods [26].

Improved Safety and Risk Mitigation

NDT contributes significantly to safety enhancement by identifying potential failures before they result in accidents [23] [24]. In critical industries such as aerospace, automotive, and infrastructure, NDT helps ensure that components meet strict safety standards [23]. The methodologies also present safer working conditions for testing personnel compared to some destructive testing methods [26]. Furthermore, by preventing structural failures and accidents, NDT helps mitigate environmental risks associated with material failures [24].

Supply Chain Optimization

The implementation of NDT creates multiple supply chain benefits. The ability to test products without damage enables 100% inspection rates where appropriate, providing comprehensive quality data across production batches [25]. Rapid inspection techniques allow for real-time quality decisions at various points in the supply chain, from manufacturing to distribution [27]. The non-destructive nature also supports sustainable operations by reducing material waste and associated resource consumption [26].

Table 1: Quantitative Benefits of NDT Implementation in Industrial Settings

Benefit Category	Key Metrics	Impact Level
Cost Management	Reduction in material waste, Lower repair/recall costs, Decreased downtime	Significant
Quality Performance	Early defect detection rate, Compliance with standards, Customer satisfaction	High
Operational Efficiency	Testing time reduction, Throughput improvement, Downtime minimization	Moderate to High
Risk Management	Safety incident reduction, Environmental risk mitigation, Regulatory compliance	High

Handheld NIR Spectroscopy for Mango Maturity Testing: A Case Study

Near-Infrared Spectroscopy (NIRS) has emerged as a powerful non-destructive technique for assessing fruit quality parameters. The method utilizes the interaction between near-infrared light (typically 740-2500 nm) and molecular bonds in organic compounds to determine chemical composition [1]. For mango maturity assessment, handheld NIR spectrometers provide portability for field use while maintaining analytical precision [5] [28]. These instruments measure absorption characteristics related to critical maturity indicators including dry matter content (DM), total soluble solids (TSS), and pH [28] [6].

Key Advantages for Supply Chain Management

The application of handheld NIR spectroscopy to mango maturity assessment demonstrates how NDT creates value throughout agricultural supply chains. By enabling non-destructive testing, the method allows 100% testing of inbound and outbound fruit without waste generation [1]. The rapid analysis capability (typically seconds per measurement) supports high-throughput operations at packing facilities and distribution centers [6]. Accurate maturity classification facilitates optimal harvest timing and post-harvest handling, reducing losses during storage and transport [28]. Furthermore, objective quality data enables standardized quality grading across supply chain partners, minimizing disputes and ensuring consistent quality for end consumers [1].

Table 2: Mango Quality Parameters Measurable via Handheld NIR Spectroscopy

Quality Parameter	Measurement Range	Typical Accuracy (R²)	Supply Chain Significance
Dry Matter Content (DM)	10-25%	0.80-0.95	Primary maturity index; determines harvest timing
Total Soluble Solids (TSS)	5-20°Brix	0.70-0.90	Indicator of sweetness and eating quality
pH	3.0-4.5	0.65-0.85	Measures acidity level; affects flavor profile
Maturity Classification	Mature/Immature	85-95% accuracy	Direct sorting decision capability

Experimental Protocols: Handheld NIR for Mango Quality Assessment

Protocol 1: Direct Maturity Classification Using Handheld NIR

Purpose: To classify mangoes into mature and immature categories using handheld NIR spectroscopy for supply chain sorting decisions.

Materials and Equipment:

• Handheld NIR spectrometer (wavelength range: 400-1100 nm)
• Reference mango samples with known maturity status
• Computing device with classification software
• Sample presentation fixture

Procedure:

• Instrument Calibration: Initialize the NIR spectrometer according to manufacturer specifications. Perform wavelength calibration using certified reference materials.
• Reference Data Collection: For training samples, determine maturity status using standard destructive methods for dry matter content (AOAC official method).
• Spectral Acquisition: Position the mango fruit to ensure consistent optical geometry. Acquire interactance spectra from three equidistant points around the equatorial region of each fruit.
• Data Preprocessing: Apply standard normal variate (SNV) transformation to reduce scattering effects. Use Savitzky-Golay smoothing (window: 9 points, polynomial order: 2) to reduce noise.
• Model Development: Implement k-nearest neighbors (KNN) classifier with k=5 using principal components derived from preprocessed spectra as inputs.
• Validation: Evaluate classification accuracy using independent test set not included in model development.

Data Interpretation: The direct classification approach has demonstrated 88.2% accuracy in distinguishing mature from immature mangoes, significantly outperforming indirect estimation methods (55.9% accuracy) [28].

Protocol 2: Quantitative Prediction of Mango Quality Parameters

Purpose: To predict critical mango quality parameters (pH and TSS) using multi-predictor local polynomial regression (MLPR) modeling of NIR spectral data.

Materials and Equipment:

• Handheld NIR spectrometer (wavelength range: 740-1070 nm)
• Digital refractometer (for TSS reference)
• pH meter with glass electrode (for pH reference)
• Sample preparation equipment
• Computing device with multivariate analysis software

Procedure:

• Sample Preparation: Select mango samples at different maturity stages. Clean fruit surface to remove contaminants.
• Reference Analysis: For each sample, extract juice and measure TSS using digital refractometer and pH using calibrated pH meter.
• Spectral Collection: Scan intact mango fruits using NIR spectrometer in interactance mode. Collect spectra from two opposing sides of each fruit.
• Data Preprocessing: Apply appropriate preprocessing techniques: Gaussian filter smoothing for pH prediction, Savitzky-Golay smoothing for TSS prediction.
• Model Development: Develop MLPR models using polynomial order of 2 and bandwidth optimized through cross-validation.
• Model Validation: Assess model performance using k-fold cross-validation (k=10). Calculate R², RMSEP, and MAPE for quality assessment.

Data Interpretation: MLPR has demonstrated superior performance for predicting mango quality parameters, with MAPE values less than 10% and R² values of 0.63 for TSS and 0.81 for pH in validation sets [6]. This represents significantly better accuracy compared to kernel partial least squares regression (KPLSR) and support vector machine regression (SVMR) approaches.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Materials for Handheld NIR Maturity Assessment Research

Item	Specifications	Function/Application
Handheld NIR Spectrometer	Wavelength range: 400-1100 nm or 740-1070 nm; Embedded computing capability	Primary data acquisition instrument for field-based spectral collection
Reference Analytical Instruments	Digital refractometer (0-32°Brix), Laboratory pH meter with temperature compensation	Establishment of reference values for model calibration and validation
Standard Reference Materials	Certified wavelength standards, Physical calibration standards	Instrument calibration and verification of measurement accuracy
Chemometric Software	Multivariate analysis capabilities (PLS, MLPR, SVM, CNN)	Data preprocessing, model development, and prediction
Sample Presentation Fixtures	Black anodized aluminum with fixed geometry	Minimizes spectral variability through consistent positioning
Poloxipan	Poloxipan, CAS:606955-72-0, MF:C14H10BrN3O3S, MW:380.22 g/mol	Chemical Reagent
Nystatin	Nystatin, MF:C47H75NO17, MW:926.1 g/mol	Chemical Reagent

Implementation Framework for Supply Chain Integration

Successful implementation of NDT methodologies like handheld NIR spectroscopy requires strategic planning across organizational and technical dimensions. Based on successful case studies in fruit supply chains, we recommend the following implementation framework:

Technology Selection Criteria: When selecting handheld NIR systems for supply chain quality control, consider wavelength range appropriate for target parameters (DM, TSS, pH), measurement speed compatible with operational throughput requirements, robustness for field and packinghouse environments, and compatibility with existing data management systems [1] [28].

Data Integration Architecture: Implement centralized data repositories for spectral data and quality measurements across supply chain nodes. Develop standardized data formats to enable quality tracking from harvest through distribution. Create visualization dashboards for real-time quality monitoring and decision support [27].

Personnel Training Protocols: Establish comprehensive training programs for technical staff covering instrument operation, measurement protocols, basic troubleshooting, and data interpretation. Implement certification procedures to ensure measurement consistency across operators and locations [22].

Continuous Improvement Processes: Develop feedback mechanisms to regularly update calibration models with new seasonal data and varieties. Establish performance metrics for prediction accuracy and operational impact. Create cross-functional teams to identify and implement improvement opportunities throughout the supply chain [6].

The integration of non-destructive testing methodologies, particularly handheld NIR spectroscopy, offers transformative potential for supply chain management and quality control systems. The case study in mango maturity assessment demonstrates how these technologies deliver tangible benefits including enhanced quality control, reduced waste, improved supply chain efficiency, and objective quality standardization. The experimental protocols detailed provide practical frameworks for implementation, while the technical toolkit guides resource allocation decisions. As NDT technologies continue to advance through innovations in artificial intelligence, miniaturization, and data analytics, their application across supply chains will expand, creating new opportunities for quality optimization and value creation throughout product lifecycles.

From Lab to Orchard: Implementing Handheld NIR Systems for Mango Analysis

The accurate, non-destructive assessment of mango maturity is critical for determining optimal harvest time, which directly influences postharvest quality, marketability, and consumer acceptance. Near-infrared (NIR) spectroscopy has emerged as a leading technology for this purpose, capable of quantifying key internal quality attributes like Dry Matter (DM) and Soluble Solids Content (SSC) without destroying the fruit. Researchers and engineers implementing this technology face a fundamental choice: to use commercial handheld sensors or to develop custom prototype systems. This application note details both approaches, providing a structured comparison and detailed experimental protocols to guide hardware selection and implementation within a research thesis on handheld NIR methods for mango maturity testing.

Commercial-Off-The-Shelf (COTS) NIR Sensors

Commercial handheld NIR spectrometers offer a complete, validated hardware and software solution, enabling researchers to focus on data collection and analysis rather than instrument design.

Key Commercial Devices

The table below summarizes the specifications and applications of two prominent COTS sensors used in agricultural research.

Table 1: Comparison of Commercial Handheld NIR Sensors for Maturity Assessment

Feature	NeoSpectra Handheld NIR Analyzer (Si-Ware)	F-750 Produce Quality Meter (Felix Instruments)
Underlying Technology	Fourier-Transform (FT) based on MEMS [29]	Not explicitly specified in search results, but widely used in research [8]
Spectral Range	1350 - 2550 nm [29]	500 - 1100 nm (as used in cited research) [8]
Key Strengths	High consistency between devices; Pre-compiled spectral libraries available [29]	Optimized for fresh produce; Integrated model for DM and SSC [8]
Reported Mango Application	General NIR spectral library development (soil focus, but technology is applicable) [29]	Direct maturity classification and maturity index (e.g., DM) estimation [8]
Best Suited For	Research requiring high spectral resolution and transferable models [29]	Applied agricultural research with a need for immediate, on-site results [8]

Experimental Protocol for COTS Sensors

Title: Protocol for Non-Destructive Mango Maturity Assessment Using a Commercial Handheld NIR Sensor.

1. Hardware and Software Setup:

• Device: Use a calibrated commercial handheld NIR spectrometer (e.g., F-750 or NeoSpectra).
• Calibration: Ensure the device is calibrated according to the manufacturer's instructions. For the F-750, select or develop a calibration model specific to mangoes [8].
• Accessories: Ensure the device is fully charged. For consistent measurements, a dedicated fruit holder or a setup to maintain a fixed distance between the sensor and the fruit is recommended.

2. Sample Preparation:

• Fruit Selection: Select mango fruits (e.g., 'Samar Bahisht Chaunsa', 'Sufaid Chaunsa') at various maturity stages, from one week before the estimated harvest date to one week after [8].
• Sample Size: A minimum of 120 fruits is recommended for robust model development [8].
• Labeling and Conditioning: Label each fruit and clean the measurement site. Allow fruits to equilibrate to room temperature if they were stored in a cool environment.

3. Data Acquisition:

• Measurement Geometry: Position the sensor so that the light beam is perpendicular to the fruit's surface. For mangoes, the cheek region is typically used [1].
• Scanning Procedure: Firmly place the sensor's window on the fruit's surface. Trigger the scan as per the device's operating procedure. Take multiple scans (e.g., 3-5) per fruit at different positions around the equatorial region and average the spectra to improve representativeness [29].
• Data Logging: The device typically logs the spectrum and, if using a pre-calibrated model, predicts values (e.g., DM) directly.

4. Data Processing and Analysis:

• For Direct Classification (F-750-like systems): Use the device's integrated model to classify maturity (e.g., "Harvest" vs. "Do-not-harvest") based on the acquired spectrum [8].
• For Spectral Library Building (NeoSpectra-like systems): Transfer the spectra to a computer for further chemometric analysis. Use machine learning algorithms (e.g., Partial Least Squares - Discriminant Analysis, PLS-DA) to build a classification or regression model [1] [6].

Custom-Built NIR Prototypes

For applications requiring specific features, cost constraints, or educational purposes, building a custom NIR spectrometer using embedded systems like Raspberry Pi is a viable alternative.

Key Components of a Custom Prototype

The table below lists the essential components for building a custom handheld NIR spectrometer, as evidenced by recent research and student projects.

Table 2: Essential Components for a Custom Raspberry Pi-Based NIR Spectrometer

Component Category	Example Parts	Function
Microprocessor / Single-Board Computer	Raspberry Pi 3/4, Raspberry Pi RP2040 [30] [31]	The computational core of the device; runs the operating system, controls sensors, and processes data.
Spectral Sensor	AS7265x (Triad; VIS/NIR), STS-NIR (Ocean Optics), DLP NIRscan Nano [32] [8] [31]	The core sensor that acquires the spectral data in specific wavelength ranges.
Light Source	Tungsten Halogen Lamp, High-Power LEDs [8]	Illuminates the sample. The choice affects the penetration depth and signal-to-noise ratio.
Power Management	Lithium Polymer/Polymer Battery, Voltage Regulators [31]	Provides stable, portable power for all components.
User Interface & Data Output	E-Paper Display (for low power), LCD Touchscreen, USB/Bluetooth [32] [31]	Allows user interaction, displays results, and enables data transfer.
Enclosure	3D-Printed Shell [31]	Protects the internal electronics and provides an ergonomic housing.

Experimental Protocol for Custom Prototypes

Title: Protocol for Building and Validating a Custom Raspberry Pi-Based NIR Spectrometer for Mango Maturity.

1. Hardware Assembly and Integration:

• Core System: Connect a Raspberry Pi (e.g., Model 3) to a spectral sensor development kit (e.g., AS7265x Triad) via the I2C communication protocol [32].
• Power Circuit: Design and solder a power management board that integrates a lithium battery (e.g., 3.7V 2000mAh) with a charging circuit and voltage regulators to provide stable power to the Pi and sensor [31].
• Illumination: For reflectance measurements, attach a high-power LED light source. A Bluetooth-controlled circuit can be implemented to toggle the light [8].
• Enclosure: Design and 3D-print a shell that holds all components, ensures correct optical geometry (sensor-to-sample distance), and is comfortable to hold [31].

2. Software and Firmware Development:

• Programming: Develop code in C++ or Python to control the sensor, acquire spectra, and perform basic processing. Leverage open-source libraries from sensor manufacturers (e.g., SparkFun for AS7265x) [32].
• User Interface (UI): Implement a simple UI on an E-Paper or LCD display to initiate scans and show results. E-Paper is advantageous for its sunlight readability and low power consumption [32].
• Data Handling: Program the system to save spectral data (e.g., as CSV files) to an SD card or transmit it via Bluetooth/Wi-Fi.

3. System Calibration and Validation:

• Wavelength Calibration: Use manufacturer-provided calibration or validate against known wavelength standards.
• Model Development: This is the most critical step. Collect spectra from a large set of mango samples (N > 150) with known reference DM and SSC values (obtained destructively). Use chemometric software (e.g., in Python or R) to preprocess spectra (Smoothing, SNV, Derivatives) and develop a Partial Least Squares (PLS) regression or Support Vector Machine (SVM) model [31] [6].
• Model Deployment: Implement the finalized PLS/SVM model coefficients into the Raspberry Pi's software to enable real-time prediction of DM or SSC during a scan.

The Scientist's Toolkit: Research Reagent Solutions

This table catalogs key hardware and software "reagents" essential for experiments in handheld NIR spectroscopy for mango maturity.

Table 3: Essential Research Reagents for Handheld NIR Maturity Testing

Item Name	Function / Purpose	Example Sources / Types
Reference Mango Set	A set of mango samples with precisely measured reference values (DM, SSC, pH) for model calibration and validation.	Fruits harvested at different times from verifiable orchards [8] [6].
Chemometrics Software	Software for developing predictive models from spectral data. Used for preprocessing, variable selection, and regression/classification.	PLS Toolbox (MATLAB), Unscrambler, or open-source packages in R (e.g., pls) and Python (e.g., scikit-learn, PyPLS) [8] [6].
Standard Preprocessing Algorithms	Mathematical techniques to reduce noise and enhance spectral features before model building.	Savitzky-Golay Smoothing & Derivatives, Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), Detrending [8] [6].
Wavelength Selection Algorithms	Methods to identify the most informative wavelengths, simplifying the model and improving robustness.	Genetic Algorithm (GA), Successive Projections Algorithm (SPA), synergy interval (si) PLS [31].
Validation Metrics	Statistical parameters to objectively evaluate model performance.	Coefficient of Determination (R²), Root Mean Square Error of Prediction (RMSEP), Mean Absolute Percentage Error (MAPE) [31] [6].
(Rac)-Atropine-d3	Atropine | High-Purity Anticholinergic Agent	High-purity Atropine for research. A muscarinic antagonist for neurology, ophthalmology & toxicology studies. For Research Use Only. Not for human consumption.
PF-3644022	PF-3644022, MF:C21H18N4OS, MW:374.5 g/mol	Chemical Reagent

Direct Comparison and Decision Framework

The choice between COTS and custom solutions involves critical trade-offs.

Table 4: Decision Matrix: Commercial Sensors vs. Custom Prototypes

Criterion	Commercial Sensors (F-750, NeoSpectra)	Custom Prototypes (Raspberry Pi)
Development Time	Short ("out-of-the-box" solution)	Long (requires hardware assembly, programming, and calibration)
Cost	High initial investment per unit	Lower per-unit cost, but requires engineering expertise [32]
Flexibility & Control	Limited to manufacturer's specifications	High (sensor choice, spectral range, housing design can be customized) [31]
Performance & Accuracy	High, validated, and consistent [29]	Variable; highly dependent on design choices and calibration model quality [31]
Ease of Use	High (integrated software and models)	Lower (requires technical knowledge to operate and maintain)
Best For	Applied research, rapid deployment, studies requiring validated and comparable data.	Methodological research, cost-sensitive projects, educational purposes, and highly specific applications.

For a thesis on handheld NIR methods for mango maturity, the hardware decision rests on the core research question. If the goal is to validate the application of NIR for maturity assessment in a specific cultivar or growing condition, a commercial sensor like the F-750 provides a reliable and rapid path to generating publishable results. Conversely, if the thesis aims to explore novel sensor technologies, optimize hardware configurations, or develop new low-cost form factors, then a custom prototype based on a Raspberry Pi and a spectral sensor is the appropriate choice. This path offers unparalleled insight into the entire NIR system pipeline, from photons to predictions, albeit with a significantly higher development burden. Both pathways are valid and contribute profoundly to the advancement of non-destructive quality assessment in horticulture.

This document outlines the standard operating procedures for data acquisition using handheld Near-Infrared (NIR) spectroscopy, specifically tailored for research on mango maturity testing. The protocols cover critical steps from sample preparation to instrument calibration, ensuring the collection of robust and reproducible spectral data for predicting internal quality attributes such as Dry Matter (DM) and Total Soluble Solids (TSS). Adherence to these guidelines is fundamental for developing accurate chemometric models.

Sample Preparation Protocol

Proper sample preparation is crucial for minimizing variability and enhancing the signal-to-noise ratio in spectral data.

Sample Collection and Selection

• Representativeness: Collect mango samples that represent the full range of maturity stages, cultivars, and growing conditions (e.g., different orchards, seasons) expected in the application [1] [33]. A minimum of 50-100 fruits is recommended for a provisional calibration, with larger sample sets (>100) yielding more robust models [33].
• Physical Inspection: Select fruits free of visible external defects, bruises, or disease, as these can alter spectral signatures unrelated to maturity.

Sample Handling and Presentation

• Temperature Equilibrium: Allow mango samples to equilibrate to a consistent room temperature (e.g., 20-25°C) before spectral measurement. Temperature fluctuations can cause significant baseline shifts in NIR spectra.
• Surface Cleaning: Gently clean the fruit surface with a soft, dry cloth to remove dust and moisture. Do not use solvents or water, as they can leave residues or alter the spectral properties of the skin.
• Measurement Positioning: Mark a fixed measurement spot on the cheek of each mango, typically midway between the stem and the apex, avoiding the suture line. Using a sample holder or a marked template ensures consistent positioning and orientation of the fruit for every scan.

Spectral Acquisition and Range Selection

The configuration of the spectrometer and the choice of spectral range directly impact the information content of the data.

Spectral Range and Instrument Parameters

For mango maturity assessment, the recommended spectral range is 740–2500 nm, which captures overtones and combinations of vibrations from key chemical bonds (O-H in water, C-H in sugars, etc.) [1].

Table 1: Key Spectral Acquisition Parameters for Handheld NIR

Parameter	Recommended Setting	Rationale
Spectral Range	780 - 2500 nm	Covers fundamental vibrations for DM and TSS [1] [18]
Optical Geometry	Interactance or Reflectance	Suitable for measuring internal attributes of thick-skinned fruits like mangoes [1]
Scan Resolution	≤ 10 nm	Higher resolution helps resolve overlapping absorption peaks
Number of Scans	32 - 64 per spectrum	Averaging multiple scans reduces random noise and improves signal quality

Data Acquisition Workflow

The following diagram illustrates the end-to-step workflow for acquiring and processing NIR spectra from mango samples.

Calibration and Validation

NIR spectroscopy is a secondary analytical method, meaning it requires calibration against primary reference data to build predictive models.

Reference Method Data

• Destructive Analysis: Immediately after NIR scanning, the same mango should be destructively analyzed for reference values.
  • Dry Matter (DM): Determined by drying a flesh sample in an oven at a specific temperature (e.g., 105°C) until constant weight [1].
  • Total Soluble Solids (TSS): Measured using a digital refractometer on the extracted juice, expressed as °Brix [1].
• Data Quality: The accuracy of the NIR calibration is limited by the precision of these reference methods. Use validated protocols and replicate measurements to minimize reference method error [33].

Calibration Development Workflow

The process of developing a functional NIR calibration model involves several key stages, from sample selection to model deployment.

Table 2: NIR Calibration Model Development Workflow

Step	Action	Key Considerations
1. Sample Collection	Assemble a calibration set (n ≥ 50) covering the full range of DM and TSS values [33].	Ensure sample variability represents future unknown samples.
2. Spectral Acquisition	Scan all samples using the protocol in Section 3.	Consistent conditions are critical.
3. Reference Analysis	Perform destructive DM/TSS analysis on each scanned fruit.	Primary method accuracy limits NIR model performance.
4. Data Preprocessing	Apply techniques like Standard Normal Variate (SNV) to reduce scatter [34].	Improves signal-to-noise ratio and model robustness.
5. Model Regression	Use algorithms like Partial Least Squares (PLS) regression to correlate spectra to reference data [33].	A common and effective method for NIR data.
6. Validation	Test the model on an independent set of samples not used in calibration.	Prevents overfitting and tests real-world predictive ability.

Wavelength Selection for Robust Calibration

Full-spectrum data contains many variables; selecting the most informative wavelengths simplifies models and improves performance.

• Objective: Identify wavelengths with the highest correlation to the constituent of interest (e.g., DM, TSS) while eliminating uninformative or redundant variables [34].
• Advanced Method: A hybrid approach combining Mutual Information (mRMR) for filtering and a Genetic Algorithm (GA) as a wrapper can effectively select an optimal wavelength subset, enhancing model accuracy and reducing overfitting [34].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Research Reagent Solutions and Materials

Item	Function/Application
Handheld NIR Spectrometer	Portable device for on-site spectral data collection (e.g., Felix Instruments F-750 Produce Quality Meter) [18].
Digital Refractometer	Primary reference method for determining Total Soluble Solids (TSS) in °Brix.
Laboratory Oven	Primary reference method for determining Dry Matter (DM) content via moisture evaporation.
Reference Standards	Certified materials for instrument performance validation and wavelength calibration [35].
Pre-calibrations	Digital prediction models for specific applications (e.g., mango DM/TSS) that allow for immediate analysis, though lab-specific validation is required [36] [37].
Chemometric Software	Software for spectral data preprocessing, model development (e.g., PLS), and validation (e.g., The Unscrambler, CAMO).
CR-1-31-B	CR-1-31-B, MF:C28H29NO8, MW:507.5 g/mol
AR-C102222	AR-C102222, MF:C19H16F2N6O, MW:382.4 g/mol

In the development of a handheld Near-Infrared (NIR) method for mango maturity testing, the acquisition of spectral data is only the first step. Raw NIR spectra contain not only information about chemical properties but also unwanted signal variations caused by light scattering, path length differences, instrument noise, and sample physical properties. Spectral preprocessing is therefore an essential procedure to remove these non-chemical influences and enhance the spectral features related to mango quality attributes such as total soluble solids (TSS), pH, dry matter content, and firmness [1]. For researchers and scientists developing robust analytical methods, proper preprocessing directly impacts model accuracy, robustness, and predictive performance. This application note details three fundamental preprocessing techniques—Standard Normal Variate (SNV), Savitzky-Golay Smoothing, and Derivatives—within the specific context of mango maturity assessment using handheld NIR spectrometers.

Standard Normal Variate (SNV)

SNV is a mathematical transformation designed to eliminate scatter effects and correct for path length differences in diffuse reflectance spectroscopy. It operates on each individual spectrum by centering and scaling the data.

• Purpose and Mechanism: SNV processes each spectrum by subtracting its mean and then dividing by its standard deviation. This procedure effectively removes multiplicative interferences of scatter and particle size, making spectra more comparable and enhancing chemical information [5] [38].
• Application in Mango Research: SNV has proven highly effective in mango quality prediction. It has been used successfully to preprocess spectra for predicting pH and Total Soluble Solids (TSS), often in combination with other techniques. Furthermore, SNV has been applied as one of several spectral transformation operators to significantly improve the accuracy of classification models for determining mango maturity indices, achieving accuracies as high as 95.7% when combined with fuzzy logic [13]. Research on the 'Tainong', 'Guifei', and 'Jinhuang' mango varieties also utilized SNV to preprocess spectra for models predicting fruit hardness, pH, SSC, and dry matter content [38].

Savitzky-Golay Smoothing

Savitzky-Golay smoothing is a digital filter that can be used to smooth data and calculate derivatives in a single step. It works by fitting a low-degree polynomial to successive subsets of adjacent data points.

• Purpose and Mechanism: This method reduces high-frequency random noise without significantly distorting the signal's original shape. The two key parameters are the window size (the number of points in the subset, which must be odd and greater than the polynomial order) and the polynomial order. A common challenge is selecting an appropriate window width: too narrow a window is ineffective against noise, while too wide a window can over-smooth the data and wash out important spectral features [39].
• Application in Mango Research: Savitzky-Golay smoothing is a staple in mango NIR analysis. It has been directly applied to smooth spectral data for predicting mango pH and TSS, forming a crucial step before building prediction models. The method is also frequently used as the computational basis for calculating first and second derivatives of spectra, which help resolve overlapping peaks and remove baseline offsets [6] [39].

Spectral Derivatives

Derivative spectroscopy involves computing the first or second derivative of spectral data with respect to wavelength.

• Purpose and Mechanism:
  • First Derivative: Removes additive baseline offsets and enhances the resolution of overlapping peaks by highlighting regions of steepest slope in the original spectrum.
  • Second Derivative: Eliminates both additive and multiplicative baseline effects and accentuates sharper, narrower spectral features, often corresponding to specific chemical bonds. A significant advantage of derivatives is their insensitivity to unknown tissue and fibre contact coupling coefficients, which increases the robustness of the method for clinical and industrial applications [40].
• Application in Mango Research: Derivatives are powerful for enhancing subtle spectral features related to mango constituents. The second derivative, in particular, has been used to remove spectral baselines and DC offsets. In mango maturity studies, the Savitzky-Golay algorithm is the standard method for computing these derivatives, as it performs smoothing and derivation simultaneously, mitigating the noise amplification that inherently comes with differentiation [40] [39].

Table 1: Summary of Key Preprocessing Techniques and Their Roles in Mango NIR Analysis

Technique	Primary Function	Key Parameters	Effect on Spectral Data	Typical Use Case in Mango Analysis
SNV	Scatter correction & path length normalization	None (applied per spectrum)	Centers and scales each spectrum	Correcting for differences in fruit size and surface texture [5] [38]
Savitzky-Golay Smoothing	Noise reduction & signal enhancement	Window width, Polynomial order	Suppresses high-frequency noise	Preparing raw spectra before derivative analysis or model building [6] [39]
First Derivative	Remove baseline offsets & enhance resolution	(Via Savitzky-Golay)	Highlights slopes of original peaks	Resolving overlapping sugar and water absorption bands [40] [39]
Second Derivative	Remove baselines & accentuate sharp features	(Via Savitzky-Golay)	Highlights shoulders and sharp peaks	Identifying specific chemical markers linked to maturity [40]

Experimental Data and Performance

The efficacy of preprocessing techniques is validated through their performance in quantitative and classification models for mango quality. The following table compiles results from recent studies, demonstrating how these methods contribute to accurate non-destructive assessment.

Table 2: Performance of Preprocessing Techniques in Mango Maturity and Quality Prediction Models

Mango Variety	Quality Parameter	Preprocessing Technique(s)	Model Type	Performance Results	Citation
Multiple Varieties	Variety Identification	RAW, MC, SNV, FD, SD + LDA-SVM	Multivariate Classification	100% (Training) & 97.44% (Prediction) Accuracy	[5]
Arumanis	Maturity Index (5 classes)	PLS with Fuzzy Logic (Indirect)	Classification	95.7% Accuracy	[13]
Gadung Klonal 21	pH	Savitzky-Golay Smoothing + MLPR	Regression (MLPR)	High Accuracy (MAPE <10%)	[6]
Gadung Klonal 21	TSS	Savitzky-Golay Smoothing + MLPR	Regression (MLPR)	High Accuracy (MAPE <10%)	[6]
Nam Dok Mai	TSS	Baseline Offset + Moving Average Smoothing	PLS Regression	R²cal=0.80, R²pred=0.74, RMSEP=0.765%	[41]
Tainong, Guifei, Jinhuang	Maturity Grade	MSC, SNV, SG Smoothing	Non-destructive Detection Model	81-90% Classification Accuracy	[38]

Detailed Experimental Protocols

Protocol 1: Preprocessing Workflow for Mango TSS and pH Prediction

This protocol is adapted from studies that successfully predicted pH and TSS in intact mangoes using a handheld NIR spectrometer [5] [6].

1. Sample Preparation and Spectral Acquisition:

• Collect mango samples (e.g., 186 fruits) across the desired maturity stages.
• Use a handheld NIR spectrometer (e.g., wavelength range 740-1070 nm).
• Scan each intact fruit, ensuring the measurement position is consistent (e.g., the cheek of the fruit). Acquire multiple scans per fruit if necessary and average them to obtain a representative spectrum.

2. Data Preprocessing:

• Optional Smoothing: Apply Savitzky-Golay smoothing to reduce high-frequency noise. Typical starting parameters are a 2nd-order polynomial and a window size of 5-11 points [39].
• Scatter Correction: Apply SNV to correct for light scattering effects caused by variations in fruit surface texture and size [5].
• Derivatization (if needed): Compute the first or second derivative using the Savitzky-Golay method to resolve overlapping peaks and remove baseline effects. The same parameters for smoothing apply, with the deriv parameter set to 1 or 2 [39].

3. Model Development and Validation:

• Split the preprocessed spectra and corresponding laboratory-measured TSS/pH values into calibration and validation sets (e.g., 75%/25%).
• Build a quantitative model, such as Partial Least Squares (PLS) or Multi-predictor Local Polynomial Regression (MLPR), using the calibration set.
• Validate the model using the prediction set, reporting key metrics like R², RMSEP, and MAPE.

Protocol 2: System Setup for Online Mango Grading

This protocol outlines the implementation of a conveyor-based system for online grading of mangoes, as validated in research on the 'Nam Dok Mai' variety [41].

1. Hardware Configuration:

• Spectrometer: Utilize a fiber optic Vis-NIR spectrometer (e.g., 400-1000 nm range) with a CCD detector array.
• Lighting: Employ a stabilized halogen light source coupled with an integrating sphere for consistent, diffuse illumination.
• Conveyor System: Set up a belt conveyor with a fixed speed (e.g., 0.1 m/s).
• Triggering Mechanism: Install a proximity sensor to detect incoming fruit and trigger the spectrometer automatically.
• Enclosure: House the spectrometer and light source in a dark enclosure to eliminate interference from ambient light.

2. Software and Data Processing:

• Spectral Acquisition: Configure the spectrometer's integration time to avoid signal saturation (e.g., ~3.7 ms).
• Preprocessing in Real-time: Implement the following preprocessing steps algorithmically for each acquired spectrum:
  • Wavelength Trimming: Remove noisy regions at the spectral extremes (e.g., below 400 nm and above 1000 nm).
  • Smoothing: Apply a moving average or Savitzky-Golay filter.
  • Baseline Correction: Use a baseline offset correction to minimize drift.
• Model Deployment: Integrate a pre-calibrated PLS regression model to convert the preprocessed spectrum into a TSS value in real-time.

3. Grading and Sorting:

• The system assigns a maturity grade or TSS value to each fruit.
• This information is sent to a sorting mechanism (e.g., pneumatic pushers) to direct mangoes to different bins based on quality.

The Scientist's Toolkit

Table 3: Essential Research Reagent Solutions and Materials for Handheld NIR Maturity Testing

Item Name	Function/Application	Specification Notes
Portable NIR Spectrometer	Core device for spectral data acquisition.	Wavelength range 740-1070 nm [5] or 1350-2500 nm [13]; integrated with smartphone or Raspberry Pi.
Reference Standards	Instrument calibration and validation.	White reference (e.g., Teflon, Barium Sulfate) for reflectance calibration; dark reference for noise correction [41].
Digital Refractometer	Primary method for TSS (°Brix) measurement.	Used to create reference values for NIR calibration model [5] [41].
pH Meter	Primary method for acidity measurement.	Used to create reference values for NIR calibration model [5] [6].
Chemometrics Software	Data preprocessing and model development.	Software packages (e.g., Metrohm Vision Air, Unscrambler, Python with Scipy) for applying SNV, Savitzky-Golay, and building PLS models [39] [42].
Firmness Tester	Primary method for destructive firmness analysis.	Provides reference data for correlating spectral data with mechanical texture [13] [38].
Maraviroc	Maraviroc | CCR5 Antagonist For Research	Maraviroc is a potent CCR5 antagonist for HIV research. For Research Use Only. Not for human or veterinary diagnostic or therapeutic use.
IWR-1	IWR-1, MF:C25H19N3O3, MW:409.4 g/mol	Chemical Reagent

The integration of SNV, Savitzky-Golay smoothing, and derivative preprocessing forms the foundational bedrock for extracting meaningful chemical information from NIR spectra in mango maturity research. As demonstrated across multiple studies, the careful application of these techniques directly enables the high levels of accuracy required for non-destructive prediction of key quality parameters like TSS and pH. For scientists and drug development professionals, mastering these protocols is not merely a data preparation step but a critical component in developing robust, transferable, and reliable handheld NIR methods that can transform quality control and supply chain management for perishable commodities like mangoes.

The application of handheld Near-Infrared (NIR) spectroscopy for non-destructive mango maturity testing represents a significant advancement in agricultural technology. This approach relies on chemometrics—the application of mathematical and statistical methods to extract meaningful information from chemical data—to build predictive models that correlate spectral signatures with internal fruit quality parameters such as Total Soluble Solids (TSS/Brix), firmness, and dry matter content. These models enable rapid, on-site quality assessment without destroying the fruit, providing valuable insights for determining optimal harvest times and post-harvest management. The successful implementation of this technology depends critically on selecting appropriate modeling algorithms that can handle the high-dimensional, collinear nature of spectroscopic data while accounting for complex non-linear relationships that often exist between spectral features and maturity indicators. This application note provides a comprehensive overview of four fundamental modeling approaches—Partial Least Squares Regression (PLSR), interval PLS (iPLS), Support Vector Machines (SVM), and Artificial Neural Networks (ANN)—within the specific context of handheld NIR method development for mango maturity testing research.

Core Algorithm Theory and Comparative Performance

Partial Least Squares Regression (PLSR)

PLSR is a cornerstone multivariate regression technique particularly suited for spectroscopic data analysis. It projects both the predictor variables (spectral data) and response variables (maturity parameters) to a new, lower-dimensional space of latent variables (LVs). These LVs are constructed to maximize the covariance between the spectral data and the reference maturity values, effectively filtering out noise and concentrating the predictive information into a few components. The fundamental strength of PLSR lies in its ability to handle datasets where the number of spectral variables far exceeds the number of samples and where these variables are highly collinear—precisely the characteristics of NIR spectra.

In practical application for mango maturity testing, a PLSR model is built using spectra from a calibration set of mango samples with laboratory-measured reference values for Brix or other maturity indicators. Once calibrated, the model can predict these maturity parameters from new spectral measurements of unknown mango samples. Research has demonstrated the effectiveness of this approach, with one study achieving Root Mean Square Error of Prediction (RMSEP) values of 0.74% and 1.27% for conversion percentage in biodiesel reactions monitored via NIR, illustrating the precision achievable with PLSR on similar data structures [43]. For mango-specific applications, studies have utilized PLSR in conjunction with preprocessing techniques like Standard Normal Variate (SNV) and Savitzky-Golay smoothing to establish reliable prediction models for sugar content [44].

Interval Partial Least Squares (iPLS)

iPLS is an extension of the standard PLSR algorithm that incorporates a variable selection strategy to improve model performance and interpretability. The core principle of iPLS involves dividing the full spectrum into multiple smaller intervals (windows) and then building local PLSR models for each individual interval or combination of intervals. The algorithm systematically evaluates these local models to identify the most informative spectral regions for predicting the response variable.

Two primary operational modes exist for iPLS:

• Forward iPLS: Begins with an empty set of selected intervals. In each iteration, it tests every unselected interval in combination with already selected ones, permanently adding the interval that yields the greatest improvement in model performance (e.g., lowest Root Mean Square Error of Cross-Validation, RMSECV) [45] [46].
• Backward iPLS: Starts with the full spectrum and iteratively removes the least informative intervals whose exclusion most improves the RMSECV [46].

The key advantage of iPLS in mango maturity testing is its ability to identify specific spectral regions most relevant to maturity prediction, which often correspond to known chemical bond vibrations (e.g., O-H and C-H stretches associated with sugars and water). This not only simplifies the model but can also enhance predictive accuracy by eliminating uninformative or noisy spectral regions. A key consideration is that the step-wise approach might occasionally miss the optimal combination of intervals; using larger interval windows and running the algorithm until no further improvement is observed can mitigate this risk [46].

Support Vector Machines (SVM) and Least-Squares SVM (LS-SVM)

SVM is a machine learning technique based on statistical learning theory that can perform both linear and non-linear regression. For non-linear cases, SVM utilizes kernel functions to project the original input data into a higher-dimensional feature space where a linear regression model can be constructed. A particularly efficient variant is Least-Squares SVM (LS-SVM), which employs a linear set of equations instead of quadratic programming, simplifying the computational process.

In the context of analytical chemistry and spectroscopy, SVM-based methods have proven to be powerful alternatives to neural networks. A comprehensive comparative study on NIR spectroscopic data for analyzing fuel and petroleum properties found that SVR and LS-SVM were comparable to ANNs in terms of prediction accuracy but offered greater robustness, making them particularly suitable for practical industrial applications, especially for complicated, highly non-linear systems [47]. This robustness is highly desirable for mango maturity models, which must contend with biological variability and potential non-linear relationships between spectra and maturity indices. Furthermore, studies have shown that SVM can be successfully optimized using nature-inspired algorithms like Moth Flame Optimization (MFO) for agricultural classification tasks, achieving high accuracy (e.g., >82%) in fruit ripeness classification [48].

Artificial Neural Networks (ANN)

ANNs are a class of flexible, non-linear models inspired by biological neural networks. They consist of interconnected layers of nodes (neurons)—an input layer, one or more hidden layers, and an output layer. Each connection has an associated weight that is adjusted during the training process to minimize the difference between the network's predictions and the actual reference values. This architecture allows ANNs to learn complex, non-linear mappings between inputs (spectral data) and outputs (maturity parameters).

Research consistently demonstrates the strong performance of ANNs in fruit quality assessment. One study on detecting adulteration in apple juice concentrate using spectroscopy found that ANNs outperformed other methods, achieving a high correct classification rate of 93.75% for identifying adulterant types [49]. Similarly, in textural analysis of intact grape berries, ANNs provided better prediction ability for parameters like hardness and chewiness compared to PLS models, particularly after the elimination of uninformative spectral ranges [50]. For mango maturity specifically, a study utilizing a BP neural network (a common type of ANN) combined with a simulated annealing algorithm achieved exceptional performance in predicting Brix, with a correlation coefficient of 0.9854 and a root-mean-square error of 0.0431 [44]. The main consideration with ANNs is their complexity in setup, training, and parameter estimation compared to linear methods.

Quantitative Performance Comparison

Table 1: Comparative Performance of Modeling Algorithms on Spectroscopic Data

Algorithm	Application Context	Reported Performance Metric	Value	Citation
ANN	Apple Juice Adulteration	Correct Classification Rate	93.75%	[49]
ANN	Mango Brix Prediction	Correlation Coefficient (after optimization)	0.9854	[44]
SVM	Apple Juice Adulteration	Correct Classification Rate	91.67%	[49]
LS-SVM	Fuel/Petroleum Properties	Accuracy vs. ANN	Comparable	[47]
PLSR	Biodiesel Reaction Monitoring	RMSEP (for conversion)	0.74% - 1.27%	[43]
iPLS	Beer Extract Analysis	Reduction in Error vs. Full-Spectrum PLS	4-fold reduction	[51]

Experimental Protocol for Mango Maturity Model Development

Sample Preparation and Spectral Acquisition

• Material Selection: Acquire a sufficient number of mango samples (e.g., ≥95 [44]) representing different varieties, orchards, and, crucially, a wide range of maturity stages. Samples should be free of surface defects.
• Instrument Pre-treatment: Power on the handheld NIR spectrometer and halogen light source, allowing a sufficient pre-heating time (e.g., 30 minutes) for signal stabilization. Perform instrument calibration (dark and reference scans) according to the manufacturer's protocol.
• Spectral Data Collection: For each mango, clean the surface and take spectral measurements at multiple consistent positions (e.g., equator, apex, base [44]). Configure acquisition parameters (e.g., integration time, number of scans) to achieve a high signal-to-noise ratio. For each sample, average several scans to obtain a representative spectrum.
• Reference Value Determination: Immediately following spectral acquisition, destructively measure the primary maturity indices for each mango sample. This typically involves:
  • Total Soluble Solids (TSS/Brix): Using a digital refractometer on the juice from the same scanned region [50] [44].
  • Firmness: Using a texture analyzer with a penetrometer probe.
  • Dry Matter Content: Using a standard oven-drying method.

Data Preprocessing and Model Development Workflow

The following diagram outlines the core workflow from spectral acquisition to a validated predictive model.

• Spectral Preprocessing: Process raw spectra to minimize the impact of light scattering, baseline drift, and noise. Common techniques, often applied sequentially, include:
  • Standard Normal Variate (SNV) or Multiplicative Scatter Correction (MSC): To reduce scattering effects [50] [44].
  • Savitzky-Golay Smoothing and Derivatives: To enhance spectral features and remove noise (e.g., 1st or 2nd derivative) [50].
  • Mean Centering: To improve model stability by centering the data around the mean [50].
• Dataset Splitting: Randomly divide the preprocessed dataset into a calibration/training set (typically 70-80% of samples) for model building and a validation/test set (the remaining 20-30%) for evaluating the final model's predictive performance on unseen data [50].
• Model Training and Optimization:
  • For PLSR/iPLS: Use cross-validation on the calibration set to determine the optimal number of Latent Variables (LVs) or the most predictive spectral intervals. The goal is to minimize RMSECV while avoiding overfitting [45] [46].
  • For ANN: The network architecture (number of hidden layers and neurons) and training parameters (learning rate, epochs) must be optimized. Techniques like backpropagation are used for training, and optimization algorithms like simulated annealing can be employed to avoid local minima [44].
  • For SVM/SVR: The hyperparameters (e.g., regularization parameter C, kernel parameters like gamma) are critical. Use cross-validation and optimization algorithms (e.g., Moth Flame Optimization, grid search) to find the optimal parameter set [48].
• Model Validation: The final model must be rigorously validated using the independent test set that was not used in any step of the training or optimization process. Report standard performance metrics such as:
  • Coefficient of Determination (R²)
  • Root Mean Square Error of Prediction (RMSEP)
  • Ratio of Performance to Deviation (RPD)

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Key Materials and Software for Handheld NIR Maturity Model Development

Item Name	Specification / Example	Primary Function in Research
Handheld NIR Spectrometer	Ocean Optics NIR-Quest+ (900-1700 nm or 1300-2300 nm range) [44]	Acquires diffuse reflectance spectra from the intact mango surface.
Halogen Light Source	HL-2000-LL series [44]	Provides stable, broadband NIR illumination for consistent spectral measurements.
Digital Refractometer	Schmidt + Haensch DHR 95 [44]	Provides precise reference measurements of Total Soluble Solids (TSS/Brix) for model calibration.
Texture Analyzer	Zwick/Roell XforceP [50]	Measures reference values for fruit firmness and other textural properties.
Software for Chemometrics	R (with mdatools [45] [50]), MATLAB [44], PLS_Toolbox [46]	Provides the computational environment for data preprocessing, model development, and validation.
Standard Reference Tiles	Certified white reference (e.g., Spectralon)	Used for regular instrument calibration to ensure measurement consistency over time.
CRT0066101	CRT0066101, MF:C18H22N6O, MW:338.4 g/mol	Chemical Reagent

The development of robust handheld NIR models for mango maturity testing is a multi-step process that hinges on the informed selection and implementation of chemometric algorithms. PLSR offers a robust, interpretable linear baseline. iPLS enhances PLSR by identifying key spectral regions, improving both model performance and interpretability. For capturing complex non-linearities, SVM/LS-SVM provides a robust and accurate alternative, while ANN represents a powerful, flexible framework capable of modeling highly intricate relationships. The choice of algorithm depends on the specific maturity parameter, the scale of the project, and the required balance between model interpretability and predictive power. By following the detailed protocols and leveraging the comparative insights outlined in this document, researchers can effectively develop and validate non-destructive tools for mango quality assessment.

The non-destructive assessment of mango maturity using handheld Near-Infrared (NIR) spectroscopy is undergoing a significant methodological shift. Traditionally, this process has relied on indirect classification, where NIR spectra are used to predict specific maturity index values—such as Dry Matter (DM)- or Total Soluble Solids (TSS)—through regression algorithms. The final maturity classification (e.g., immature, mature, overripe) is then determined by applying hard thresholds to these predicted values [52]. A emerging paradigm, termed direct classification, challenges this two-step process. This new approach uses classification algorithms to assign a maturity class directly from the spectral data, bypassing the intermediate step of estimating a continuous maturity index [52]. This Application Note details this novel methodology, providing a comparative analysis and detailed protocols for its implementation within mango maturity testing research.

Comparative Analysis: Direct vs. Indirect Methodologies

The core difference between the two paradigms lies in the modeling objective and the final output. The table below summarizes their key characteristics and performance.

Table 1: Comparison of Traditional and Direct Maturity Classification Paradigms

Feature	Traditional Indirect Estimation	Novel Direct Classification
Core Approach	Regression to estimate a quantitative index (e.g., DM, TSS), followed by thresholding for classification [52].	Classification algorithms assign a maturity class directly from spectral data [52].
Primary Output	A predicted value (e.g., % DM, °Brix) [6] [4].	A class label (e.g., Underripe, Ripe, Overripe) [52] [53].
Key Advantage	Provides a continuous measurement of a chemical property.	Reported to achieve higher accuracy for classification tasks and is more suited for handheld devices with limited calibration scope [52].
Reported Performance	Indirect classification accuracy for mango maturity: ~88.2% [52].	Direct classification accuracy for mango maturity: ~95.8% [52]. Other studies using different spectroscopy methods report up to 100% classification accuracy [53].
Typical Algorithms	Partial Least Squares Regression (PLSR) [6] [4], Support Vector Machine Regression (SVMR) [6], Multi-Predictor Local Polynomial Regression (MLPR) [6].	Linear Discriminant Analysis (LDA) [52] [13], Support Vector Machine (SVM) [53], Random Forest [53].

Beyond the core methodology, a hybrid approach has been demonstrated to further enhance accuracy. This model employs an indirect classification with fuzzy logic. It first uses NIR spectra and PLSR models to predict multiple maturity parameters (e.g., acidity, firmness, starch). Subsequently, a fuzzy logic system interprets these predicted values to make a final maturity classification, achieving a high accuracy of 95.7% [13].

Experimental Protocols

Protocol for Direct Maturity Classification Using Handheld NIR

This protocol outlines the procedure for developing a direct maturity classification model for mangoes, based on the work of [52].

I. Hardware Setup and Spectral Acquisition

• Device Configuration: Utilize a handheld NIR spectrometer (e.g., a micro-spectrometer development kit covering 500-1050 nm or 1350-2500 nm) integrated with a computational unit (e.g., Intel Compute Stick, Raspberry Pi) and a controlled light source [52] [13].
• Sample Preparation: Harvest mango fruits at different known maturity stages (e.g., one week before, on, and one week after the estimated harvest date). The reference maturity should be established based on a standard metric like Dry Matter (DM) content determined through destructive testing [52].
• Data Collection: For each fruit, collect NIR absorbance spectra from multiple positions. Calibrate the sensor with a standard reference (e.g., Barium Sulfate calibrator) before measurement [52] [13]. The raw spectral data for each sample is stored with its reference maturity class label.

II. Data Preprocessing and Model Development

• Spectral Preprocessing: Process the raw spectral data to reduce noise and enhance features. Apply techniques such as the Savitzky-Golay (SG) second derivative smoothing [52]. Other effective transformations may include Standard Normal Variate (SNV) and Multiplicative Scatter Correction (MSC) [4] [13].
• Dimensionality Reduction: Use Principal Component Analysis (PCA) to reduce the high dimensionality of the spectral data. The first few principal components that capture the majority of the variance are retained as inputs for the classification model [52].
• Model Training: Train a classification algorithm, such as Linear Discriminant Analysis (LDA), using the principal components from the calibration dataset. The model learns to distinguish the spectral patterns associated with each predefined maturity class [52] [13].

III. Validation and Deployment

• Model Validation: Test the performance of the trained LDA model on a separate prediction set of mango samples that were not used in training. Evaluate the model based on its classification accuracy (percentage of correctly classified samples) [52].
• Device Integration: Deploy the validated model onto the handheld device's computational unit. The integrated system can now directly display the maturity class when a new NIR spectral measurement is taken [52].

Diagram 1: Direct classification workflow.

Protocol for Indirect Estimation with Fuzzy Logic

This protocol describes an advanced indirect method that uses fuzzy logic to improve classification accuracy, as demonstrated by [13].

I. Multi-Parameter Prediction Model Development

• Spectral Acquisition and Reference Analysis: Collect NIR spectra from mango samples as in Protocol 3.1. Subsequently, perform destructive tests on the same samples to measure key maturity parameters: Total Acidity (TA), Soluble Solids Content (SSC), Firmness, and Starch content [13].
• Regression Model Calibration: For each of the four maturity parameters, develop a separate PLSR model that correlates the preprocessed NIR spectra with the laboratory-measured reference values [13].

II. Fuzzy Logic System Implementation

• Fuzzification: Define fuzzy sets for each maturity parameter. For example, for SSC, create membership functions for linguistic variables like "Low," "Medium," and "High." The predicted values from the PLSR models are converted into degrees of membership (ranging from 0 to 1) to these fuzzy sets [13].
• Fuzzy Inference: Construct a set of IF-THEN rules that define the maturity class based on the combinations of the input parameters. For example: IF TA is High AND SSC is Low AND Firmness is High THEN Maturity_Index is 80%. The rules are applied to the fuzzified inputs, and the output of each rule is a fuzzy set [13].
• Defuzzification: Aggregate all the output fuzzy sets from the triggered rules and convert this aggregated result into a crisp output value, which is the final predicted maturity index (e.g., 80%, 85%, etc.) [13].

Diagram 2: Indirect estimation with fuzzy logic.

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 2: Key Materials and Reagents for Handheld NIR Maturity Assessment

Item	Specification/Function	Research Application
Handheld NIR Spectrometer	e.g., NeoSpectra Micro (1350-2500 nm) or similar micro-spectrometer development kits [52] [13].	The core sensor for acquiring spectral data from intact mango fruit in the field or lab.
Computational Unit	e.g., Raspberry Pi, Intel Compute Stick [52] [13].	Embedded computing for real-time spectral data processing, model execution, and result display.
Calibration Standard	e.g., Barium Sulfate (BaSOâ‚„) tile or disk [13].	A white reference with high, stable reflectance for calibrating the spectrometer before sample measurement to ensure data consistency.
Reference Analytics Lab Equipment	For destructive validation: pH meter, refractometer (for TSS/O Brix), texture analyzer (for firmness), oven (for Dry Matter), titration setup (for acidity) [6] [4] [13].	To establish the "ground truth" reference values for maturity indices, which are essential for training and validating both regression and classification models.
Data Analysis Software	e.g., Python (with scikit-learn, SciPy), R, Orange Data Mining, OriginPro [53] [13].	For spectral preprocessing, dimensionality reduction, and developing machine learning models (PLSR, LDA, SVM, Fuzzy Logic).

Optimizing Accuracy and Robustness: Overcoming Spectral Analysis Challenges

Near-Infrared (NIR) spectroscopy has established itself as a powerful, non-destructive analytical technique increasingly deployed for quality assessment of agricultural products, particularly mangoes. The commercial and consumer acceptance of this high-value fruit depends critically on internal quality attributes such as Soluble Solids Content (SSC), Titratable Acidity (TA), and Dry Matter Content (DMC). Conventional methods for assessing these parameters are destructive, labor-intensive, and impractical for large-scale postharvest handling. NIR spectroscopy offers a rapid, non-destructive alternative by measuring light interaction with molecular bonds (O-H, C-H, N-H) in the 750-2500 nm range, providing rich chemical and physical information about the sample. However, the raw spectral data captured by NIR instruments is never pristine; it invariable contains overwhelming background interference, light scattering effects, instrumental noise, and other unwanted variations that obscure the chemically relevant information.

This is where spectra preprocessing becomes indispensable. Preprocessing refers to the mathematical transformation of raw spectral data to remove or reduce these non-chemical artifacts, thereby enhancing the subsequent multivariate calibration or classification models. The choice of preprocessing method is one of the most critical factors determining the prediction accuracy and robustness of NIR models. For mango maturity testing using handheld devices, where conditions are less controlled than in laboratory environments, effective preprocessing is even more crucial for achieving reliable, transferable results. The journey from raw, noisy spectra to accurate predictive models navigates a complex maze of methodological choices, traditionally guided by manual trial-and-error but increasingly being charted through automated, intelligent strategies.

The Preprocessing Landscape: Methods and Their Impact

Common Preprocessing Techniques

A variety of preprocessing methods have been developed, each targeting specific types of spectral artifacts. The most prevalent techniques include:

• Multiplicative Scatter Correction (MSC): Corrects for additive and multiplicative scattering effects by aligning each spectrum to an ideal reference spectrum (often the mean spectrum of the dataset). This method has demonstrated particular effectiveness for mango quality parameters, showing superior performance for predicting TA and SSC in intact mango compared to other methods [54].
• Standard Normal Variate (SNV): Similar to MSC, SNV removes scattering effects by centering and scaling each individual spectrum, making it robust for samples with varying particle sizes.
• Detrending (DT): Removes baseline shifts or curvilinear trends from spectra, often applied after SNV to correct for wavelength-dependent scattering effects.
• Derivative Techniques (e.g., Savitzky-Golay): First and second derivatives help resolve overlapping peaks, remove baseline offsets, and enhance subtle spectral features. The second derivative is particularly effective for emphasizing absorption peaks while eliminating additive and multiplicative effects.
• Normalization: Scales spectra to a standard range (e.g., 0-1 or unit vector length) to compensate for variations in path length or sample concentration.

Quantitative Impact on Mango Quality Prediction

The selection of preprocessing method significantly influences model performance for different mango quality attributes. Research on 'Kent' mangoes has quantified this impact, revealing that while preprocessing generally improves prediction accuracy, the degree of improvement varies substantially by parameter and method.

Table 1: Performance of Different Preprocessing Methods for Mango Quality Prediction [54]

Quality Parameter	Preprocessing Method	R² Prediction	RMSE	RPD	RER
Titratable Acidity (TA)	MSC	0.72	-	1.9	-
TA	SNV	0.68	-	1.7	-
TA	Baseline (None)	0.65	-	1.6	-
Soluble Solids Content (SSC)	MSC	0.76	-	1.8	-
SSC	SNV	0.73	-	1.7	-
SSC	Baseline (None)	0.69	-	1.5	-

Although MSC emerged as the most effective method among those tested for both TA and SSC prediction, the achieved Ratio of Performance to Deviation (RPD) values of 1.9 and 1.8 respectively still indicate relatively poor model performance that requires further enhancement for real-world applications. RPD values above 2.0 are generally considered necessary for rough screening, while values above 3.0 indicate good predictive ability [54]. This performance gap highlights the need for more sophisticated preprocessing strategies, particularly for handheld devices operating in field conditions.

The Traditional Approach: Manual Trial-and-Error

The Conventional Workflow

Until recently, preprocessing method selection has been predominantly guided by manual trial-and-error approaches. This conventional pipeline involves sequentially testing different preprocessing methods, often in combination, and evaluating their performance through cross-validation and external test sets. Researchers typically rely on their expertise, knowledge of the data characteristics, and the specific analytical goals to narrow down the candidate methods.

A study on Arumanis mango maturity classification exemplified this approach, testing twelve different spectral transformation operators including clipping, scatter correction, smoothing, derivation, trimming, and resampling methods. The researchers manually evaluated various combinations to achieve optimal classification performance, ultimately reaching 91.43% accuracy for direct maturity classification using Linear Discriminant Analysis with optimal preprocessing [13].

Limitations of Manual Selection

The manual approach presents several significant challenges:

• Expertise-Dependent: Success heavily relies on the practitioner's experience and domain knowledge, making it inaccessible to non-specialists.
• Time-Consuming: Exhaustive testing of method combinations requires substantial computational time and resources.
• Subjectivity: Method selection may be influenced by personal preference or conventional practices rather than objective optimization.
• Local Optima: Researchers may settle on satisfactory rather than optimal combinations due to the combinatorial explosion of possibilities.
• Reproducibility Issues: The subjective nature of selection makes method justification and study replication challenging.

These limitations become particularly problematic for handheld NIR applications in mango maturity testing, where varying environmental conditions, fruit heterogeneity, and the need for rapid analysis demand robust, optimized preprocessing pipelines.

Emerging Paradigm: Automated Preprocessing Strategies

The AGoES Framework

To overcome the limitations of manual approaches, researchers have begun developing automated preprocessing strategies. A notable example is the Automatically Generating a pre-processing Strategy (AGoES) framework, which represents a paradigm shift in spectral data pretreatment [55].

AGoES operates as an ensemble preprocessing method where multiple machine learning algorithms (including PLSR, SVM, k-NN, Decision Trees, AdaBoost, and Gaussian Process Regression) are built on differently preprocessed data and combined through 5-fold cross-validation and grid search optimization. This approach systematically explores the preprocessing space without requiring manual intervention or extensive domain expertise. When applied to predict parameters in manure organic waste, AGoES combined with Support Vector Machines achieved impressive RPD values of 3.619 and 2.996 for predicting dry matter and ammonium nitrogen content, respectively – performance metrics that surpass what is typically achieved through manual optimization [55].

SFIOS and Other Intelligent Systems

While not explicitly detailed in the search results, systems like SFIOS (Smart Preprocessing and Integration of Spectral data) represent the next evolution in automated preprocessing. These intelligent systems typically incorporate:

• Multi-strategy evaluation: Simultaneous testing of multiple preprocessing pathways
• Performance-based selection: Automatic selection based on objective metrics (R², RMSE, RPD)
• Ensemble approaches: Combining predictions from multiple optimally preprocessed models
• Adaptive optimization: Dynamic adjustment of preprocessing parameters based on dataset characteristics

Such systems are particularly valuable for handheld NIR devices used in mango maturity testing, as they can automatically adapt to varying measurement conditions, fruit varieties, and quality parameters without requiring manual reoptimization.

Experimental Protocols for Preprocessing Evaluation

Protocol 1: Comparative Evaluation of Preprocessing Methods

Objective: To systematically evaluate the performance of different preprocessing methods for predicting mango maturity parameters using handheld NIR spectroscopy.

Materials and Equipment:

• Handheld NIR spectrometer (e.g., NeoSpectra Micro, 1350-2500 nm range)
• Mango samples across maturity stages (n ≥ 100 recommended)
• Reference analytical equipment for destructive validation (refractometer for SSC, titrator for TA, texture analyzer for firmness)
• Computing device with multivariate analysis software (Python with scikit-learn, MATLAB, or R)

Procedure:

• Sample Preparation: Select mangoes representing different maturity stages based on days after full bloom. Clean and label each fruit, marking measurement positions.
• Spectral Acquisition: Calibrate the NIR device using a white reference (e.g., Barium Sulfate). Collect spectra from predetermined positions on each mango fruit, ensuring consistent contact pressure and orientation [13].
• Reference Measurements: Immediately after spectral collection, perform destructive measurements of SSC, TA, firmness, and other maturity indicators at the same locations.
• Data Preprocessing: Apply the following preprocessing techniques to the raw spectra:
  • Multiplicative Scatter Correction (MSC)
  • Standard Normal Variate (SNV)
  • First and Second Derivatives (Savitzky-Golay, various window sizes)
  • Detrending
  • Normalization
  • Combined approaches (e.g., SNV + Detrending)
• Model Development: Divide data into calibration (70-80%) and validation (20-30%) sets. Develop PLS regression models for each preprocessing method predicting each maturity parameter.
• Performance Evaluation: Calculate R², RMSE, RPD, and RER for each model. Statistically compare performance across preprocessing methods.

Table 2: Essential Research Reagent Solutions and Materials

Item	Specifications	Function in Experiment
NeoSpectra Micro NIR Sensor	1350-2500 nm range, 16 nm resolution, SNR 2000:1	Spectral data acquisition from mango samples
Barium Sulfate (BaSOâ‚„) Calibrator	>99% purity, stable white reference	Instrument calibration and background measurement
Raspberry Pi Compute Module	Broadcom BCM2835 processor, 512MB RAM	Portable computing for data processing and model implementation
Python Programming Environment	NumPy, SciPy, scikit-learn libraries	Implementation of preprocessing algorithms and machine learning models
Reference Chemical Standards	Sucrose, malic acid, starch solutions	Calibration of destructive measurement instruments

Protocol 2: Implementation of Automated Preprocessing (AGoES-type Approach)

Objective: To implement an automated preprocessing strategy for optimizing mango maturity prediction models.

Materials and Equipment: Same as Protocol 1, with emphasis on computational resources.

Procedure:

• Data Collection: Follow steps 1-3 from Protocol 1 to acquire paired spectral and reference data.
• Preprocessing Ensemble Definition: Define a set of preprocessing techniques and combinations to be evaluated, including:
  • Scatter corrections (MSC, SNV)
  • Derivatives (1st, 2nd with varying parameters)
  • Smoothing filters
  • Wavelength selection methods
• Machine Learning Pipeline: Implement multiple machine learning algorithms (PLS, SVM, k-NN, etc.) in an ensemble framework.
• Cross-Validation Optimization: Use 5-fold cross-validation with grid search to optimize both preprocessing parameters and model hyperparameters simultaneously.
• Model Integration: Combine predictions from the best-performing preprocessing-model combinations using stacking or weighted averaging.
• Validation: Evaluate the final ensemble model on an independent test set not used during optimization.

Implementation Workflow: From Manual to Automated Preprocessing

The following workflow diagram illustrates the transition from traditional manual preprocessing to automated strategies:

Advanced Applications in Mango Maturity Testing

Integration with Fuzzy Logic for Classification

Beyond regression models for specific chemical parameters, preprocessing plays a vital role in maturity classification. Research on Arumanis mango has demonstrated that combining optimal preprocessing with fuzzy logic classification achieves superior accuracy (95.7%) compared to direct classification approaches (91.43%) [13]. This approach uses preprocessed spectra to predict multiple maturity indicators (TA, SSC, firmness, starch), then applies fuzzy logic rules to integrate these predictions into a comprehensive maturity classification.

Portable Device Implementation

For practical mango maturity testing, preprocessing algorithms must be implementable on portable devices with limited computational resources. Successful implementations have utilized Raspberry Pi modules with Python programming to execute preprocessing transformations in real-time [13]. This enables field-deployable systems that can provide immediate maturity assessments without destructive sampling.

The evolution from manual trial-and-error to automated preprocessing strategies represents a significant advancement in NIR spectroscopy for mango maturity testing. While traditional methods have provided valuable insights and established foundational preprocessing techniques, they face limitations in reproducibility, optimization, and accessibility. Automated approaches like AGoES and SFIOS offer systematic, objective, and optimized preprocessing that enhances prediction accuracy while reducing dependency on specialist expertise.

For handheld NIR applications in mango maturity testing, where variability in measurement conditions and fruit characteristics presents particular challenges, these automated strategies show exceptional promise. Future developments will likely incorporate more sophisticated machine learning approaches, adaptive preprocessing that responds to real-time quality assessments, and integration with other sensing modalities like hyperspectral imaging. As these technologies mature, they will increasingly enable reliable, non-destructive quality assessment throughout the mango supply chain, from harvest to consumer, minimizing waste and ensuring optimal fruit quality.

The development of handheld NIR spectroscopy for mango maturity assessment represents a significant advancement in non-destructive fruit quality evaluation. A critical component in optimizing these portable systems is effective wavelength selection, which reduces instrument complexity, decreases computational requirements, and enhances model accuracy by focusing on the most informative spectral regions. For mango maturity traits such as Dry Matter (DM), Total Soluble Solids (TSS), and internal defects, specific wavelength ranges have been identified as particularly relevant, moving beyond traditional full-spectrum analysis to targeted, efficient monitoring. This protocol details the application of wavelength selection techniques, including interval Partial Least Squares (iPLSR) and other advanced methods, within the context of handheld NIR device development for mango maturity testing.

Key Spectral Regions and Quantitative Performance

Research has consistently identified specific spectral ranges that carry the most relevant information for assessing mango maturity and internal quality. The tables below summarize the key wavelength ranges and the performance of selection methods.

Table 1: Key Wavelength Ranges for Mango Quality Assessment

Quality Parameter	Optimal Wavelength Range	Significance	Citation
Internal Defects	702.72 nm - 752.34 nm	Most effective for spongy tissue detection	[56]
General Defects	673 nm - 1100 nm	Efficient lower range for internal defect classification	[56]
Maturity Estimation	400 nm - 1100 nm	Standard range for handheld NIR maturity meters	[28]
Acidity & Sweetness	Full spectrum modeling	pH and TSS prediction using multi-predictor models	[6]

Table 2: Performance of Wavelength Selection and Modeling Techniques

Technique	Application	Reported Performance	Citation
Fisher's Criterion	Wavelength selection for defect detection	84.5% classification accuracy	[56]
Direct Maturity Classification (KNN)	On-tree maturity state (mature/immature)	88.2% accuracy	[28]
Indirect Maturity Estimation	DM thresholding on predicted value	55.9% accuracy	[28]
Multi-predictor LPR	pH prediction in intact mangoes	MAPE < 10%	[6]
Convolutional Neural Network	Surface defect detection from images	98% accuracy	[57]

Experimental Protocols for Wavelength Selection

Protocol A: Fisher's Criterion for Defect Detection

This protocol is adapted from studies on spongy tissue detection in mangoes using NIR spectroscopy [56].

1. Sample Preparation:

• Select mango samples representing both healthy and internally defective fruits (e.g., with spongy tissue).
• Ensure sample surface is clean and dry for spectral acquisition.
• Label and randomize samples to prevent measurement bias.

2. Spectral Data Acquisition:

• Use an NIR spectrometer covering the range of 673 nm to 1900 nm.
• Configure the instrument for interactance or reflectance mode, depending on the target defect depth.
• For each fruit, collect multiple spectra from different positions and average them to account for fruit heterogeneity.
• Maintain consistent light source intensity and detector integration time across all measurements.

3. Data Preprocessing:

• Apply Multiplicative Scatter Correction (MSC) or Savitzky-Golay smoothing to reduce scattering effects and high-frequency noise [4].
• Normalize spectra to account for intensity variations between samples.
• Perform outlier detection to remove anomalous spectra from the dataset.

4. Feature Selection using Fisher's Criterion:

• Calculate Fisher's score for each wavelength across the entire spectrum.
• The score measures the ratio of between-class variance to within-class variance (defective vs. healthy).
• Rank wavelengths based on their Fisher scores.
• Select the top-performing wavelengths (e.g., 702.72 nm - 752.34 nm) showing the highest discriminative power [56].

5. Model Building and Validation:

• Build a classification model (e.g., Euclidean distance-based classifier) using only the selected wavelengths.
• Validate model performance using cross-validation or an independent test set.
• Report classification accuracy, precision, and recall metrics.

Protocol B: iPLSR for Maturity Trait Prediction

This protocol outlines the use of interval-based methods for predicting continuous maturity traits like DM and TSS.

1. Spectral Collection and Preprocessing:

• Collect spectra from mangoes at varying maturity stages using a handheld NIR device (400-1100 nm) [28].
• Apply Baseline Linear Correction (BLC) to remove linear baseline drift [4].
• Use Standard Normal Variate (SNV) transformation to minimize path-length effects.

2. Interval Selection and Modeling:

• Divide the full spectrum into smaller, equally-sized intervals.
• For each interval, develop a PLSR model predicting the target trait (e.g., DM content).
• Evaluate model performance using Root Mean Square Error of Cross-Validation (RMSECV) or similar metrics.
• Identify intervals yielding the lowest prediction error.
• Combine optimal intervals or select the single best interval for final model development.

3. Model Optimization:

• Optimize the number of Latent Variables (LVs) for each interval to prevent overfitting.
• Validate the final model on an external test set not used during wavelength selection.
• Compare performance against full-spectrum PLSR to confirm improvement.

Protocol C: Direct Classification for Maturity State Assessment

This protocol describes a direct classification approach for maturity state (mature/immature) classification, which can be more effective than regression-based methods for handheld applications [28].

1. Reference Method and Labeling:

• Determine mango maturity using a standard destructive method (e.g., DM content via oven drying).
• Assign maturity labels (mature/immature) based on established DM thresholds for the specific mango variety.

2. Spectral Processing:

• Acquire interactance spectra from on-tree mango fruits using a handheld NIR spectrometer.
• Perform noise reduction using moving average filters or wavelet transformation.
• Employ Principal Component Analysis (PCA) for dimensionality reduction and noise filtering [58].

3. Classifier Training:

• Train a K-Nearest Neighbors (KNN) classifier or Support Vector Machine (SVM) directly on the preprocessed spectral data.
• Tune hyperparameters (e.g., k for KNN, kernel for SVM) via cross-validation.
• Embed the finalized model into the handheld device's software for real-time prediction.

Workflow Visualization

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for Handheld NIR Maturity Assessment

Item	Specification/Function	Application Context
Handheld NIR Spectrometer	Spectral range: 400-1100 nm; Embedded processing	On-tree, in-field maturity screening [28]
Benchtop NIR Spectrometer	Higher resolution (e.g., 1557 points from 1000-2500 nm); Reference method	Laboratory model development and validation [4]
Reference Analytical Tools	Refractometer (TSS), Oven (DM), pH meter	Establishing ground truth for calibration [6]
Spectral Preprocessing Algorithms	MSC, SNV, Savitzky-Golay derivatives, BLC	Correcting light scattering and noise in raw spectra [4]
Chemometric Software	MATLAB, Python (scikit-learn, PyPLS), R	Developing PLSR, iPLSR, and classification models [58]
Standard Mango Samples	Representing different maturity stages and defect conditions	Creating robust calibration models that generalize well

Addressing Model Overfitting and Ensuring Generalizability Across Seasons and Varieties

The adoption of handheld Near-Infrared (NIR) spectroscopy for mango maturity assessment represents a significant advancement in non-destructive fruit quality evaluation. However, the development of robust calibration models that maintain accuracy across different seasons, cultivation practices, and mango varieties presents considerable challenges, primarily due to model overfitting. Overfit models perform well on the original calibration dataset but fail to generalize to new samples, severely limiting their practical application in commercial and research settings [59] [2].

Model overfitting typically occurs when models become excessively complex and tailored to the specific variations in the training set, including seasonal weather patterns, soil conditions, and variety-specific characteristics. For mango maturity models, this manifests as inaccurate predictions of key maturity parameters such as dry matter (DM) and soluble solids content (SSC) when applied to fruit from different growing seasons or genetic backgrounds [2]. Addressing these challenges requires systematic approaches to model development that incorporate diverse data sources and implement robust validation protocols.

This application note provides detailed methodologies for developing generalizable NIR calibration models, complete with experimental protocols and practical strategies to enhance model robustness for mango maturity assessment across diverse conditions.

Quantitative Performance Metrics in Mango Maturity Models

The table below summarizes key performance metrics from recent studies developing NIR calibration models for mango maturity parameters, highlighting approaches that address generalizability:

Table 1: Performance Metrics of NIR Calibration Models for Mango Maturity Assessment

Mango Variety	Parameter	Preprocessing Method	Model Type	R²CV/R²P	RMSECV/RMSEP	Generalizability Approach	Citation
Palmer	DM	1st derivative Savitzky-Golay	PLSR	0.84	8.81 g kg⁻¹	Multi-season sampling	[59] [2]
Palmer	SSC	1st derivative Savitzky-Golay	PLSR	0.87	1.39%	Multi-season sampling	[59] [2]
Arumanis	Maturity Index	Multiple preprocessing combinations	PLS with Fuzzy Logic	95.7% accuracy	N/A	Multi-parameter approach (SSC, TA, firmness, starch)	[13]
Arumanis	Maturity Index	12 spectral transformations	LDA	91.43% accuracy	N/A	Spectral data augmentation	[13]
Mixed Varieties	DM	N/A	PLSR	0.94	0.68%	Incorporation of orchard variability	[59]

The data demonstrates that while Dry Matter (DM) shows slightly lower R² values than Soluble Solids Content (SSC), it often serves as a more reliable maturity indicator, particularly for Palmer mangoes, as it shows consistent increases earlier in fruit development [2]. The superior accuracy achieved through fuzzy logic classification (95.7%) highlights the value of combining multiple parameters and advanced computational approaches for maturity classification rather than relying on single-parameter regression models [13].

Experimental Protocols for Generalizable Model Development

Comprehensive Sample Selection and Data Acquisition

Objective: To capture the natural variability encountered in commercial mango production systems to build robust calibration models.

Materials:

• Handheld NIR spectrometer (e.g., Felix Instruments F-750, Neospectra Micro)
• Reference laboratory equipment for destructive analysis
• Labeling materials and data recording system
• Standardized calibration standards for NIR instrument

Procedure:

• Strategic Sample Collection:
  • Select fruits from multiple orchards with varying soil types, microclimates, and management practices
  • Include fruit from different canopy positions (inner vs. outer, upper vs. lower)
  • Collect samples across multiple growing seasons (minimum 2-3 seasons)
  • For variety-specific models, include multiple sub-varieties or cultivars from the same genetic group

• Developmental Stage Coverage:

  • Harvest samples at regular intervals during fruit development (e.g., 7-day intervals from 60 days after bloom to commercial harvest)
  • For Palmer mango, specifically include critical developmental windows (91, 98, 105, 112, 119, and 126 days after bloom) [59] [2]

• Spectral Data Acquisition:

  • Clean fruit surface and remove any foreign materials
  • Take measurements at multiple positions on each fruit (typically 3-4 positions) to account for within-fruit variability
  • Ensure consistent contact pressure and orientation between fruit and sensor
  • Maintain consistent instrument settings (integration time, number of scans) throughout data collection
  • Include instrument recalibration every 20-30 samples using standardized reference materials

• Reference Data Collection:

  • Immediately after NIR scanning, destructively analyze the same tissue regions for reference parameters
  • For DM: Use oven-dry method (105°C until constant weight) on fruit flesh samples
  • For SSC: Use digital refractometer on extracted juice
  • For firmness: Use texture analyzer or penetrometer with appropriate probes
  • For titratable acidity: Use automated titrator with standardized NaOH solution

Data Preprocessing and Model Development Protocol

Objective: To transform spectral data and develop calibration models resistant to overfitting.

Materials:

• Chemometric software (e.g., MATLAB, Unscrambler, Python with scikit-learn)
• High-performance computing resources for complex model training

Procedure:

• Spectral Data Preprocessing:
  • Apply multiple preprocessing techniques to different data segments:
    • Standard Normal Variate (SNV) for scatter correction
    • Savitzky-Golay derivatives (1st and 2nd order) to enhance spectral features
    • Multiplicative Scatter Correction (MSC) to compensate for light scattering
    • Detrending to remove baseline shifts
  • Systematically compare preprocessing combinations to identify optimal approaches [13]

• Feature Selection:

  • Employ wavelength selection algorithms (e.g., Genetic Algorithms, CARS, SPA) to identify informative spectral regions
  • Focus on regions known to correlate with chemical bonds of interest:
    • O-H bonds in water (970-1,150 nm)
    • C-H bonds in carbohydrates and oils (1,100-1,250 nm, 1,600-1,800 nm)
    • O-H bonds in sugars (1,400-1,450 nm) [60]

• Model Training with Validation:

  • Utilize Partial Least Squares Regression (PLSR) as primary algorithm for continuous parameters (DM, SSC)
  • Implement Least Squares-Support Vector Machines (LS-SVM) for non-linear relationships
  • Apply fuzzy logic classification for maturity index classification [13]
  • Employ k-fold cross-validation (k=10 recommended) during model development
  • Reserve a completely independent validation set (20-30% of total samples) not used in model training or cross-validation

• Model Robustness Evaluation:

  • Test models on external validation sets from different seasons and orchards
  • Calculate RPD (Ratio of Performance to Deviation) values, with RPD >2.0 indicating good predictive ability [2]
  • Monitor RMSEP (Root Mean Square Error of Prediction) values compared to RMSECV (Root Mean Square Error of Cross-Validation) - similar values indicate good generalizability

The following workflow diagram illustrates the comprehensive model development process:

The Scientist's Toolkit: Essential Research Reagents and Equipment

Table 2: Essential Materials for Handheld NIR Maturity Model Development

Category	Item	Specification/Recommendation	Application Purpose
Instrumentation	Handheld NIR Spectrometer	Felix Instruments F-750 or F-751 (310-1100 nm); Neospectra Micro (1350-2500 nm)	Field and lab spectral data collection
	Reference Analyzers	Digital refractometer (±0.1% SSC); Texture analyzer; Laboratory oven; Automated titrator	Reference value determination for model calibration
Software & Analysis	Chemometric Software	MATLAB with PLS Toolbox; The Unscrambler; Python with scikit-learn	Data preprocessing and model development
	Spectral Preprocessing	Standard Normal Variate (SNV); Savitzky-Golay derivatives; Multiplicative Scatter Correction (MSC)	Spectral data optimization and noise reduction
Sample Management	Calibration Standards	Barium Sulfate (BaSOâ‚„) or certified reference materials	Instrument calibration and validation
	Sample Tracking System	Barcode labels; Database management system	Sample identification and data integrity
Model Validation	Independent Sample Sets	Fruits from different seasons, orchards, and varieties	Model generalizability testing

Developing generalizable handheld NIR models for mango maturity testing requires meticulous attention to experimental design, sample selection, and validation protocols. By implementing the comprehensive approaches outlined in this application note—specifically, incorporating multi-season and multi-orchard sampling, employing sophisticated data preprocessing techniques, and applying rigorous validation procedures—researchers can significantly enhance model robustness and commercial applicability. Continued refinement of these protocols will further advance the adoption of NIR technology as a reliable, non-destructive tool for mango maturity assessment across diverse production environments.

In the field of agricultural produce quality assessment, the mango industry faces significant challenges in determining fruit maturity accurately and non-destructively. Traditional methods often involve destructive sampling, which is labor-intensive, impractical for large-scale operations, and leads to product waste [1]. In recent years, handheld Near-Infrared (NIR) spectroscopy has emerged as a powerful tool for non-destructive quality evaluation, offering rapid analysis without damaging the fruit [5] [61]. The effectiveness of these handheld systems, however, heavily depends on the sophisticated chemometric models that interpret the complex spectral data obtained from fruit scanning.

This application note focuses on two advanced analytical approaches that significantly enhance the predictive performance of handheld NIR systems for mango maturity testing: ensemble methods and Multi-predictor Local Polynomial Regression (MLPR). Ensemble methods combine multiple machine learning models to improve overall prediction accuracy and robustness, while MLPR offers a flexible nonparametric approach for modeling complex relationships between spectral data and maturity indicators [62] [63]. These techniques are particularly valuable for addressing the biological variability inherent in agricultural products and for modeling the non-linear relationships between spectral features and mango quality parameters.

Within the broader context of handheld NIR research for mango maturity testing, these advanced statistical approaches enable more accurate prediction of key maturity indicators such as Total Soluble Solids (TSS), pH, Dry Matter (DM) content, and firmness [5] [1] [28]. The integration of these sophisticated algorithms with portable NIR technology represents a significant advancement toward practical, reliable, and scalable solutions for the fruit industry.

Theoretical Framework

Handheld NIR Spectroscopy Fundamentals

Near-Infrared Spectroscopy operates on the principle that when NIR light (typically in the range of 740-1070 nm for handheld devices) interacts with organic materials, specific chemical bonds undergo vibrational energy transitions that result in characteristic absorption patterns [5] [61]. In mango maturity assessment, these absorption features correspond to molecular vibrations of bonds in compounds such as sugars (O-H), organic acids (C-H), and water (O-H) [1]. The resulting spectra serve as a chemical fingerprint that can be correlated with quality parameters through multivariate calibration.

Handheld NIR spectrometers designed for field use typically employ interactance geometry, where the detector captures light that has penetrated the fruit surface and undergone partial internal scattering [1] [28]. This configuration provides information about both chemical composition and physical properties, making it particularly suitable for assessing internal quality attributes without destructive sampling. The portability of these instruments enables in-situ measurements directly in orchards, packinghouses, and throughout the supply chain, providing real-time decision support for harvest timing and quality grading.

Ensemble Methods in Spectral Analysis

Ensemble methods represent a powerful paradigm in machine learning that combines multiple base models to produce improved predictive performance compared to any single constituent model. The fundamental principle behind ensemble learning is that by aggregating predictions from several models, the overall bias and variance can be reduced, leading to better generalization and robustness [63]. In the context of NIR spectroscopy for mango quality assessment, this approach is particularly valuable due to the high dimensionality of spectral data and the complex, non-linear relationships between spectral features and maturity parameters.

The most common ensemble strategies include:

• Bagging (Bootstrap Aggregating): Creates multiple versions of the training set through bootstrapping and aggregates their predictions, effectively reducing variance. The Random Forest algorithm is a prominent example that builds multiple decision trees and combines their outputs through voting or averaging [5] [1].

• Boosting: Sequentially builds models where each new model focuses on correcting errors made by previous ones, thereby reducing both bias and variance. Gradient Boosting Regression (GBR) has shown exceptional performance in predicting ripening indicators from spectral data [64].

• Stacking: Combines multiple different types of models using a meta-learner that learns how to best weight the predictions of the base models. This approach can leverage the complementary strengths of diverse algorithms [63].

Research has demonstrated that ensemble methods consistently outperform single-model approaches in mango quality prediction. One study reported that ensemble models achieved superior scalability and accuracy in dynamic agricultural environments compared to traditional classification methods [63].

Multi-Predictor Local Polynomial Regression (MLPR)

Multi-Predictor Local Polynomial Regression (MLPR) is a nonparametric regression technique that extends traditional polynomial regression by fitting local models to subsets of data. Unlike global parametric models that assume a specific functional form for the entire dataset, MLPR adapts to local variations in the relationship between predictors and response variables, making it particularly suitable for modeling complex, non-linear relationships in spectral data [62].

The mathematical foundation of MLPR involves fitting a polynomial of degree (d) to a neighborhood of data points around a target point (x_0) using weighted least squares. For a multivariate predictor scenario (as with NIR spectra containing hundreds of wavelengths), the model takes the form:

[ yi = \beta0 + \sum{j=1}^p \betaj(x{ij} - x{0j}) + \sum{j=1}^p \sum{k=j}^p \beta{jk}(x{ij} - x{0j})(x{ik} - x{0k}) + \cdots + \varepsiloni ]

where the coefficients (\beta) are estimated locally for each prediction point, and the contribution of each observation is weighted according to its distance from the target point, typically using a kernel function [62].

The biresponse multipredictor local polynomial nonparametric regression variant has shown exceptional performance in simultaneously predicting multiple mango quality parameters (pH and TSS) from NIR spectra, achieving a Mean Absolute Percentage Error (MAPE) of 4.473%, which indicates high prediction accuracy [62]. This approach is particularly valuable for handheld NIR applications because it does not require strict assumptions about the underlying functional form of the relationship between spectra and quality parameters, allowing it to adapt to the natural biological variability in mango fruits.

Experimental Protocols

Sample Preparation and Spectral Acquisition

Table 1: Sample Preparation Protocol for Mango Maturity Assessment

Step	Parameter	Specifications	Purpose
Sample Selection	Variety	Uniform mango varieties (e.g., 'Keitt')	Minimize biological variability
	Maturity Stages	Multiple harvest dates (e.g., 1 week before, at, and after optimal harvest)	Capture full maturity range [15]
	Sample Size	Minimum 120-198 fruits [5] [15]	Ensure statistical robustness
Sample Handling	Transportation	Temperature-controlled transport within 48h of harvest	Maintain fruit integrity
	Storage Conditions	24±1°C [15]	Standardize pre-measurement conditions
	Surface Preparation	Clean, dry, and label measurement spots	Ensure consistent spectral acquisition
Spectral Acquisition	Instrument	Handheld NIR spectrometer (400-1100 nm or 740-1070 nm) [5] [28]	Portable field measurements
	Geometry	Interactance mode [28]	Probe internal quality attributes
	Measurement Points	3 positions per fruit (top, middle, bottom) [65]	Account for natural variability
	Reference Standards	White reference before each session	Maintain calibration

Proper sample preparation is critical for obtaining reliable NIR spectra. Fruits should be selected to represent the full maturity continuum, with maturity stages verified through destructive reference methods. The measurement positions on each fruit should be marked to ensure consistency in repeated measurements, and environmental conditions should be controlled throughout the process [15] [65].

Reference Methodologies for Quality Parameters

Table 2: Reference Methods for Mango Quality Parameter Validation

Quality Parameter	Reference Method	Protocol	Purpose
Total Soluble Solids (TSS)	Digital Refractometer	Extract juice from scanned areas, measure in °Brix [5]	Quantify sugar content as maturity indicator
pH	pH Meter	Calibrate electrode, measure juice from scanned areas [5]	Assess acidity changes during maturation
Dry Matter (DM) Content	Oven Drying	Dry tissue samples at 105°C to constant weight [28]	Establish maturity classification threshold
Firmness	Penetrometer	Measure flesh resistance with standardized probe	Assess textural changes during ripening

For model development, reference measurements must be performed immediately after spectral acquisition on the exact same positions scanned by the NIR instrument. This temporal and spatial alignment is essential for building accurate calibration models [5] [28].

Data Preprocessing Workflow

Spectral preprocessing is essential for enhancing the signal-to-noise ratio and removing physical light scattering effects unrelated to chemical composition. The optimal preprocessing workflow typically includes:

• Noise Reduction: Apply Savitzky-Golay filtering to smooth spectral noise while preserving meaningful features [64].
• Scatter Correction: Use Standard Normal Variate (SNV) or Multiplicative Scatter Correction (MSC) to minimize the effects of light scattering due to surface irregularities and path length variations [5].
• Spectral Derivatives: Calculate first or second derivatives to resolve overlapping peaks and enhance spectral features, while simultaneously removing baseline offsets [5].

Additional preprocessing techniques may include normalization, detrending, and wavelength selection, depending on the specific instrument characteristics and mango varieties being analyzed.

Implementation Protocols

Ensemble Method Implementation

The implementation of ensemble methods for mango maturity assessment follows a systematic protocol:

• Feature Selection: Apply variable selection algorithms such as Competitive Adaptive Reweighted Sampling (CARS) or Successive Projections Algorithm (SPA) to identify the most informative wavelengths and reduce data dimensionality [15]. This step is crucial for handling the high dimensionality of NIR spectra and improving model interpretability.

• Base Model Training: Develop multiple diverse base models using different algorithms:

  • Partial Least Squares (PLS): A linear regression method that projects predictors and response variables to a new space [5]
  • Support Vector Machine (SVM): Effective for handling non-linear relationships using various kernel functions [5]
  • Random Forest (RF): An ensemble method itself that builds multiple decision trees [5]
  • Neural Networks (NN): Capable of modeling complex non-linear relationships in high-dimensional data [5]

• Ensemble Aggregation: Combine predictions from base models using strategies such as:

  • Weighted Averaging: Assign weights to models based on their individual performance
  • Stacked Generalization: Train a meta-model to learn how to best combine base model predictions
  • Majority Voting: For classification tasks (e.g., mature/immature), use the majority class prediction [28]

Research has demonstrated that ensemble methods can achieve classification accuracy of up to 97.44% for mango variety identification and 88.2% for direct maturity classification, significantly outperforming single-model approaches [5] [28].

MLPR Implementation Protocol

The implementation of Multi-predictor Local Polynomial Regression for mango quality prediction involves the following detailed protocol:

• Bandwidth Selection: Determine the optimal bandwidth parameter (h) that defines the size of the local neighborhood for each prediction point. This can be achieved through cross-validation techniques, selecting the bandwidth that minimizes the prediction error on validation samples. The bandwidth controls the bias-variance tradeoff, with larger bandwidths increasing bias but reducing variance.

• Local Weighting: Implement a kernel weighting function (e.g., Gaussian, Epanechnikov, or tricube kernel) to assign higher weights to observations closer to the target point (x0). The kernel function (Kh(xi - x0)) determines how rapidly weights decrease with distance from the target point.

• Local Polynomial Fitting: For each target point (x_0) in the predictor space:

  • Select the local neighborhood of data points within bandwidth (h)
  • Calculate weights for each observation using the kernel function
  • Fit a polynomial of specified degree (d) using weighted least squares: [ \min{\beta} \sum{i=1}^n Kh(xi - x0) \left[ yi - \beta0 - \sum{j=1}^p \betaj(x{ij} - x_{0j}) - \cdots \right]^2 ]
  • Use the fitted local model to predict the response at (x_0)

• Biresponse Extension: For simultaneous prediction of multiple quality parameters (e.g., pH and TSS), extend the MLPR framework to a multivariate response scenario. This approach leverages correlations between response variables to improve prediction accuracy [62].

The MLPR approach has demonstrated exceptional performance in mango quality prediction, achieving a Mean Absolute Percentage Error (MAPE) of 4.473% for simultaneous prediction of pH and TSS, which is considered highly accurate in agricultural applications [62].

Performance Analysis

Comparative Performance of Advanced Techniques

Table 3: Performance Comparison of Advanced Algorithms for Mango Quality Prediction

Algorithm	Application	Performance Metrics	Advantages	Limitations
Ensemble Methods (Random Forest, GBR)	Maturity classification, TSS prediction	Direct maturity classification: 88.2% accuracy [28]; TSS prediction: R²=0.82, RMSE=0.92 [64]	Robust to outliers, handles non-linear relationships, reduces overfitting	Computational complexity, model interpretability challenges
MLPR (Biresponse)	Simultaneous pH and TSS prediction	MAPE=4.473% for pH and TSS prediction [62]	Adapts to local data structure, no assumptions about global functional form	Computational intensity for large datasets, bandwidth selection sensitivity
Si-PLS (Synergy Interval PLS)	TSS and pH prediction	TSS: R²=0.63, RMSEP=1.83; pH: R²=0.81, RMSEP=0.49 [5]	Effective wavelength selection, improved interpretability	Limited to linear relationships, suboptimal for complex non-linearities
LDA-SVM Hybrid	Variety identification	100% accuracy (training), 97.44% accuracy (prediction) [5]	Combines dimensionality reduction with classification	Primarily suitable for classification tasks
AutoMLP	Quality classification	98.46% accuracy for freshness classification [63]	Automated architecture optimization, handles complex patterns	Black-box nature, extensive data requirements

The performance comparison reveals that each advanced technique offers distinct advantages for specific applications in mango quality assessment. Ensemble methods excel in classification tasks and handling complex, non-linear relationships in spectral data. MLPR provides exceptional accuracy for continuous parameter prediction and adapts well to the natural variability in biological samples. The selection of an appropriate algorithm depends on the specific application requirements, including the need for interpretability, computational constraints, and the nature of the prediction task (classification vs. regression).

Optimization Guidelines

Based on empirical research, the following optimization guidelines can enhance model performance:

• Data Quality and Representation: Ensure the calibration set encompasses the full biological variability expected in the target population, including different varieties, maturity stages, growing conditions, and seasonal variations [1].

• Feature Selection: Implement aggressive variable selection to identify the most informative wavelengths and reduce model complexity. Techniques such as CARS, SPA, and interval PLS can significantly improve model performance and transferability [15].

• Model Validation: Employ appropriate validation strategies including cross-validation, external validation sets, and validation across multiple seasons to ensure model robustness and prevent overfitting [5] [28].

• Computational Efficiency: For real-time applications, balance model complexity with computational requirements. Ensemble methods and MLPR can be computationally intensive, so consider optimized implementations for embedded systems in handheld devices [28].

The Scientist's Toolkit

Table 4: Essential Research Reagent Solutions for Handheld NIR Mango Maturity Studies

Category	Item	Specifications	Application Purpose
Reference Materials	Standard pH buffers	pH 4.0, 7.0, and 10.0 solutions	Calibration of pH meter for reference measurements [5]
	Sucrose standards	5-25° Brix solutions	Verification of refractometer accuracy for TSS measurement [5]
Sample Preparation	Digital balance	0.001g precision	Sample weighing for dry matter determination [28]
	Forced-air oven	105°C constant temperature	Dry matter content determination [28]
	Fruit corer	Stainless steel, standardized size	Tissue sampling for reference analysis
Spectral Standards	White reference tile	Ceramic, spectrally flat	Regular instrument calibration [61]
	Wavelength standards	Rare earth oxides (e.g., Holmium oxide)	Wavelength accuracy verification
Quality Control	Control mango samples	Characterized for key parameters	Method validation and instrument performance tracking
	Temperature logger	±0.5°C accuracy	Environmental monitoring during experiments

This toolkit represents the essential reagents and materials required for developing and validating handheld NIR methods for mango maturity assessment. Proper selection and consistent use of these materials is critical for generating reliable, reproducible results that enable robust model development and meaningful performance comparisons across studies and research groups.

Ensemble methods and Multi-predictor Local Polynomial Regression represent significant advancements in the analytical framework supporting handheld NIR spectroscopy for mango maturity assessment. These sophisticated algorithms enhance the capability to extract meaningful information from complex spectral data, enabling accurate, non-destructive prediction of key quality parameters including TSS, pH, dry matter content, and maturity classification.

The implementation protocols outlined in this document provide researchers with comprehensive methodologies for applying these advanced techniques in practical research scenarios. Through appropriate experimental design, careful data preprocessing, and rigorous model validation, these approaches can deliver performance levels suitable for commercial application in mango quality assessment throughout the supply chain.

As handheld NIR technology continues to evolve, further research opportunities exist in optimizing these algorithms for specific mango varieties, growing regions, and supply chain applications. The integration of these advanced analytical techniques with portable spectroscopy represents a powerful combination that can significantly enhance quality control, reduce postharvest losses, and improve market satisfaction for mango producers worldwide.

Benchmarking Performance: Validating NIR Models Against Traditional Methods

Within the scope of thesis research on handheld Near-Infrared (NIR) methods for mango maturity testing, the selection of appropriate model validation metrics is paramount for developing robust, field-deployable tools. This protocol details the application of key regression metrics—R-squared (R²), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE)—for calibrating and validating predictive models for maturity indices such as Soluble Solids Content (SSC) and firmness. Matthew's Correlation Coefficient (MCC), while noted, is identified as unsuitable for this continuous prediction task and is excluded from the core analytical framework. The guidelines provide researchers with standardized methodologies for model evaluation, ensuring reliable and interpretable outcomes for industrial application.

Near-Infrared (NIR) spectroscopy has emerged as a powerful, non-destructive technique for assessing internal fruit quality attributes [66]. For climacteric fruits like mangoes, determining optimal harvest maturity (the stage at which fruit can be picked and subsequently ripen to acceptable quality) is crucial for minimizing postharvest losses, which can account for 20-40% of produce [66]. Handheld NIR spectrometers operate by irradiating the fruit and measuring the resulting reflected or transmitted radiation in the 780-2500 nm range. The resulting spectra, influenced by the fruit's chemical composition and light-scattering properties, are then correlated with key maturity indices like SSC and firmness using chemometric models [66] [67].

However, the development of a robust calibration model is only half the challenge. As noted in a recent study, "a machine learning model for fruit maturity estimation for one variety may not directly be applicable to other varieties of the same fruit" [66]. This underscores the necessity for rigorous validation using a suite of performance metrics that collectively describe a model's predictive accuracy, precision, and practical utility. The selection of inappropriate metrics can lead to models that perform well in the laboratory but fail under real-world, in-field conditions.

Performance Metrics for Regression Models

In the context of NIR-based mango maturity testing, the model's output (e.g., predicted SSC in °Brix) is a continuous variable. Therefore, regression metrics are the appropriate tools for validation. The following section defines, interprets, and contextualizes the primary metrics used within this framework.

Core Metric Definitions and Applications

Table 1: Summary of Key Regression Performance Metrics for NIR Model Validation

Metric	Mathematical Formula	Interpretation	Advantages	Disadvantages
R-Squared (R²)Coefficient of Determination	( R^2 = 1 - \frac{SS{res}}{SS{tot}} ) [68]	Proportion of variance in the dependent variable explained by the model. Closer to 1 indicates a better fit [68] [69].	Intuitive; scale-free; provides a relative measure of model fit [70].	Does not indicate bias; can be artificially inflated by adding irrelevant variables [68] [71].
Root Mean Square Error (RMSE)	( RMSE = \sqrt{\frac{1}{n} \sum{(yi - \hat{y}i)^2}} ) [68]	Standard deviation of the prediction errors (residuals). Closer to 0 indicates better accuracy [68].	In same units as the response variable (e.g., °Brix, N); penalizes large errors more heavily [68] [69].	Sensitive to outliers [68].
Mean Absolute Percentage Error (MAPE)	( MAPE = \frac{1}{n} \sum{ \left| \frac{yi - \hat{y}i}{y_i} \right| } \times 100\% ) [68]	Average percentage error of the predictions relative to the actual values [68].	Easy to interpret as a percentage; useful for comparing models across different scales [68].	Undefined for zero values; can be biased towards low forecasts [68] [70].

Metric Selection and Synergistic Use

No single metric provides a complete picture of model performance. A robust validation protocol requires their synergistic use:

• R² vs. RMSE: While R² indicates how well the model explains the variance in the data, RMSE provides an absolute measure of the average prediction error in the original units. A model can have a high R² but a high RMSE if the overall pattern is captured but predictions are consistently off by a large margin [71]. For practical decision-making in mango harvesting, the RMSE value for SSC (e.g., in °Brix) gives a direct indication of the prediction accuracy growers can expect.
• The Role of MAPE: MAPE expresses the error as a percentage, making it highly interpretable for stakeholders. For instance, a MAPE of 5% for SSC prediction is intuitively understood as a 5% average deviation from the actual value [68]. This is particularly useful for communicating model performance to non-technical audiences.
• Comparative Guidance: Research suggests that RMSE is often a better choice for model comparison and is the metric minimized by many common regression algorithms [72]. Furthermore, a 2021 study argues that R-squared is "more informative and truthful" than other symmetric percentage errors because it provides a normalized measure of performance that is independent of the data scale [70].

The Inapplicability of Matthew's Correlation Coefficient (MCC)

Matthew's Correlation Coefficient (MCC) is a metric for evaluating binary or multi-class classification models. It is not suitable for regression tasks like predicting continuous maturity parameters (SSC, firmness). Using MCC for a regression problem is technically incorrect. The NIR maturity prediction models discussed here are fundamentally regression-based, and thus, MCC is excluded from this validation protocol.

Experimental Protocol for NIR Model Validation

This protocol outlines the procedure for developing and validating a PLS regression model to predict mango dry matter (DM) or SSC using a handheld NIR device.

The following diagram illustrates the end-to-end workflow for model development and validation.

Detailed Methodology

Sample Preparation and Spectral Acquisition

• Sample Collection: Procure a minimum of 120 mangoes from a target cultivar to ensure a representative sample size covering the expected maturity range (e.g., from immature to ripe) [67].
• Acclimatization: Condition all fruit at a stable temperature (e.g., 25°C) for at least one hour prior to scanning to minimize the effect of temperature on spectral features [73].
• Spectral Acquisition: Using a calibrated handheld NIR spectrometer (e.g., Felix F-750), collect diffuse reflectance spectra from multiple positions on each fruit (e.g., 3 longitudes ~120° apart). This accounts for fruit anisotropy, which is critical for improving model performance [67]. For each scan, ensure firm contact between the fruit and the sensor head or use a consistent distance.

Reference Analysis and Data Pairing

• Destructive Reference Testing: Immediately after NIR scanning, destructively measure the maturity index at the exact scanned location.
  • For Dry Matter (DM): Extract a core of flesh from the scanned area and dry it in an oven to a constant weight [73].
  • For Soluble Solids Content (SSC): Juice the flesh from the scanned area and measure °Brix using a digital refractometer [73].
• Data Logging: Create a dataset where each NIR spectrum is paired with its corresponding, precisely measured reference value (DM or SSC).

Data Pre-processing and Modeling

• Spectral Pre-processing: Apply chemometric pre-processing techniques to the raw spectra to reduce noise and enhance the chemical signal. Common methods include:
  • Multiplicative Scatter Correction (MSC) or Standard Normal Variate (SNV) to correct for light scattering effects [66] [67].
  • Savitzky-Golay Derivatives to resolve overlapping peaks and remove baseline offsets [66].
• Dataset Splitting: Randomly split the paired dataset into a calibration set (e.g., 2/3 of samples) for building the model and a validation set (e.g., 1/3 of samples) for testing its performance on unseen data [67].
• Model Calibration: Develop a Partial Least Squares (PLS) regression model using the calibration set. PLS is the most common technique in NIR spectroscopy as it projects the spectral data to latent variables that maximize covariance with the reference values [66] [67].

Model Validation and Performance Calculation

• Prediction: Use the calibrated PLS model to predict the DM or SSC values for the samples in the withheld validation set.
• Metric Calculation: Calculate the performance metrics by comparing the predicted values against the destructively measured reference values.
  • Formula for R²: ( R^2 = 1 - \frac{\sum{(yi - \hat{y}i)^2}}{\sum{(yi - \bar{y})^2}} ) [68]
  • Formula for RMSE: ( RMSE = \sqrt{\frac{1}{n} \sum{(yi - \hat{y}i)^2}} ) [68]
  • Formula for MAPE: ( MAPE = \frac{100\%}{n} \sum{ \left| \frac{yi - \hat{y}i}{yi} \right| } ) [68]

Table 2: Interpretation of Metric Values for a Maturity Prediction Model (Example for SSC)

Metric	Excellent	Acceptable	Poor	Context for Mango SSC
R²	> 0.80	0.65 - 0.80	< 0.65	An R² of 0.74 for DM prediction was reported in a kiwifruit study, indicating a good fit [73].
RMSE	As close to 0 as possible.	Model-specific.	Model-specific.	An RMSE of 0.5 °Brix means predictions are, on average, 0.5 °Brix away from the true value.
MAPE	< 10%	10% - 20%	> 20%	A MAPE of 5% implies predictions are off by 5% on average, which is highly accurate.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Materials and Equipment for Handheld NIR Maturity Research

Item	Function / Rationale	Example Specifications
Handheld NIR Spectrometer	The core instrument for rapid, non-destructive spectral data collection in the field or lab.	Felix F-750 Produce Quality Meter; Range: 800-2500 nm [73].
Digital Refractometer	Provides the destructive reference measurement for Soluble Solids Content (SSC) in °Brix for model calibration.	Accuracy: ±0.1 °Brix [73].
Laboratory Oven	Used for determining Dry Matter (DM) content via the oven-drying method, a key maturity index.	Capable of maintaining 105°C [73].
Texture Analyzer / Penetrometer	Measures fruit firmness (in Newtons, N) as a destructive reference for a firmness calibration model.	Equipped with a standard Magness-Taylor probe [67].
Chemometric Software	For spectral pre-processing, PLS model development, and calculation of validation metrics.	Thermo TQ Analyst, CAMO Unscrambler, or open-source R/Python packages (e.g., scikit-learn) [68] [67].

The successful deployment of a handheld NIR method for mango maturity testing hinges on a rigorously validated calibration model. By employing a combination of R², RMSE, and MAPE, researchers can comprehensively assess a model's explanatory power, absolute accuracy, and relative error. This protocol provides a standardized framework for achieving this, ensuring that developed models are not only statistically sound but also reliable for making critical harvest and postharvest decisions, ultimately contributing to reduced food waste and improved fruit quality.

The determination of mango maturity and internal quality is paramount for optimizing harvest timing, ensuring consumer satisfaction, and minimizing post-harvest losses. Traditional methods for assessing key quality parameters like Total Soluble Solids (TSS), an indicator of sweetness, and pH, representing acidity, are predominantly destructive, labor-intensive, and impractical for large-scale operations [6] [74]. Near-Infrared (NIR) spectroscopy has emerged as a rapid, non-destructive, and environmentally friendly analytical technique ideal for fruit quality evaluation [75]. However, the efficacy of NIR spectroscopy hinges on the robust regression models used to decipher spectral data and predict physicochemical properties.

This application note provides a structured framework for researchers and scientists conducting a comparative analysis of three regression modeling techniques—Partial Least Squares Regression (PLSR), Support Vector Machine Regression (SVMR), and Multi-Predictor Local Polynomial Regression (MLPR)—for predicting TSS and pH in mangoes, with a specific focus on handheld NIR systems.

Theoretical Background of Regression Models

The development of accurate calibration models is a critical step in NIR spectroscopy. The broad, overlapping bands in NIR spectra necessitate multivariate regression techniques to extract meaningful information [75].

• Partial Least Squares Regression (PLSR): PLSR is one of the most widely used linear algorithms in chemometrics. It works by projecting the predicted variables and the observable variables to a new space, identifying latent variables that maximize the covariance between the spectral data and the reference measurements. This makes it particularly effective when the number of explanatory variables (wavelengths) far exceeds the sample size and when these variables are highly collinear [75] [67].
• Support Vector Machine Regression (SVMR): SVMR is a non-linear learning algorithm based on statistical learning theory. It operates by mapping input data into a high-dimensional feature space using kernel functions, wherein a linear regression is performed. SVMR is particularly powerful for handling complex, non-linear relationships between spectral data and quality parameters, often providing robust models even with limited samples [6] [76].
• Multi-Predictor Local Polynomial Regression (MLPR): MLPR is a nonparametric regression approach that offers great flexibility. Instead of assuming a global functional form, it fits separate polynomial regressions for each point of estimation based on neighboring data points defined by a bandwidth. This local fitting technique allows MLPR to capture inherent nonlinearities in the data without being unduly influenced by outliers, making it highly adaptable to complex spectral patterns [6].

Comparative Analysis of Model Performance

A direct comparison of the predictive performance of PLSR, SVMR, and MLPR for mango TSS and pH is presented below, synthesizing findings from recent research.

Table 1: Comparative Performance of Regression Models for Predicting Mango pH and TSS [6]

Quality Parameter	Regression Model	Spectral Pre-processing	R²	RMSE	MAPE (%)
pH	MLPR	Gaussian Filter Smoothing	Best Reported	Lowest Reported	< 10
	KPLSR*	Various	Lower than MLPR	Higher than MLPR	> 10
	SVMR	Various	Lower than MLPR	Higher than MLPR	> 10
TSS	MLPR	Savitzky-Golay Smoothing	Best Reported	Lowest Reported	< 10
	KPLSR*	Various	Lower than MLPR	Higher than MLPR	> 10
	SVMR	Various	Lower than MLPR	Higher than MLPR	> 10

*KPLSR: Kernel PLS, a nonlinear variant of PLSR.

Table 2: Model Performance for Other Fruits Using Handheld NIR (Contextual Reference)

Fruit	Quality Parameter	Best Model	R²	RMSEP	Citation
Strawberry	TSS	SVMR (with HSV colorspace)	0.792	Not Specified	[76]
Grape	TSS	Gradient Boosting Regression (GBR)	0.82	0.92	[64]
Grape	Anthocyanin	Decision Tree (DT)	0.87	87.81	[64]
Mulberry	pH	PLS / LS-SVM / MLR	~0.90	Not Specified	[74]

Key Findings from Comparative Studies

• Superiority of MLPR: A study on intact mangoes demonstrated that MLPR consistently provided the best predictive performance for both pH and TSS, achieving the highest R² and the lowest Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE < 10%) compared to KPLSR and SVMR [6]. The local fitting nature of MLPR likely allows it to better capture the complex, non-linear relationships between mango spectra and its internal qualities.
• Performance of SVMR: SVMR has shown strong performance in various fruit quality applications. For instance, it was the best model for predicting TSS and pH in strawberries using image-derived color features [76]. Its ability to handle non-linearity makes it a robust alternative to linear models.
• Role of PLSR: While potentially outperformed by more flexible models in some specific mango studies, PLSR remains a highly reliable and widely used workhorse in NIR spectroscopy. It has proven effective for determining pH in bumpy fruits like mulberry and for assessing firmness in peaches, underscoring its general utility [74] [67].
• Criticality of Spectral Pre-processing: The performance of all models is heavily dependent on appropriate spectral pre-processing. The best model for mango pH (MLPR with Gaussian filter) used different pre-processing than the best model for TSS (MLPR with Savitzky-Golay smoothing), highlighting the need to optimize this step for each specific application [6].

Experimental Protocol for Handheld NIR-based Maturity Testing

The following section outlines a standardized protocol for developing and validating regression models for mango TSS and pH prediction.

Sample Preparation

• Sourcing: Collect a sufficient number of mango samples (e.g., >150) from the target cultivar (e.g., Gadung Klonal 21, Palmer) to ensure a robust model [6] [2].
• Maturity Stages: Ensure the sample set encompasses fruits from at least three distinct maturity stages, from unripe to fully ripe, to capture the full physiological and biochemical variability [6] [64].
• Acclimatization: Allow fruits to acclimatize to a stable room temperature (e.g., 25°C) before scanning to minimize the effect of temperature on spectral readings [67].

Spectral Data Acquisition

• Instrumentation: Use a validated handheld NIR spectrometer. Examples include the Felix Instruments F-750 Produce Quality Meter or similar devices, typically operating in a range such as 740-1070 nm [5] [2].
• Measurement Geometry: Given the anisotropic nature of fruit, collect spectra from multiple positions on each fruit (e.g., 3 longitudes ~120° apart and 2-3 latitudes). This practice significantly improves model performance by accounting for natural variation [67].
• Scanning Procedure: Place the handheld probe firmly and consistently against the fruit's skin. Take multiple scans per position and average them to improve the signal-to-noise ratio. Ensure a white reference scan is performed regularly as per the manufacturer's guidelines [2].

Reference Measurement

• Destructive Analysis: Immediately after NIR scanning, destructively analyze the same tissue region scanned by the NIR probe.
• TSS Measurement: Extract juice from the scanned area and measure TSS (°Brix) using a digital refractometer [5] [76].
• pH Measurement: Use the same juice to measure pH with a calibrated digital pH meter [5].

Data Analysis and Modeling

• Spectral Pre-processing: Apply necessary pre-processing techniques to the raw spectra to remove physical light scattering effects and enhance chemical information. Common methods include:
  • Standard Normal Variate (SNV) or Multiplicative Scatter Correction (MSC) for scatter correction [75] [67].
  • Savitzky-Golay derivatives (1st or 2nd) for baseline correction and resolution enhancement [6] [77].
• Dataset Splitting: Divide the dataset into a calibration set (e.g., 70-80% of samples) for model training and a prediction set (e.g., 20-30%) for external validation [67].
• Model Development:
  • PLSR: Use cross-validation (e.g., leave-one-out) on the calibration set to determine the optimal number of latent variables to avoid overfitting [67] [77].
  • SVMR: Optimize hyperparameters such as the kernel type (e.g., Radial Basis Function), cost parameter (C), and epsilon (ε) through grid search and cross-validation [76].
  • MLPR: Optimize the bandwidth parameter and the degree of the local polynomial, as these critically influence the model's flexibility and prediction accuracy [6].
• Model Evaluation: Evaluate and compare all models based on the following metrics on the independent prediction set:
  • Coefficient of Determination (R²)
  • Root Mean Square Error of Prediction (RMSEP)
  • Mean Absolute Percentage Error (MAPE)

Figure 1: Experimental Workflow for NIR Model Development

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Materials for Handheld NIR-based Maturity Testing

Item Category	Specific Examples	Function in Research
Handheld NIR Spectrometer	Felix Instruments F-750 / F-751; \ntec5 VIS/NIR system; \nMicroNIR1700	Primary device for rapid, non-destructive spectral data collection from intact mango fruits. [5] [2] [64]
Reference Analytical Instruments	Digital Refractometer; \nDigital pH Meter	Provides destructive reference measurements (TSS and pH) for model calibration and validation. [6] [5]
Chemometric Software	PLS Toolbox (MATLAB); \nThe Unscrambler; \nR or Python with scikit-learn	Platform for performing spectral pre-processing, developing regression models (PLSR, SVMR, MLPR), and model validation. [6]
Data Pre-processing Algorithms	Standard Normal Variate (SNV); \nSavitzky-Golay Smoothing/Derivatives; \nMultiplicative Scatter Correction (MSC)	Critical for "cleaning" spectral data by removing noise, correcting baseline drift, and minimizing light scattering effects. [6] [75]

This application note delineates a comprehensive protocol for the comparative evaluation of PLSR, SVMR, and MLPR models in predicting mango TSS and pH using handheld NIR spectroscopy. Current research indicates that while PLSR remains a fundamental and reliable tool, advanced non-linear methods like MLPR can offer superior predictive accuracy for this specific application. The successful implementation of this technology hinges on a rigorous experimental design, including representative sampling, multi-point spectral acquisition, appropriate spectral pre-processing, and thorough model validation. The adoption of these robust, non-destructive methods empowers breeders, growers, and food scientists to make data-driven decisions, ultimately enhancing mango quality and promoting sustainable production practices.

Application Notes

The accurate, non-destructive classification of mango maturity is a critical challenge in post-harvest management, directly impacting fruit quality, shelf life, and market value. This document provides detailed application notes and protocols for the implementation of three prominent machine learning classifiers—Linear Discriminant Analysis combined with Support Vector Machine (LDA-SVM), Random Forest (RF), and K-Nearest Neighbors (KNN)—within the context of handheld Near-Infrared (NIR) spectroscopy for mango maturity testing. Research demonstrates that direct maturity classification significantly outperforms traditional regression-based index estimation, with classification accuracy reaching up to 88.2% for KNN and 97.44% for LDA-SVM in practical settings, compared to a maximum of 55.9% accuracy for indirect estimation methods [28]. These models leverage NIR spectral data (typically in the 400-1100 nm or 1350-2500 nm ranges) to non-destructively predict maturity states based on critical internal quality parameters such as Dry Matter (DM), Total Soluble Solids (TSS), and acidity [5] [78] [28]. The following sections offer a comprehensive comparison of model performance, detailed experimental protocols, and a catalog of essential research tools to empower researchers and development professionals in implementing these robust classification solutions.

Performance Comparison of Classification Models

The selection of an appropriate classifier is paramount for achieving high accuracy and robustness in mango maturity prediction. The table below summarizes the documented performance of LDA-SVM, Random Forest, and K-Nearest Neighbors classifiers across various studies.

Table 1: Performance Comparison of Machine Learning Classifiers for Mango Maturity Assessment

Classification Model	Reported Accuracy	Key Strengths	Notable Applications
LDA-SVM (Hybrid)	97.44% - 100% (Prediction/Training) [5]	High accuracy in variety identification; effective for complex, high-dimensional spectral data [5].	Identification of mango varieties using a handheld NIR spectrometer (740-1070 nm) [5].
Random Forest (RF)	98.1% [79]	High accuracy; handles non-linear relationships; provides feature importance metrics [79] [80].	Grading mango quality (G1, G2, G3) based on external features and weight [79].
K-Nearest Neighbors (KNN)	88.2% - 100% [53] [28]	Simple implementation; effective for direct maturity classification; robust non-parametric method [53] [28].	Direct on-tree maturity state (mature/immature) classification using handheld NIR [28]; Ripeness classification using Raman spectroscopy [53].

Model Selection Guidelines

Choosing the optimal model depends on the specific research goals, dataset characteristics, and computational constraints.

• LDA-SVM Hybrid Model: This classifier is ideal for projects focused on distinguishing between different mango varieties or where the spectral data exhibits complex, non-linear patterns that LDA alone cannot separate effectively. Its high accuracy, as demonstrated in variety identification, makes it suitable for high-precision tasks [5].
• Random Forest: RF is an excellent choice for general-purpose maturity or quality grading and for understanding which spectral features most contribute to classification. Its ability to model complex interactions without overfitting is a significant advantage [79] [80].
• K-Nearest Neighbors: KNN is highly effective for the direct classification of maturity state (e.g., mature/immature), especially when the research aims to replace traditional index estimation. It is simple to implement and has proven highly effective in practical, on-tree applications using handheld devices [28].

Experimental Protocols

Protocol 1: Direct Maturity Classification using Handheld NIR Spectrometry

This protocol outlines the procedure for direct maturity state classification of mangoes using a handheld NIR spectrometer and machine learning classifiers, adapting methodologies from recent research [13] [28].

Research Reagent Solutions and Materials

Table 2: Essential Research Tools for Handheld NIR-based Maturity Classification

Item	Specification/Function	Example Products/Models
Handheld NIR Spectrometer	Acquires spectral data from fruit samples.	NeoSpectra Micro (1350-2500 nm) [13], NIRFlex N-500 [80], SCIO [78]
Embedded Computational Hardware	Controls the spectrometer, processes data, and runs classification models.	Raspberry Pi [13], Intel Compute Stick [13]
Calibration Standard	Used for spectrometer calibration to ensure measurement accuracy.	Barium Sulfate (BaSOâ‚„) [13]
Reference Analytical Tools	Provides ground truth data for model training and validation.	Digital Refractometer (for TSS/BRIX), pH Meter, Oven (for Dry Matter) [5] [78]
Software & Libraries	For spectral analysis, model development, and deployment.	Python (Scikit-learn), C++, Orange Data Mining [79] [13] [80]

Sample Preparation and Spectral Acquisition

• Sample Collection: Harvest mango fruit samples representing the target maturity stages (e.g., immature, mature, overripe). The number of days after full bloom (DAF) is a common physiological reference [13] [80]. A minimum sample size of 150-200 fruits is recommended for robust model training [13] [80].
• Labeling: Assign ground truth maturity labels to each sample. This can be based on:
  • Direct Classification Approach: A binary label (e.g., mature/immature) based on a standard threshold for a reference parameter like Dry Matter (e.g., DM ≥ 14%) [28].
  • Indirect Fuzzy Approach: Classify into multiple maturity indices (e.g., 80%, 85%, 90%, 95%, 100%) based on a combination of destructive measurements including TSS, Titratable Acidity (TA), firmness, and starch content [13].
• Spectral Scanning: Calibrate the handheld NIR spectrometer using the calibration standard [13]. For each fruit, collect interactance or reflectance spectra from multiple marked locations (e.g., top, middle, bottom) to account for natural variation [13]. The typical wavelength range for mango analysis is between 400-1100 nm or 1350-2500 nm [13] [28].

Data Preprocessing and Model Training

• Spectral Preprocessing: Process raw spectral data to reduce noise and enhance features. Apply a combination of the following techniques:
  • Scatter Correction: Standard Normal Variate (SNV) or Multiplicative Scatter Correction (MSC).
  • Smoothing: Savitzky-Golay filter to reduce high-frequency noise.
  • Derivation: First or second derivatives (Savitzky-Golay) to resolve overlapping peaks and remove baseline effects [5] [13].
• Data Splitting: Randomly split the preprocessed dataset and corresponding labels into a training set (e.g., 80%) and a hold-out testing set (e.g., 20%) [79] [80].
• Model Training and Optimization:
  • LDA-SVM: First, use LDA to reduce the dimensionality of the spectral data. Then, train an SVM classifier (e.g., with a non-linear kernel like Radial Basis Function) on the transformed data [5].
  • Random Forest: Train an ensemble of decision trees. Use hyperparameter optimization (e.g., GridSearchCV) to tune parameters like the number of estimators (n_estimators) and the maximum depth of trees [79] [80].
  • K-Nearest Neighbors: Train the KNN model by storing the preprocessed training spectra. Optimize the number of neighbors (k) via cross-validation [28].
• Model Evaluation: Evaluate the performance of the trained classifiers on the hold-out test set using metrics such as accuracy, precision, recall, and the confusion matrix.

The following workflow diagram illustrates the direct classification protocol:

Protocol 2: Maturity Classification Based on External Features

This protocol details an alternative approach that utilizes a computer vision system to capture external features for quality grading, which can be combined with weight to infer internal quality [79].

Hardware Setup and Data Collection

• System Setup: Construct a grading system comprising a roller conveyor, a vision chamber with a camera, and a load-cell weight sensor [79].
• Image and Data Acquisition: As mangoes rotate on the conveyor, capture images from multiple angles. Process these images using conventional image processing algorithms to extract external features: length, width, and defect area [79].
• Data Integration: Simultaneously, record the weight of each mango using the load-cell sensor. Combine the extracted image features (length, width, defect) with weight to create a structured dataset [79].

Data Preparation and Model Implementation

• Data Normalization and Cleaning: Apply Data Normalization and Elimination of Outliers (DNEO) to create a clean, reliable dataset [79].
• Feature Engineering: Calculate derived features such as fruit density, which is a critical factor correlating with internal sweetness and quality [79].
• Model Training with Nested Cross-Validation: To ensure an unbiased performance estimate, employ a Nested Cross-Validation (NCV) method while training classifiers like Random Forest, LDA, SVM, and KNN on the structured dataset [79].

The following diagram outlines the comparative analysis workflow for evaluating the three classification models:

The integration of handheld NIR spectroscopy with robust machine learning classifiers presents a transformative solution for non-destructive mango maturity assessment. The direct classification approach, which bypasses the less accurate step of predicting a continuous maturity index, is highly recommended. The choice between LDA-SVM, Random Forest, and KNN should be guided by the specific application: LDA-SVM for high-precision variety or complex pattern discrimination, Random Forest for high-accuracy quality grading and feature interpretation, and KNN for efficient and effective direct maturity state classification. By adhering to the detailed protocols and utilizing the specified research toolkit, scientists and developers can significantly advance the efficiency and reliability of mango quality control within the supply chain.

This application note presents a comparative analysis of direct classification versus indirect estimation methods for determining mango maturity using handheld Near-Infrared (NIR) spectroscopy. Within the context of advancing non-destructive fruit quality assessment, empirical results demonstrate that a direct classification approach achieves significantly higher accuracy (88.2%) in maturity grading compared to traditional indirect estimation models (55.9%). The protocols detailed herein provide researchers and development professionals with reproducible methodologies for implementing these techniques, highlighting the impact of model selection on the efficacy of handheld NIR systems for supply chain optimization.

The mango (Mangifera indica L.) is a high-value tropical fruit whose commercial acceptance is critically dependent on accurate maturity and ripeness assessment [1]. Key internal quality attributes include Dry Matter Content (DMC) and Total Soluble Solids (TSS), which are well-established indicators of eating quality and are strongly correlated with final maturity stages [1]. Traditional methods for assessing these traits are destructive, labor-intensive, and impractical for large-scale postharvest handling.

Near-Infrared (NIR) spectroscopy has emerged as a prominent non-destructive technology for evaluating fruit quality, capable of rapidly capturing chemical information based on molecular bond interactions (O-H, C-H, N-H) in the 740-2500 nm spectral range [1]. This study systematically evaluates two predominant data modeling paradigms within the framework of handheld NIR method development for mango maturity testing: indirect estimation of chemical traits and direct classification of maturity stages.

The following table summarizes the key performance metrics comparing the indirect estimation and direct classification methodologies for mango maturity assessment using handheld NIR spectroscopy.

Table 1: Performance Comparison of Indirect Estimation vs. Direct Classification for Mango Maturity Testing

Performance Metric	Indirect Estimation	Direct Classification
Overall Accuracy	55.9%	88.2%
Model Approach	Regression (PLS-R)	Classification (CNN/Transformers)
Primary Output	Continuous Value (e.g., DMC %)	Discrete Class (e.g., Mature/Immature)
Key Preprocessing Steps	SNV, Detrending, 1st/2nd Derivative [1]	SNV, Detrending, 1st/2nd Derivative [1]
Typical Model Complexity	Moderate	High
Implementation Workflow	Multi-stage	Single-stage
Robustness to Spectral Noise	Moderate	High

Experimental Protocols

Protocol A: Indirect Estimation of Maturity via DMC

Principle: This method indirectly classifies maturity by first using a regression model to predict a continuous chemical trait (Dry Matter Content), followed by a rule-based assignment into maturity categories based on established DMC thresholds [1].

Materials:

• Handheld NIR spectrometer (e.g., with a range of 740-2500 nm)
• Representative mango fruit samples (e.g., Kensington Pride cultivar)
• Laboratory reference equipment: Forced-air oven, analytical balance, refractometer

Procedure:

• Spectral Data Acquisition:
  • Calibrate the handheld NIR spectrometer according to manufacturer specifications.
  • Configure the instrument to operate in interactance optical geometry to capture signals from the inner fruit pulp [1].
  • For each mango fruit, take three spectral measurements at equidistant points around the equatorial region.
  • Ensure consistent, firm contact between the fruit and the spectrometer's measurement window.

• Reference DMC Analysis (Destructive):

  • Following NIR scanning, peel and core each mango sample.
  • Weigh a representative portion of the mango flesh to obtain the initial wet weight.
  • Dry the sample in a forced-air oven at 70°C until a constant weight is achieved (typically 24-48 hours).
  • Weigh the dried sample to obtain the final dry weight.
  • Calculate the DMC using the formula: DMC (%) = (Dry Weight / Wet Weight) × 100.

• Data Preprocessing & Model Development:

  • Preprocess the raw spectral data using techniques such as Standard Normal Variate (SNV) and Detrending to minimize light-scattering effects [1].
  • Employ Savitzky-Golay 1st or 2nd derivative preprocessing to enhance spectral features and remove baseline offsets [1].
  • Split the dataset (spectra and reference DMC values) into calibration and validation sets (e.g., 70/30 split).
  • Develop a Partial Least Squares Regression (PLS-R) model to correlate the preprocessed spectra with the reference DMC values.
  • Validate the model's performance using the independent validation set.

• Maturity Classification:

  • Apply the validated PLS-R model to predict the DMC of unknown samples.
  • Classify mangoes into maturity categories (e.g., "Immature," "Mature," "Over-mature") based on predetermined DMC thresholds (e.g., Mature ≥ 14% DMC).

Protocol B: Direct Classification of Maturity Stage

Principle: This end-to-end approach uses deep learning classification models to directly map raw or preprocessed spectral data into discrete maturity classes, bypassing the need for intermediate chemical estimation [1].

Materials:

• Handheld NIR spectrometer (as in Protocol A)
• Representative mango fruit samples with pre-labeled maturity classes (e.g., via expert visual assessment and firmness testing)

Procedure:

• Spectral Data Acquisition & Labeling:
  • Acquire NIR spectra as described in Protocol A, Section 3.1.1.
  • Assign each mango fruit to a maturity class (e.g., "Immature," "Mature," "Over-mature") based on a consensus from expert assessors using external criteria (skin color, firmness). This establishes the ground-truth labels for the model.

• Data Preprocessing & Dimensionality Reduction:

  • Apply identical spectral preprocessing (SNV, Detrending, Derivatives) as in Protocol A.
  • Optionally, use dimensionality reduction algorithms like Principal Component Analysis (PCA) to reduce the number of spectral variables and mitigate multicollinearity [1].

• Deep Learning Model Development:

  • Architect a Convolutional Neural Network (CNN) or an attention-augmented/transformer-based model designed for 1D spectral data analysis [1].
  • The model's input is the preprocessed spectrum, and the final output layer uses a softmax activation function to generate probabilities for each maturity class.
  • Train the model using the labeled spectral data, optimizing for categorical cross-entropy loss.
  • Employ techniques like k-fold cross-validation to ensure model generalizability.

Research Reagent Solutions & Essential Materials

Table 2: Essential Research Materials for Handheld NIR-based Mango Maturity Testing

Item	Function/Application in Research
Handheld NIR Spectrometer (740-2500 nm)	Core device for non-destructive spectral data acquisition in field or lab settings [1].
Forced-Air Oven	Required for destructive reference analysis to determine Dry Matter Content (DMC) for model calibration [1].
Analytical Balance (High-precision)	Used for measuring wet and dry weights of mango samples for accurate DMC calculation [1].
Refractometer	Validates Total Soluble Solids (TSS) content, another key maturity indicator, for correlative analysis.
Standard Reference Tiles (e.g., Spectralon)	For consistent calibration and validation of the NIR spectrometer before measurement sessions.

Experimental Workflow Diagrams

Diagram 1: Workflow comparison of the two NIR maturity assessment methods.

Diagram 2: Data processing pipeline showing the divergence point for the two modeling approaches.

This application note details a novel methodology for classifying mango maturity using a handheld Near-Infrared (NIR) spectrometer coupled with fuzzy logic analysis. The presented indirect classification approach achieved a 95.7% accuracy in distinguishing between five maturity indices by integrating four key physicochemical parameters: total acidity (TA), soluble solids content (SSC), firmness, and starch content [13]. This protocol provides researchers and agricultural technologists with a complete framework for implementing this high-accuracy, non-destructive testing method, which significantly outperforms traditional direct classification models [13] [28].

Accurate maturity determination in climacteric fruits like mango is critical for supply chain distribution, shelf-life prediction, and meeting consumer preferences. Traditional methods are often destructive, slow, and subjective [13]. While handheld NIR devices offer a non-destructive alternative, many existing solutions rely on direct classification or regression of a single parameter, which can limit accuracy [28]. The Arumanis mango variety poses a particular challenge, as its skin color does not change significantly during maturation, making visual assessment unreliable [13].

The innovation documented herein lies in an indirect classification strategy. Instead of using spectral data to directly assign a maturity class, the NIR data is first used to predict multiple internal quality parameters. These quantitative predictions are then fused using a fuzzy logic system to make the final maturity classification, resulting in superior performance [13].

Research Reagent Solutions and Essential Materials

Table 1: Key materials and equipment required for protocol implementation.

Item Category	Specific Example / Specification	Function / Role in Experiment
Handheld NIR Spectrometer	Neo Spectra Micro (NSM) Development Kit [13]	Spectral data acquisition in the 1350–2500 nm range.
Embedded Computing Unit	Raspberry Pi (Broadcom BCM2835, 512MB RAM) [13]	On-device data processing, model execution, and control.
Software & Programming	Python Programming Language [13]	Data processing, model implementation, and user interface control.
Reference Lab Equipment	Digital Refractometer, pH Meter, Texture Analyzer [5] [6]	Destructive measurement of reference parameters (SSC, pH, Firmness).
Chemical Reagents	Barium Sulfate (BaSOâ‚„) [13]	Used as a calibration standard for the NIR spectrometer.
Fruit Samples	Mango (e.g., Arumanis variety), 35 samples per maturity index [13]	Biological material for model development and validation.

The following tables summarize the core quantitative findings from the research, providing a clear overview of the maturity indices and model performance.

Table 2: Reference maturity indices and corresponding physicochemical parameter guidelines for Arumanis mango [13].

Maturity Index	Days After Full Bloom (DAF)	Shelf Life (Days)	Key Flesh Color	Taste Profile
80%	90 - 95	21 - 25	Butter yellow around seeds	Sweet, Sour, Fresh
85%	~105	14 - 17	Evenly butter yellow	Sweet, Sour, Fresh
90%	~108	~7	Yellow orange	Sweet, Fresh
95%	~112	~5	Orange	Sweet, Fresh
100%	~115	~1	Reddish yellow	Sweet, Fresh

Table 3: Performance comparison of direct versus indirect classification models for mango maturity [13].

Classification Approach	Core Methodology	Best-Performing Model	Reported Accuracy
Direct	Spectral data directly mapped to maturity class	Linear Discriminant Analysis (LDA)	91.43%
Indirect	Spectral data used to predict TA, SSC, firmness, and starch, with Fuzzy Logic for final classification	Partial Least Squares (PLS) + Fuzzy Logic	95.7%

Experimental Protocols

Protocol A: Sample Preparation and Spectral Acquisition

Objective: To collect consistent and reliable NIR spectral data from mango fruit samples of known maturity.

Materials: Handheld NIR device (e.g., Neo Spectra Micro), calibrator (Barium Sulfate), mango samples, sample holder [13].

Steps:

• Sample Harvesting & Labeling: Harvest mangoes according to known Days After Full Bloom (DAF) and label each fruit with its maturity index based on DAF and other agronomic records [13].
• Sample Cleaning: Clean the fruit surface to remove dirt and sap, which can interfere with spectral readings [13].
• Marking Measurement Points: Mark four equidistant locations around the equator of each mango. This ensures spectral data is representative of the whole fruit and accounts for natural variation [13].
• Sensor Calibration: Calibrate the NIR sensor using the BaSOâ‚„ calibrator prior to measuring samples to establish a baseline reference [13].
• Spectral Acquisition: Place the sensor firmly against each marked location on the fruit. Acquire spectral data in the 1350–2500 nm range. The integration time can be adjusted up to 2 seconds to optimize the signal-to-noise ratio [13].
• Data Storage: Save the raw spectral data for each sample, linked to its unique ID and maturity index label.

Protocol B: Reference Parameter Measurement (Destructive)

Objective: To obtain ground-truth data for the four parameters (TA, SSC, firmness, starch) for model training.

Materials: Texture analyzer, digital refractometer, pH meter/TA titration kit, starch assay kit.

Steps:

• Firmness Measurement: Immediately after NIR scanning, use a texture analyzer with a cylindrical probe to perform a puncture test on the fruit flesh. Record the firmness value in Newtons (N) [13] [81].
• Juice Extraction: Homogenize the fruit flesh and extract juice for chemical analysis.
• Soluble Solids Content (SSC): Use a digital refractometer to measure the SSC (ºBrix) of the extracted juice [5] [6].
• Total Acidity (TA): Use a pH meter or titration method to determine the TA, typically expressed as a percentage of citric acid [13] [6].
• Starch Content: Determine starch content using a standard destructive chemical assay (e.g., enzymatic or polarimetric method) on a sample of the homogenized flesh [13].

Protocol C: Model Development and Fuzzy Logic Integration

Objective: To develop a robust model that translates spectral data into a maturity classification via prediction of the four physicochemical parameters.

Materials: Spectral and reference data, computational environment (e.g., Python with scikit-learn, NumPy).

Steps:

• Spectral Preprocessing: Apply a suite of preprocessing techniques to the raw spectral data to reduce noise and enhance features. The successful study tested 12 spectral transformations, including clipping, scatter correction (e.g., Standard Normal Variate - SNV), smoothing (e.g., Savitzky-Golay), and derivatives [13].
• Regression Model Development:
  • Split the dataset into training and testing sets (e.g., 90%/10%) [13].
  • Use Partial Least Squares (PLS) regression to build models that predict each of the four reference parameters (TA, SSC, firmness, starch) from the preprocessed spectral data [13].
  • Validate the regression models using the test set to ensure predictive accuracy for each parameter.
• Fuzzy Logic System Design:
  • Fuzzification: Define fuzzy sets (e.g., "Low," "Medium," "High") for each of the four input parameters (TA, SSC, firmness, starch) and for the output (maturity index) [13].
  • Rule Base Construction: Create a set of IF-THEN rules that map combinations of the input parameters to a maturity index. Example: "IF TA is High AND SSC is Low AND Firmness is High AND Starch is High, THEN Maturity is 80%" [13].
  • Inference and Defuzzification: For a new sample, the system uses the predicted parameter values from the PLS models, applies the fuzzy rules, and aggregates the results to produce a crisp output (one of the five maturity classes) [13].

Workflow and System Architecture Visualization

The following diagram illustrates the complete experimental and analytical workflow, from sample preparation to final classification.

Diagram 1: Indirect mango maturity classification workflow.

Critical Experimental Considerations

• Spectral Preprocessing Optimization: The selection of the optimal combination of preprocessing methods is crucial and should be treated as a hyperparameter. The high performance of the 95.7% accuracy model was achieved by systematically testing and combining 12 different spectral transformations [13].
• Device Agnosticism: The core methodology is not limited to a specific spectrometer. The principles can be applied to other handheld NIR devices operating in similar wavelength ranges (e.g., 740-1070 nm) [5] [28], though model recalibration may be necessary.
• Model Generalizability: To ensure robustness, the model should be validated with samples harvested across different seasons and from different geographical origins. This tests the model's ability to handle natural variation and prevents overfitting to a specific batch of fruit [6].

Conclusion

Handheld NIR spectroscopy has matured into a powerful, non-destructive tool for mango maturity assessment, moving beyond simple regression to sophisticated direct classification and hybrid models. The synthesis of research confirms that methodologies like LDA-SVM and fuzzy logic, which leverage multiple maturity parameters, can achieve accuracy rates exceeding 95%, significantly outperforming traditional indirect estimation. Critical to success are robust optimization strategies for data preprocessing and wavelength selection, which mitigate noise and enhance model focus. Future directions should prioritize the development of automated, real-time systems using embedded hardware and explore the transfer of these rapid, non-destructive analytical principles to biomedical and clinical research, particularly in the quality control of pharmaceutical raw materials and the characterization of biological samples. The continued evolution of ensemble models and deep learning promises even greater accuracy and robustness for practical, in-field applications.

References

• 1. Mango Quality Assessment Using Near-Infrared ... [https://www.mdpi.com/2073-4395/15/10/2271]
• 2. Predictive Modeling for Palmer Mango | NIRS & ... [https://felixinstruments.com/blog/nirs-modeling-for-palmer-mango/]
• 3. The Application of Handheld Near-Infrared Spectroscopy ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10708147/]
• 4. Enhanced near infrared spectral data to improve prediction ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7200245/]
• 5. Application of handheld NIR spectrometer for simultaneous ... [https://www.sciencedirect.com/science/article/pii/S2772375523001855]
• 6. Mango quality prediction based on near-infrared... [https://f1000research.com/articles/12-656]
• 7. Using Dry Matter as a Measure of Maturity & Quality in ... [https://felixinstruments.com/blog/using-dry-matter-as-a-measure-of-maturity-quality-in-mangos/]
• 8. A new approach towards handheld NIR spectroscopy [https://www.sciencedirect.com/science/article/abs/pii/S1350449521000116]
• 9. Building a 'Kent' Mango Firmness Model with NIR ... [https://felixinstruments.com/blog/building-a-kent-mango-firmness-model-with-nir-spectroscopy/]
• 10. Determination of mango physiological indices by near- ... [https://www.sciencedirect.com/science/article/abs/pii/S0925521400001022]
• 11. Rapid and non-destructive determination of quality parameters in the... [https://pubmed.ncbi.nlm.nih.gov/26675859/]
• 12. Near infrared spectroscopy as a fast and non-destructive ... [https://www.sciencedirect.com/science/article/abs/pii/S0168169921006748]
• 13. Using Fuzzy Logic to Increase Accuracy in Mango Maturity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9781779/]
• 14. Comparison of reflectance, interactance and transmission ... [https://www.sciencedirect.com/science/article/abs/pii/S0925521400001307]
• 15. Non-invasive prediction of mango quality using near- ... [https://www.sciencedirect.com/science/article/abs/pii/S0260877423002510]
• 16. Comparison of Diffuse Reflectance and ... [https://www.mdpi.com/2076-3417/14/21/10001]
• 17. Advancement of non-destructive spectral measurements for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10471194/]
• 18. Advances in NIR Spectroscopy for Root Crops [https://felixinstruments.com/blog/advances-in-nir-spectroscopy-for-root-crops/]
• 19. Nondestructive Quality Detection of Characteristic Fruits ... [https://www.mdpi.com/2077-0472/15/20/2167]
• 20. Improving fruit variety classification using near-infrared ... [https://www.sciencedirect.com/science/article/abs/pii/S0889157525000572]
• 21. Implementing near infrared spectroscopy for the online ... [https://pubmed.ncbi.nlm.nih.gov/40267577/]
• 22. What Is Non-Destructive Testing (NDT)? A Complete Guide [https://www.uti.edu/blog/ndt/what-is-non-destructive-testing]
• 23. What Are the Benefits of Non-Destructive Testing? [https://chemresearchco.com/non-destructive-testing-benefits/]
• 24. 8 Benefits Of Non Destructive Testing For Your Business [https://nexxis.com/8-benefits-ndt/]
• 25. Key Advantages Of Non-Destructive Testing [https://www.themanufacturingoutlook.com/news/key-advantages-of-nondestructive-testing--nwid-1012.html]
• 26. 9 Key Benefits Non-Destructive Testing Brings to the Table [https://revolutionized.com/benefits-of-non-destructive-testing/]
• 27. Exploring the Latest Innovations in NDE/NDT for Quality ... [https://movley.com/blog/exploring-the-latest-innovations-in-nde-ndt-for-quality-control-and-manufacturing-insights]
• 28. A new approach towards handheld NIR spectroscopy [https://pure.atu.ie/en/publications/mango-maturity-classification-instead-of-maturity-index-estimatio/]
• 29. Near infrared (NIR) soil spectral library using ... [https://www.si-ware.com/technical-publications/near-infrared-nir-soil-spectral-library-using-the-neospectra-handheld-nir-analyzer-by-si-ware]
• 30. Handheld Smart Spectrometer device based on the ... [https://www.si-ware.com/technical-publications/handheld-smartspectrometer-device-based-on-the-neospectra-micro-development-kit-to-determine-the-degree-of-maturity-the-reflection-spectrum-is-recorded-non-destructively-and-evaluated-directly-in-the-embedded-device]
• 31. A portable NIR system for nondestructive assessment of ... [https://www.sciencedirect.com/science/article/pii/S0023643822007447]
• 32. RS Student Project: A low-cost portable VIS/NIR Spectrometer [https://www.rs-online.com/designspark/rs-student-project-a-low-cost-portable-visnir-spectrometer]
• 33. What is an NIR Calibration? [https://www.kpmanalytics.com/blog/what-is-an-nir-calibration]
• 34. Wavelength selection method for near-infrared ... [https://www.sciencedirect.com/science/article/abs/pii/S0039914025000591]
• 35. Extending certified spectral fluorescence standards for the ... [https://link.springer.com/article/10.1007/s00216-024-05723-w]
• 36. NIR spectroscopy pre-calibrations: Immediate results [https://www.metrohm.com/en/discover/blog/2024/nir-spectroscopy-pre-calibrations.html]
• 37. A Guide to Accelerating Implementation with Pre-Calibrations [https://www.azom.com/article.aspx?ArticleID=23384]
• 38. The Mango core quality non-destructive testing and ... [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5549953]
• 39. Savitzky–Golay smoothing method [https://nirpyresearch.com/savitzky-golay-smoothing-method/]
• 40. Application of spectral derivative data in visible and near ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3116517/]
• 41. Application of a Vis-NIR Spectroscopic Technique to ... [https://www.jstage.jst.go.jp/article/hortj/89/5/89_UTD-168/_html/-char/en]
• 42. How to implement NIR spectroscopy in your laboratory ... [https://www.metrohm.com/en/discover/blog/2024/nir-spectroscopy-in-your-laboratory.html]
• 43. A PLS regression model using NIR spectroscopy for on- ... [https://www.sciencedirect.com/science/article/pii/S0016236111003437]
• 44. Neural Network-Based Analysis and Its Application to ... [https://www.mdpi.com/2076-3417/14/6/2402]
• 45. Interval PLS | Getting started with mdatools for R [https://mda.tools/docs/ipls.html]
• 46. Interval PLS (IPLS) for Variable Selection [https://www.eigenvectordocs.com/index.php?title=Interval_PLS_(IPLS)_for_Variable_Selection]
• 47. Support vector machine regression (SVR/LS-SVM) [https://pubmed.ncbi.nlm.nih.gov/21350755/]
• 48. Optimizing Support Vector Machine for Avocado Ripeness ... [https://jeeemi.org/index.php/jeeemi/article/view/652]
• 49. Comparative performance of artificial neural networks and ... [https://www.sciencedirect.com/science/article/abs/pii/S0963996924016879]
• 50. Use of Artificial Neural Networks and NIR Spectroscopy for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8834220/]
• 51. Interval Partial Least-Squares Regression (iPLS) [https://opg.optica.org/as/abstract.cfm?uri=as-54-3-413]
• 52. A new approach towards handheld NIR spectroscopy [https://www.sciencedirect.com/science/article/pii/S1350449521000116]
• 53. Machine learning classification of mango maturity based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12637035/]
• 54. Impact of spectra pre-processing on prediction accuracy [https://www.sciencedirect.com/science/article/pii/S2666833524001680]
• 55. (Open Access) An automatic generation of pre - processing ... strategy [https://scispace.com/papers/an-automatic-generation-of-pre-processing-strategy-combined-33xk9exn]
• 56. Mango internal defect detection based on optimal ... [https://www.sciencedirect.com/science/article/pii/S258972172100012X]
• 57. Computer Vision System for Mango Fruit Defect Detection Using Deep... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9655355/]
• 58. Exploring NIR spectroscopy data: A practical chemometric ... [https://www.sciencedirect.com/science/article/pii/S0169743924002314]
• 59. Determination of 'Palmer' mango maturity indices using ... [https://www.sciencedirect.com/science/article/abs/pii/S0925521416306858]
• 60. Recent advancements in NIR spectroscopy for assessing the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9597448/]
• 61. Non-Destructive Detection of Fruit Quality - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC12191955/]
• 62. Prediction of pH and total soluble solids content of mango ... [https://scik.org/index.php/cmbn/article/view/7941]
• 63. Non-Destructive Mango Quality Prediction Using Machine ... [https://www.mdpi.com/2673-4591/107/1/116]
• 64. Incorporation of visible/near-infrared spectroscopy and ... [https://www.nature.com/articles/s41598-024-81694-3]
• 65. Machine learning classification of mango maturity based ... [https://peerj.com/articles/20288/]
• 66. Towards fruit maturity estimation using NIR spectroscopy [https://www.sciencedirect.com/science/article/abs/pii/S1350449520305272]
• 67. Application of NIR spectroscopy for firmness evaluation ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2443351/]
• 68. Essential Regression Evaluation Metrics: MSE, RMSE, MAE, R², and Adjusted R² | by FARSHAD K | Medium [https://farshadabdulazeez.medium.com/essential-regression-evaluation-metrics-mse-rmse-mae-r%C2%B2-and-adjusted-r%C2%B2-0600daa1c03a]
• 69. Evaluation Metrics for Your Regression Model - Analytics Vidhya [https://www.analyticsvidhya.com/blog/2021/05/know-the-best-evaluation-metrics-for-your-regression-model/]
• 70. The coefficient of determination R-squared is more ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8279135/]
• 71. MAE, MSE, RMSE, Coefficient of Determination, Adjusted ... [https://medium.com/analytics-vidhya/mae-mse-rmse-coefficient-of-determination-adjusted-r-squared-which-metric-is-better-cd0326a5697e]
• 72. RMSE or MAPE? which one to choose for accuracy? [https://stats.stackexchange.com/questions/554432/rmse-or-mape-which-one-to-choose-for-accuracy]
• 73. NIR Spectroscopy to Assess Edible Fruit Coatings & Maturity [https://felixinstruments.com/blog/nir-spectroscopy-to-assess-edible-fruit-coatings-maturity/]
• 74. Internal quality determination of fruit with bumpy surface ... [https://www.sciencedirect.com/science/article/abs/pii/S153751101100081X]
• 75. The Role of Near-Infrared Spectroscopy in Food Quality ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11544831/]
• 76. Prediction of Total Soluble Solids and pH of Strawberry ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9318015/]
• 77. A Comparative Analysis of Vis-NIR and NIR Spectroscopy [https://www.mdpi.com/2304-8158/13/14/2274]
• 78. A Quality Meter for the Mango Supply Chain [https://felixinstruments.com/blog/a-quality-meter-for-the-mango-supply-chain/]
• 79. Using Machine Learning to Grade the Mango's Quality ... [https://www.mdpi.com/2076-3417/10/17/5775]
• 80. Classification of Harvesting Age of Mango Based on NIR ... [https://www.iieta.org/journals/mmep/paper/10.18280/mmep.100123]
• 81. Integrated Fruit Ripeness Assessment System Based on an ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10931156/]