Mango Quality Assessment Using Near-Infrared Spectroscopy and Hyperspectral Imaging: A Systematic Review

Abstract

Mango is considered a high-value tropical fruit, and its commercial and consumer acceptance depends on internal and external quality attributes such as Total Soluble Solids (TSS), Dry Matter Content (DMC), firmness, ripeness, and surface defects. In recent years, non-destructive sensing technologies such as Near-Infrared Spectroscopy (NIRS) and Hyperspectral Imaging (HSI) have gained prominence for their ability to quickly and accurately evaluate mango quality. In this study, 101 articles published within the last ten years, were systematically retrieved, and 85 research papers were selected for detailed analysis. The review focuses on statistical analysis, conventional machine learning, deep learning, and transformer-based methods applied to mango quality assessment. The objective is to systematically review and analyse data-driven models for non-destructive mango grading using NIRS and HSI technologies, with particular emphasis on data collection methods, preprocessing techniques, dimensionality reduction, and predictive modelling approaches. This review aims to identify the most effective and widely adopted machine learning and deep learning methods, especially transformer models, for accurate and real-time mango quality assessment. Furthermore, it highlights key quality traits evaluated, current research gaps, and future opportunities to advance intelligent, real-time, and automated mango grading systems for practical use in the fruit industry.

1. Introduction

The mango (Mangifera indica L.) is commonly recognised as one of the most cultivated and economically important tropical fruit crops with an exceptional worldwide reputation due to its unique aroma, taste, and high commercial and nutritional value [1]. Mango is the third most internationally traded tropical fruit after bananas and pineapples, generating an estimated USD 55–60 billion annually [2]. Firmness (FI), Soluble Solids Content (SSC), Dry Matter Content (DMC) and Titratable Acidity (TA) are important indicators of ripeness, sweetness, and eating quality [3], all affecting the marketability and consumer acceptance of the fruit. Conventional approaches to assess these quality traits involve destructive sampling, which is labour intensive and impractical for large-scale postharvest processes. This has led to growing interest in non-destructive, rapid, and reliable methods for the effective grading, sorting, and quality monitoring of mangoes throughout the supply chain [4]. Among these approaches, Near-Infrared Spectroscopy (NIRS) and Hyperspectral Imaging (HSI) have shown great potential as advanced technologies for the simultaneous capture of chemical and visual information and, therefore, the comprehensive evaluation of fruit internal and external parameters [5]. As noted in the review of [6,7], there has been an evolution in the chemometric techniques used in NIRS fruit quality evaluation, evolving from Multiple Linear Regression (MLR) through Partial Least Square (PLS) and Support Vector Machine (SVM) to deep learning techniques such an Artificial Neural Network (ANN). New entries to the field include Convolutional Neural Network (CNN) and transformers. This literature review seeks to highlight recent developments in the application of NIRS, HSI, and data-driven modelling techniques to mango quality assessment, as well as current practices, research gaps, and future possibilities in the field.

The NIR region is defined as 740–2500 nm. Vibrational transitions and combinations of bonds such as C-H, O-H, and N-H result in absorption features, providing information about the internal chemical properties, while light scattering can provide information on internal physical properties, without destroying the sample [8].

HSI combines conventional imaging and spectroscopy. Instead of capturing just spatial information or just spectral information, it collects both. Each pixel in a hyperspectral image contains a full spectrum across many narrow contiguous wavelength bands. This hypercube of spatial and spectral data allows detection of both external features and internal properties of a sample [8,9].

Therefore, the purpose of this study is to address the identified gaps by reviewing the quality attributes of mangoes relevant to non-destructive grading using NIRS and HSI technologies. Furthermore, the major contributions of this work include the following:

• To investigate the data collection methods and system configurations used for acquiring NIRS and HSI data in non-destructive mango quality assessment.
• To review and compare preprocessing and dimensionality reduction techniques applied to NIRS and HSI datasets for enhancing model accuracy and efficiency in mango grading.
• To identify and evaluate the effectiveness of attention-augmented deep learning and transformer-based models for non-destructive mango quality assessment, in comparison with traditional statistical and machine learning methods.
• To assess the internal and external mango quality traits that can be accurately measured using NIRS and HSI technologies.
• To analyse the current challenges, limitations, and future opportunities in applying intelligent data-driven models for real-time large-scale mango grading, with a focus on improving postharvest management and supply chain sustainability.

2. The Approach of the Survey

This study aims to conduct a comprehensive review of recent advancements in mango quality assessment using NIRS and HSI. These non-destructive techniques have gained considerable attention over the past decade for their ability to assess various fruit quality attributes such as ripeness, maturity, and internal defects. To identify the relevant literature, a systematic search was performed using three major academic sources; Scopus, Web of Science, and Google Scholar; focusing on studies published between 1 January 2015 and 30 June 2025. The PRISMA guidelines were followed, as shown in Figure 1, to ensure a transparent and replicable process for article selection, screening, and inclusion [10].

2.1. Eligibility Criteria

To ensure relevance and quality, the inclusion criteria were strictly defined. Only peer-reviewed journal articles published in English were considered. The studies focused on mangoes and utilised NIRS or HSI methods for assessing quality parameters such as ripeness, maturity, or classification. Review articles, meta-analyses, and literature overviews were excluded, along with studies unrelated to mango or not employing spectral techniques. For Google Scholar, additional keyword-based exclusions were used to filter out review-related terms and non-fruit-specific content like leaves and plant parts.

2.2. Search Strategy and Information Sources

The search was conducted using a structured query to retrieve the relevant literature from Scopus, Web of Science, and Google Scholar. The following keywords were used to construct the search string:

(“near-infrared spectroscopy” OR NIRS OR “hyperspectral imaging” OR HSI) AND mango AND (quality OR ripeness OR maturity OR detection OR classification OR grading).

The filters applied included publication years, English language, and article type. Scopus and Web of Science have built-in filters for articles and language. However, Google Scholar does not provide an article-type filter, so a Boolean query was created, which excluded review-related terms such as review, “review article”, “systematic review”, “meta-analysis”, and non-fruit-specific terms like leaf and leaves.

2.3. Selection Process

The selection process followed PRISMA guidelines and consisted of multiple phases. After removing duplicate and ineligible records using automated tools and manual checks, 115 articles were retained for abstract screening. Fourteen of these were excluded due to a lack of relevance based on the titles and abstracts. The remaining 101 articles underwent full-text evaluation, where 16 additional studies were removed for not meeting the inclusion criteria. Ultimately, 85 articles were identified as eligible and were included in the final analysis.

The following research questions were considered:

• RQ1: How are NIRS and HSI data collected for mango quality assessment?

This question aims to explore the various data acquisition methods, instruments, and experimental setups used to collect NIRS and HSI data for the non-destructive evaluation of mango quality.

• RQ2: What preprocessing and dimensionality reduction techniques are used for NIRS and HSI data analysis?

This question investigates the common preprocessing methods and dimensionality reduction algorithms applied to NIRS and HSI data to enhance the data quality, reduce noise, and improve model performance in mango quality assessment.

• RQ3: How effective are attention-augmented deep learning and transformer-based models for non-destructive mango classification and grading compared to traditional statistical and machine learning methods?

The research question examines whether attention-augmented deep learning and transformer models outperform traditional statistical and machine learning methods in non-destructive mango classification and grading.

• RQ4: What quality traits are assessed in mangoes using NIRS and HSI data?

This question aims to determine the specific internal and external quality attributes that can be evaluated non-destructively using NIRS and HSI technologies in mango grading systems.

• RQ5: What are the key challenges and future opportunities in applying advanced machine learning and deep learning techniques for mango grading using NIRS and HSI data?

This question investigates the technical limitations, data-related issues, and research gaps, while also identifying promising directions and innovations for future intelligent mango quality assessment systems.

3. Background

3.1. Optical Geometry

Optical geometry describes the 3D arrangement of the light source, the fruit sample, and the detector in spectroscopic and imaging-based systems. It plays a crucial role in determining the type and extent of the data recorded when evaluating the fruit quality. Reflectance, transmittance, and interactance are three of the geometries commonly used. In the reflectance mode, the detector absorbs light reflected from the external or near-surface layers of the fruit, which is suitable for assessing external characteristics such as skin colour, firmness, or bruises. Transmittance geometry (light passing through the fruit and detected on the opposite side) is especially suited to measure internal traits like sugar, dry matter, or internal browning. A compromise is a partial transmittance, also known as interactance geometry, which involves captures of light that has entered the fruit and passed through part of the fruit. Geometries are selected depending on the desired attributes of quality and the physical properties of the fruit being analysed [11].

3.2. Spectroscopy

As a non-destructive technique, spectroscopy has been adopted in the commercial analysis of the quality of fruits, in context of sugar content, firmness, ripeness, moisture content, and even the existence of defects or diseases, etc. Spectroscopy is non-destructive, allowing for repeated measurements without sample alteration, making it suitable for postharvest monitoring and grading applications, unlike conventional destructive methodologies for sugar measurement. NIRS, which is based on the principle of light and molecular bond interactions, primarily in the 700–2500 nm range, can be used to gain valuable information on the chemical composition of fruits, in particular, the O-H, C-H, and N-H bonds in water, sugars, and proteins [11]. Hyperspectral imaging takes one step further by acquiring spatial and spectral data in an integrated manner to facilitate the mapping of fruit quality at both the surface and internal layers [12].

Automation in sorting and grading processes has seen significant advancements with the integration of spectroscopy in the fruit industry. For example, NIR devices are currently employed in packing lines to evaluate the internal quality of various fruits, including mangoes, apples, and kiwis. HSI systems are used to detect defects such as bruises, fungal infections, or skin blemishes that are not yet detectable by the human eye. These technologies provide support for more sustainable practices in agriculture, reduce waste, and lead to a better consumer experience with homogeneous product quality. In conjunction with machine learning models, spectroscopy can better classify fruits and predict their quality, thereby increasing the classification accuracy and prediction robustness [13]. With reported improvements in sensor technology and data analytics, spectroscopy has tremendous potential for increased use in precision agriculture and smart farming systems.

3.3. Near-Infrared Spectroscopy (NIRS)

NIRS is a widely used method for fruit quality assessment because of its non-destructive, rapid, and chemical-free features. This methodology enables the assessment of significant fruit quality traits, including TSS, firmness, dry matter, acidity, and internal damage, all of which can be measured without cutting or parting the fruit. In this regard, NIRS has been applied to evaluate the maturity, sweetness, and shelf life of fruits, including mangoes, apples, kiwifruits, and peaches, over the past few years [11]. This enables NIRS to be one of the preferred quality control techniques in postharvest handling and processing.

The development of portable and handheld NIR devices, combined with multivariate analysis and machine learning algorithms in NIRS, has made it a promising approach for grading and classification of fruit. The models receive the spectral data and are used to predict quality traits and classify fruits according to their internal composition and maturity levels. For instance, NIRS has been employed to predict ripeness and internal quality indices in mangoes, showing a high correlation with traditional measurements and therefore is suitable for both field and industrial settings [14].

3.4. Hyperspectral Imaging (HSI)

HSI is a state-of-the-art technique in the field of fruit quality assessment, a fusion of imaging and spectroscopy that simultaneously obtains spatial and spectral information. HSI differs from conventional spectroscopy, which provides only average spectrum data for samples, as HSI collects detailed reflectance spectra from each pixel of an image, allowing both external and internal details of fruit characteristics to be visualised and analysed. It is particularly useful for detecting surface defects, bruises, fungal infections, colour uniformity, and minor differences in ripeness, among other parameters. It enables non-destructive quality measurement, predicting internal parameters such as sugar content, firmness, and moisture distribution. For example, maturity stages and quality variation levels in mangoes, apples, and citrus fruits are detected using HSI very efficiently [9].

However, recent developments in HSI, such as faster data acquisition systems, portable devices, and integration with Artificial Intelligence (AI), have greatly improved its practicality for commercial applications. Partial least squares regression (PLSR), SVM, and CNN-based machine learning algorithms have been widely applied to recognise the large number of datasets from HSI as well as predict the qualitative characteristics of fruit accurately [15]. HSI has been previously employed for fruit ripeness classification and the early detection of spongy tissue or bruises in mango fruit with high performance. The ability of HSI to non-invasively monitor visible and invisible quality factors makes it a perfect tool for use in automated sorting lines, quality control, and traceability of the postharvest supply chain. Looking ahead, boosted by the lower prices of sensor technologies and increasing computational power, hyperspectral imaging is poised to transform quality inspection in the fruit industry [16].

3.5. Machine Learning (ML)

ML has been a game changer in fruit quality assessment, enabling the automatic and precise visual evaluation of different-quality parameters, including fruit ripeness, fruit size, colour, firmness, sweetness, and defects. Fruit sorting methods that are state of the art are subjective, lead to waste and are labour intensive, while ML-based methods are scalable, objective and non-destructive by nature. A variety of data sources, such as images, NIRS, HSI, and other sensor outputs, have been shown to enable ML models to accurately classify and predict quality traits in fruits. From analysing various attributes of these fruits like external appearance and internal elements, it is found that the most prominent techniques are SVM, decision trees, K-Nearest Neighbours (KNN), and ANN [15].

Machine learning has one of the greatest advantages in discovering complex patterns in high-dimensional data sets, which are common in the monitoring of fruit quality. ML algorithms have also used materials such as images and NIR spectra to classify maturity stages, predict SSC, and polishing and/or internal defects (for mangoes and apples) [13,17]. According to the literature, deep learning, specifically CNNs, have performed extremely well in the domain of image-based classification problems such as bruise detection, skin defects, and shape abnormalities. These parametric non-linear models can learn hierarchical features from raw image data, making manual feature extraction redundant, while also yielding accurate classification results for real-time industrial applications [18].

3.6. Deep Learning (DL)

DL is a highly successful modern method of defining the quality of fruit based on both internal and external features. More importantly, deep learning models, particularly CNNs, can autonomously learn the important features from raw input data (e.g., images or spectral inputs) without the need to predefine the features beforehand, in contrast to traditional machine learning. They excel in applications including fruit ripeness classification, defect detection, firmness estimation, and sugar content prediction. CNN is one of the widely used techniques for detecting the surface defects of apples, ripening stages of mangoes, and freshness of citrus fruits and outperformed the traditional methods in terms of accuracy and speed [15,18,19]. DL now plays a critical role in smart agriculture and automated postharvest quality control systems in a world where deep learning penetrates each corner of imaging technologies and edge computing.

3.7. Transformers

Transformers, initially developed for natural language processing, are now gaining traction in the field of fruit quality assessment due to their ability to model long-range dependencies and handle complex high-dimensional data. Unlike traditional convolutional architectures, transformer models use self-attention mechanisms to capture global patterns in data, making them particularly effective for analysing hyperspectral images and time-series sensor data related to fruit ripeness, firmness, or disease detection. Recent studies have demonstrated the effectiveness of Vision Transformers (ViTs) and hybrid CNN-transformer architectures in classifying fruit maturity stages, detecting surface defects, and predicting internal quality attributes with high accuracy, outperforming conventional deep learning methods in some cases [20,21]. As transformer models continue to evolve and become more efficient, their integration into smart agriculture tools is expected to enhance real-time decision-making and automate quality assessment processes across the fruit supply chain.

3.8. Evaluation Metrics

In mango quality assessment using NIRS and HSI data, the performance of regression and classification models is typically evaluated using well-established evaluation metrics. For regression tasks, metrics, such as the coefficient of determination (R²), Root Mean Square Error (RMSE), and Residual Predictive Deviation (RPD) are widely used.

The R² indicates the proportion of variance explained by the model and is calculated as

$$R^2=1-{\displaystyle \frac{\sum {(y_i-{\mathrm{ŷ}}_i)}^2}{\sum {(y_i-\mathrm{ȳ})}^2}}.$$

(1)

The RMSE measures the average magnitude of prediction error and is given by

$$RMSE= \sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{(y_i-{ \mathrm{ŷ}}_i)}^2}.$$

(2)

The RPD is the ratio of the standard deviation of the reference values to the RMSE:

$$RPD = SD/RMSE.$$

(3)

For classification tasks, performance is assessed using metrics such as Accuracy, Precision, Recall, and F1-score. These are derived from the confusion matrix:

$$Accuracy = (TP + TN)/(TP + TN + FP + FN)$$

(4)

$$Precision=TP/(TP+FP)$$

(5)

$$Recall=TP/(TP+FN)$$

(6)

$$F1-score=2 * (Precision * Recall)/(Precision+Recall)$$

(7)

where TP = True Positives; TN = True Negatives; FP = False Positives; and FN = False Negatives. These metrics ensure a comprehensive understanding of how well a model performs in real-world grading and prediction tasks.

4. Mango Quality Assessment Using NIRS and HSI

The taxonomy of mango quality assessment using NIRS and HSI datasets categorises approaches based on data acquisition methods, preprocessing and dimensionality reduction techniques, analysis models, and quality traits.

4.1. Data Collection

NIR reflectance or absorbance spectra may be used as-is or preprocessed with various techniques before use in developing predictive models for quantitative estimation of attributes such as dry matter content or classifications tasks, such as ripeness level. Numerous studies have considered the use of this technique in the evaluation of the properties of mango fruit. For example, a study found that NIRS outperformed HSI in predicting most internal traits such as firmness, TSS, and TA, while HSI showed better performance for β-carotene and pH, likely due to its coverage of the visible wavelength range [22].

Another study evaluated the performance of four miniaturised Near-Infrared (NIR) spectrometers: SCiO, Linksquare, DLP NIRscan Nano, and Neospectra, for the non-destructive prediction of key quality parameters in Nam Dok Mai mangoes, including Dry Matter (DM), TSS, TA, pH, and firmness. Calibration models were developed using PLSR. The SCiO and Linksquare spectrometers performed well for DM, TSS, and pH (R² > 0.80), with Linksquare also showing good results for TA. However, all spectrometers struggled with predicting the firmness accurately. The DLP NIRscan Nano and Neospectra spectrometers demonstrated only moderate performance for most parameters, with Neospectra slightly outperforming DLP. Overall, the SCiO and Linksquare instruments showed strong potential for low-cost field-based mango quality assessment [23]. Another study presented the design, construction, and evaluation of a low-cost NIR spectrometer prototype using the Hamamatsu C14384MA-01 sensor for the non-destructive assessment of mango quality [24].

While varying in the instrumentation used in the collection of spectra, all studies used a pipeline involving preprocessing and dimensionality reduction, followed by model development.

4.2. Data Preprocessing and Dimensionality Reduction Techniques

Preprocessing and dimensionality reduction are essential steps in the analysis of NIRS and HSI datasets, as they significantly influence the accuracy, efficiency, and interpretability of the models [25]. Preprocessing addresses common issues in raw spectral data, such as noise, baseline drift, and light scattering, which can result from sensor variability, environmental factors, or sample heterogeneity. Techniques such as smoothing, normalisation, derivative transformations, and scatter correction help enhance meaningful spectral features while reducing irrelevant variability [26].

Similarly, dimensionality reduction methods such as Principal Component Analysis (PCA), t-Distributed Stochastic Neighbour Embedding (t-SNE), or feature selection techniques are crucial for handling the high dimensionality and redundancy inherent in NIRS and HSI data [27]. These methods reduce the computational complexity, mitigate overfitting, and highlight the most informative features for analysis.

4.2.1. Data Preprocessing Techniques

Removing outliers using Hotelling’s T² and Q statistics and augmenting spectral data through stacking various preprocessing methods, one study demonstrated the improved performance of a 1D Convolutional Neural Network (1D-CNN). This highlights that integrating chemometric techniques with DL can significantly enhance the spectral model robustness and accuracy, suggesting a promising direction for future spectral data analysis [28]. In a similar vein, another study showed that enhancing NIR spectral data using Multiplicative Scatter Correction (MSC), Baseline Correction (BLC), and their combination significantly improved the accuracy of predicting total acidity and vitamin C in intact mangoes, thereby making the process more robust, reliable, and non-destructive [29]. Furthermore, a separate investigation into the effect of spectral preprocessing on NIRS prediction accuracy for TA and SSC in mangoes revealed that although MSC enhanced the model performance, the prediction accuracy (RPD < 2) remained insufficient for reliable real-time applications [26].

Complementing these findings, another study tested two key hypotheses: (1) that common preprocessing methods like scatter correction might degrade the model accuracy by eliminating useful scattering information and (2) that DL can more effectively model raw NIR spectra, which contains both absorption and scattering components, compared to PLS. Using a large dataset of 11,420 NIR spectra, the results supported both hypotheses. Raw absorbance data achieved the best prediction accuracy for both PLS and DL, with DL delivering the lowest RMSEP of 0.76%, outperforming all the preprocessed models. While some preprocessing, such as Variance Stabilising Normalisation (VSN) transformation, contributed complementary information, raw spectral data proved more effective for modelling, especially when using DL [25].

4.2.2. Dimensionality Reduction Techniques

A study evaluating Spectral Reflectance Indices (SRIs) to estimate the fruit quality parameters of mango and strawberries at different ripening stages found that newly developed SRIs demonstrated significantly higher predictive power than traditional indices. This was particularly evident for key traits such as SPAD, TSS, and firmness in mango and L*, b*, TSS, and firmness in strawberry. In mango, SRIs, such as RSI764,766 and RSI768,770 achieved R² values up to 0.91, while in strawberry, RSI666,636 and RSI660,620 reached R² values up to 0.84. These findings underscore the effectiveness of carefully selected SRIs as non-destructive indicators of fruit biochemical quality [30]. Similarly, another study employed NIR spectroscopy combined with feature selection methods to detect internal defects in mangoes. Using Fisher’s criterion, the optimal wavelength range of 702–752 nm was identified, achieving the highest classification accuracy of 84.5%. Moreover, lower wavelengths (673–1100 nm) proved more effective than higher wavelengths in distinguishing between healthy and defective fruits [27].

4.3. Algorithms Used for Data Analysis

Various techniques are used in NIRS and HSI data analysis to ensure accurate and efficient interpretation of complex spectral data. These include statistical methods, such as PCA and Linear Discriminant Analysis (LDA), for feature extraction and classification, as well as machine learning algorithms like SVM, k-NN, and Random Forest (RF) for predictive modelling. Additionally, deep learning approaches, including CNN and Transformers, are utilised for automated feature learning. Together, these techniques enable reliable quality assessment, classification, and grading of agricultural products.

4.3.1. Statistical Analysis Methods

A wide range of statistical and chemometric models have been employed to evaluate mango quality using NIRS and HSI technologies across multiple cultivars as summarised in

Table 1. Summary of statistical analysis.

Mango cv	Sample Info	NIRS/ HSI	Spectral Range (nm)	Algorithm Used	Quality Trait	Accuracy Achieved	Ref.
Bolibo	Five batches (each split into three subsamples of 500 g)	VIS-NIR	400–2500	PCA, ASCA	Drying effect on chlorophyll, carotenoids, water, and sugar	VIS-NIR: batch effect 47.5%, dryer effect 23.6%; NIR: dryer effect 38.3%, batch 29.4% (based on ASCA significance)	[31]
Calypso, Honey Gold, Keitt	828 samples (2206 spectra from 15 populations at various temperatures)	NIRS	729–975	PLSR + EPO, GLSW, bias correction, global modelling, repeatability file	Dry matter content (DMC)	Best: RMSEP = 1.05% (EPO), R2 = 0.82; Original model: RMSEP = 1.43%, R2 = 0.68	[32]
Nam Dok Mai Sithong	182	VIS-NIR	600–1000	PLSR (with MAS + baseline offset)	Total soluble solids (TSS)	R = 0.80 (calibration), r = 0.74 (prediction), RMSEC = 0.690%, RMSEP = 0.765%	[33]
Tommy Atkins	Not specified	HSI	880–1720	PLSR (with SNV + mean centring)	Moisture content (MC) and drying uniformity	R2 = 0.995; RMSEC = 1.881%; RMSEP = 1.408%	[34]
Nam Dok Mai	207 (for DM); up to 223 for firmness	NIRS	740–2500	PLSR	DM, TSS, titratable acidity (TA), pH, firmness	Best results: SCiO: R2cv = 0.92 (DM), 0.84 (TSS), 0.74 (pH) Linksquare: R2c = 0.91 (TSS, TA), 0.93 (pH), 0.81 (DM)	[23]
Nam Dok Mai (Si Thong)	188	NIRS + HSI	NIRS: 800–2500 HSI: 450–998	PLSR, MLR	FI, TSS, TA, pH, β-carotene, RPI	NIRS-PLSR: FI: R2V = 0.84, RPD = 2.57 TSS: R2V = 0.85, RPD = 2.64 TA: R2V = 0.89, RPD = 3.07 pH: R2V = 0.88, RPD = 2.99 β-carotene: R2V = 0.86, RPD = 2.65 RPI: R2V = 0.90, RPD = 3.16	[22]
Nam Dok Mai	600 total (two periods)	NIRS	640–1050	PLSR (with SNV, SG deriv.)	TSS, TA, pH, DM, firmness	TSS: R2c = 0.84, R2cv = 0.76, R2p = 0.81, RMSEP = 1.07° Brix TA: R2c = 0.92, R2cv = 0.90, R2p = 0.84, RMSEP = 0.36% pH: R2p = 0.80, RMSEP = 0.45 DM: R2p = 0.66 Firmness: R2p = 0.09	[24]
Alphonso	708 for training/validation + 90 test	VIS-NIRS	350–2500; best: 670–970	PCA + SIMCA	Spongy tissue (detection/classification)	SIMCA binary classification (healthy vs. spongy): 900–970 nm: 96.7% overall (100% healthy, 93.3% affected) 670–750 nm: 94.4% overall (97.8% healthy, 91.1% affected)	[35]
Guifeimang	134	HSI (VIS-NIR)	400–1000	SNV + CARS + PLSR	Soluble solid content (SSC)	Prediction: R2p = 0.90, RMSEP = 0.616, using 11.3% of bands (37 wavelengths); Calibration: R2c = 0.902, RMSEC = 0.486	[36]
Not specified (from Hainan, China)	270 (four defect types)	HSI	450–1000	PCA + band ratio (Q744/942) + I-Otsu + linear stretch	Skin defect detection (black spot, scab, bruised)	Overall accuracy: 96.67% (No false positives; nine false negatives)	[37]
Irwin (Japan)	122 fruits/576 spectra	NIRS	570–1030 (used: 600–1000)	PLSR (second derivative + CV)	Soluble solid content (SSC), skin colour	SSC: R2cv = 0.76, SECV = 0.70, RPD = 2.1 Skin Colour: R2cv = 0.78, SECV = 2.97, RPD = 2.1	[38]
Nam Dok Mai Si Thong	294 (99 + 98 + 97 @ DAFS)	NIRS	700–1100	PLSR + SNV/derivatives + PCA + FQI modelling	Fruit quality Indices (FQI1, FQI2) combining TSS, TA, DM, pH, firmness	FQI1: Q2 = 0.84, RMSEP = 9.39 FQI2: Q2 = 0.85, RMSEP = 8.38	[39]
Nam Dokmai	100 (200 spectra)	NIRS	600–1100	PLSR + SNV + empirical correction (ΔA method)	Soluble solids content (SSC)	Flesh model: r = 0.88, RMSEP = 1.27° Brix Unpeeled model: r = 0.84, RMSEP = 1.50° Brix Empirical-corrected model: r = 0.87, RMSEP = 1.36° Brix	[40]
Kent	58	NIRS	1000–2500	PLSR + MSC + BLC (enhanced spectra)	Total acidity (TA), vitamin C	TA: R2 = 0.976, RMSE = 19.35, RPD = 6.77 Vit C: R2 = 0.958, RMSE = 0.417, RPD = 3.14 (MSC + BLC enhanced)	[29]
Cengkir, Kweni, Kent, Palmer	186	NIRS	1000–2500	PLSR + EMSC (best), SNV, MAS	TSS (Brix), vitamin C (mg/100 g FM)	TSS: r = 0.86, RMSEP = 1.58, RPD = 2.25, RER = 9.72 Vit C: r = 0.86, RMSEP = 6.79, RPD = 2.19, RER = 8.87 (prediction dataset, EMSC model)	[41]
Not specified (from Tianjin, China)	240 (60 mangoes × three heights + control)	HSI	900–1700	PLSR + SNV + CARS	Firmness, TSS, TA, chroma (∆b*) and damage degree	Firmness: R2 = 0.84, RMSEP = 31.6 g TSS: R2 = 0.9, RMSEP = 0.49° Brix TA: R2 = 0.65, RMSEP = 0.1% Chroma: R2 = 0.94, RMSEP = 0.96	[42]
Kent	50 (× five days = 250 total)	VIS-NIRS	700–1130	iPLSR, PLSR + MSC + SG	Firmness during ripening	iPLSR: R2p = 0.75, RMSEP = 5.92 Hz2g2/3 PLSR: R2p = 0.67, RMSEP = 6.88 Hz2g2/3	[43]
Tommy Atkins	275	VIS-NIRS	600–1750	PLSR with sensor fusion	Ripening index (RPI) (based on TSS, TA, Fmax)	R2p = 0.832, RMSEP = 0.520 (using full sensor fusion: two probes + two accelerometers)	[44]
Alphonso	76 (43 defective, 33 healthy)	NIRS	673–1900 (analysed: 673–1100)	FLD + Fisher’s ratio + Euclidean distance (feature selection)	Internal defect detection (spongy tissue)	84.5% (at 723.35 nm using Fisher’s criterion); 83.71% (at 722.88–723.82 nm fusion)	[27]
Calypso™	>1000 samples across treatments/seasons	NIRS	729–975 (used for PLSR)	PLSR (Savitzky–Golay + MSC)	Dry matter content (DM), Brix	DM: R2p = 0.82, RMSEP = 0.52% DM-Brix correlation: R2 = 0.72–0.90 across years	[45]
Kweni, Cengkir, Palmer, Kent	186	FT-NIRS	1000–2500	PLSR + MSC, DT	Vitamin C, SSC, total acidity (TA)	Vitamin C: r = 0.86, RPD = 2.00 (with MSC) Raw data: r = 0.82, RPD = 1.75	[46]
Kaew Kamin (Thailand)	105	NIR-HSI	935–1720	PLSR + SG smoothing/original spectrum	TSS and SO2 in dehydrated mango	TSS: R2p = 0.82, RMSEP = 2.42% SO2: R2p = 0.83, RMSEP = 56.40 mg/kg	[47]
Carabao	1200 (green and ripe)	NIRS	700–990	PLSR, PCA-LDA + MSC/SNV	Maturity (DAFI), dry matter (DM), TSS, eating quality (OA)	DM: R2 = 0.774, RMSECV = 1.091% TSS (ripe): R2 = 0.839, RMSECV = 1.282%, RPD = 2.53 DAFI: R2 = 0.946, RMSECV = 2.229 days OA classification: 72–70% accuracy (cal/val)	[48]
Banganpalli	Not explicitly stated (multiple sets A-E)	NIRS	600–1100	PCA, PLS, SPA (with ICP-MS validation)	Detection of artificial ripening via CaC2 (arsenic marker)	PLS: R2 = 0.96 (0–20 ng/g As), 0.90 (51–280 ng/g); RMSE = 0.89–41.65 ng/g	[49]
Nam Dokmai (Si Thong)	120	NIRS	700–2500	PLSR, MLR + SNV, MSC, SG, SG00	β-carotene content in edible part	Long-wave PLSR: R2 = 0.879, SEP = 11.6 RE/100 g EP Short-wave MLR: R2 = 0.812, SECV = 17.63 RE	[50]
Irwin (Green Mature)	23 (14-day storage; sampled at 2-day intervals)	HSI	380–1000	PLSR + second derivative (Savitzky–Golay)	Anthocyanin in skin, SSC in flesh	Anthocyanin: R = 0.88, RMSECV = 2.96 mg/100 g f.w. SSC: R = 0.73, RMSECV = 0.98%	[51]
Kent, Keitt	450+ (Kent and Keitt)	VIS-NIRS	350–2500	PLSR (MSC + SG + derivative), IQI modelling	Internal quality index (IQI), TSS, SIS, firmness, sugar profiles	IQISIS: R2p = 0.729, RMSEP = 0.532, RPD = 1.937 TSS: R2p = 0.639 Firmness: R2p < 0.5 Skin Glucose: R2p = 0.810	[52]
Mahachanok	96 mango purée samples	FT-NIRS	800–2500	PLSR + preprocessing (Min–Max, MSC)	TSS and TA in mango purée	TSS: r2 = 0.955, RMSECV = 0.5, RPD = 4.7 TA: r2 = 0.817, RMSECV = 0.048, RPD = 2.2	[53]
Nam Dokmai (Si Thong)	160 fruits × three sections = 4800 spectra	HSI	450–998	PLSR, MLR, MSC, SNV, SG, SG00	Firmness, TSS, TA	Firmness: R2 = 0.81, RMSE = 2.83 N TA: R2 = 0.81, RMSE = 0.24% TSS: R2 = 0.50, RMSE = 2.0%	[54]
Tommy Atkins	224 samples (four shapes × seven times × eight replicates)	NIR-HSI	951–1630	PLSR, MLR + SNV + second der + mean–centre + RCV	Moisture content (MC)	PLSR (2nd Der + MC): Rp2 = 0.995, RMSEP = 1.121% RCV-MLR (7 wavelengths): Rp2 = 0.993, RMSEP = 1.282%	[55]
Gadung Klonal 21	186 (165 after outlier removal)	NIRS	900–1650	Biresponse local polynomial nonparametric regression (SG + PCA)	pH and TSS	Overall: MAPE = 4.476% Training: MAPE = 3.729% Testing: MAPE = 7.466%	[56]
Cogshall	92	NIRS	600–2300	PLSR + IPLS (backward/stepwise) + SG derivatives	Shelf life, TSS, TA, dry matter, pulp colour after ripening	TSS: RMSEP = 1.1% DM: RMSEP = 1.26% Shelf life: RMSEP = 1.78 days TA: RMSEP = 0.52% PC: RMSEP = 1.86°	[57]
Kent	192	HSI	398–1004	Radial grid correction vs. Lambertian method	Reflectance correction uniformity	Radial grid reduced pixel variation significantly vs. Lambertian (p < 10−126); execution time ~5.53 s (vs. 5.92 s)	[58]
Tommy Atkins	250 (calibration), 22 (monitoring)	NIRS	950–1650	PLSR + SNV/Jack-Knife/PCA/SG derivatives	SS, DM, TA, PF; ripening monitoring	SS: R2p = 0.92, RMSEP = 0.55° Brix DM: R2p = 0.67, RMSEP = 0.51% TA: R2p = 0.50, RMSEP = 0.17% citric acid PF: R2p = 0.72, RMSEP = 12.2 N	[59]
Tommy Atkins, Irwin, Chené, Haden, Joa	240 (168 train, 72 test)	VIS-NIRS	474–1047	PLSR + SG Derivative + interval/wavelength optimisation	Moisture content (MC)	R2p = 0.916–0.987 RMSEP = 3.97–6.61% RPD = 3.48–5.37 depending on treatment	[60]
Cogshall	250 (from three seasons)	NIRS	800–2300	PLSR + Savitzky–Golay + interval PLS (backward/stepwise)	TSS, DM, ATT, flesh colour (hue angle)	TSS: RMSEP = 0.89° Brix DM: RMSEP = 1.23% Colour: RMSEP = 3.16° TA: RMSEP = 5.90 meq/100 g FM	[61]
Nam Dokmai Si Thong	592 (three seasons: 2009, 2012, 2013)	NIRS	700–1100	PLSR + SG, SG00, SNV, MSC + DA (ripeness classification)	TSS, TA, firmness, RPI, ripeness class	TSS: R2 = 0.90, SEP = 1.2% Firmness: R2 = 0.82, SEP = 4.22 N TA: R2 = 0.74, SEP = 0.38% RPI: R2 = 0.80, SEP = 0.80	[62]
Keitt, Kent	529 (final used); 540 initially	VIS-NIRS	684–990	irPLS, irCovSel, PLS (SG derivative)	Dry matter content (DM%)	RMSEP (irCovSel) = 0.89% (on unseen cultivars + new instrument) RMSEP (PLS) = 2.06% (same test set); Bias reduction: 3.40 → 0.05%	[63]
Harivanga (Bangladesh)	120	SW-NIRS	900–1650	PLSR + SG second derivative + SPA/RC	Soluble solid content (SSC, %Brix)	SPA-PLS: rp = 0.78, SEP = 0.67%, RPD = 2.12, RER = 8.93 Full model: rp = 0.74, SEP = 0.78%, RPD = 1.78	[64]
Kent and Palmer	62 (various ripenesses)	FT-NIRS	1000–2500	PLSR + EMSC, PCR	Vitamin C (mg/100 g FM)	PLSR + EMSC: r = 0.99, RMSE = 1.33, RPD = 5.40 PCR + EMSC: r = 0.92, RMSE = 5.26, RPD = 1.36	[65]
Tainung No.1	150 (90 Na-I, 60 In-I)	VIS-NIRS	400–1000	PLS-DA, PCA + multi-omics correlation	Anthracnose disease (early detection)	In-I (early): 100.00% accuracy Na-I (early): 89.92% accuracy	[66]
Tommy Atkins	162 (two batches × 81)	HSI	400–1000 and 880–1720	PLSR + RC/SW/CARS for wavelength selection	Moisture content (MC)	Best model (RC-PLS-2): R2p = 0.972, RMSEP = 4.611%	[67]
Golek (China)	232 spectra (116 mangoes × 2)	NIRS	317–1115 nm (common: 350–1115)	PLSR + REA + calibration transfer (SST, SBC, PDS)	Soluble solids content (SSC)	Best Model (SST + REA): RMSEP = 1.243%, R2 = 0.894, 48.91% RMSEP reduction over unoptimised transfer	[68]

. For instance, PCA and ANOVA-Simultaneous Component Analysis (ASCA) models were used to study the effect of drying on chlorophyll, carotenoids, water, and sugar, revealing significant batch and dryer effects, with Visible Near-Infrared (VIS-NIR) spectroscopy showing 47.5% batch and 23.6% dryer influence, while NIR showed 38.3% and 29.4%, respectively [31]. In Calypso, Honey Gold, and Keitt varieties, a global PLSR model with enhancements like External Parameter Orthogonalisation (EPO) and bias correction achieved an RMSEP of 1.05% and R² of 0.82 for DMC prediction [32]. Nam Dok Mai Sithong mangoes showed promising results using PLSR for soluble solids prediction, achieving RMSEP = 0.765% and r = 0.74 in prediction [33], while the moisture content and drying uniformity in Tommy Atkins were evaluated with very high accuracy (R² = 0.995) using HSI data and PLSR [34].

Nam Dok Mai was frequently studied across multiple datasets. For example, spectra collected with SCiO and Linksquare devices were used to develop PLSR models for the prediction of DM, TSS, TA, and pH, with R² values ranging from 0.74 to 0.93 [23]. A combination of NIRS and HSI data was used to develop models for the firmness index, β-carotene, and Ripening Prediction Index (RPI), with R² values over 0.84 and RPDs exceeding 3.0 reported [22]. Another large-scale NIRS study of Nam Dok Mai over two time periods showed a high prediction performance for TSS (R²p = 0.81), TA (R²p = 0.84), and pH (R²p = 0.80), though the firmness prediction was weaker (R²p = 0.09) [24].

A number of studies reported use of disease and defect detection. Alphonso mangoes were effectively classified for spongy tissue using PCA + SIMCA, with an accuracy of over 96% in the 900–970 nm range [35]. Defect detection via HSI in unlabelled varieties achieved 96.67% overall accuracy using band ratio and Otsu thresholding methods [37]. Irwin mangoes from Japan showed moderate to good predictive power for SSC and skin colour with R²cv = 0.76 and 0.78, respectively [38]. For Nam Dok Mai Si Thong, PLSR and PCA-based fruit quality index modelling (FQI1 and FQI2) achieved Q² values above 0.84 [39] and early detection of anthracnose disease in Tainung No.1 mangoes achieving up to 100% classification accuracy using PLS-DA and PCA with multi-omics correlation [66].

Numerous studies focused on key nutritional traits. For example, Kent mangoes showed exceptional R² values of 0.976 (TA) and 0.958 (Vitamin C) using MSC and BLC enhanced PLSR models [29], while in Cengkir, Kweni, Kent, and Palmer varieties, PLSR + EMSC models predicted TSS and Vitamin C with r = 0.86 and RPDs over 2.0 [41]. In Guifeimang, an HSI-based PLSR model reduced the band usage to just 11.3%, yet maintained a high prediction accuracy (R²p = 0.90) for SSC [36]. HSI-based predictions of firmness, TSS, TA, and chroma from a study in Tianjin achieved R² values ranging from 0.65 to 0.94 [42]. Various regression methods also demonstrated the ability to predict vitamin C (r = 0.99), anthocyanin, firmness, and SSC with high precision across Kent, Palmer, Cogshall, Ma-hachanok, and Gadung Klonal 21 mangoes [46,51,53,56,65].

Another group of studies considered the prediction of the ripeness level. Ripening progression in Kent mangoes was evaluated with iPLSR, resulting in R²p = 0.75 for firmness prediction [43]. Tommy Atkins mangoes were also evaluated with sensor fusion techniques achieving R²p = 0.832 for the ripening index [44]. The detection of artificial ripening in Banganpalli mangoes was achieved using PLSR with arsenic markers (R² = 0.96) [49].

4.3.2. Conventional Machine Learning Approaches

ML techniques have been widely applied for mango quality assessment using NIRS and HSI, covering spectral ranges from 306 to 2500 nm as shown in

Table 2. Summary of conventional machine learning algorithms.

Mango cv	Sample Info	NIRS/ HSI	Spectral Range (nm)	Algorithm Used	Quality Trait	Accuracy Achieved	Ref.
Mango (imported from Brazil and Spain, scanned in Germany)	58	NIRS	1000–2500	BPSO + PLSR/SVR (with preprocessing and feature selection)	Total acidity (mg/100 g), vitamin C (mg/100 g)	Acidity: R2cv = 0.93, R2test = 0.97, RMSEP = 17.40% Vitamin C: R2cv = 0.66, R2test = 0.46, RMSEP = 0.848%	[72]
Namdokmai Sithong (Thailand)	104	NIRS	800–2500	PLS-DA, ANN (with SNV, MSC, derivative)	Early detection of anthracnose	ANN: 98.1% at 24 h. PLS-DA: 95.2% at 24 h. ANN reached 100% by 96 h. PLS-DA: R2test = 0.676 (SNV), RMSE test = 0.280	[73]
Multiple cultivars (Calypso™, KP, HG, R2E2, etc.) across four seasons	4675	NIRS	684–990 (optimised)	PLSR (global and cultivar-specific), ANN	Dry matter content (DMC)	ANN: RMSEP = 0.89% (global) PLSR: RMSEP = 0.86% (specific), 1.01% (global)	[74]
Multiple cultivars (Calypso™, KP, HG, R2E2, etc.) across four seasons	4675	NIRS	684–990	ANN, GPR, LPLS, LPLS-S, LOVR, MBL, Cubist, SVR, LOCAL, DataRobot, Hone Create	Dry matter content (DMC)	Best Individual: LOVR (RMSEP = 0.881%), Best Ensemble: ANN + GPR + LPLS-S (RMSEP = 0.839%), Global PLSR = 1.014%	[75]
Keitt, Haden, Local (Ghana)	198	NIRS	740–1070	SVM, LDA, RF, NN, LDA-SVM for classification; IPLS, Bi-PLS, Si-PLS for regression	TSS (° Brix), pH, variety identification	LDA-SVM: 100% (train), 97.44% (test) classification accuracy; Si-PLS (TSS: R2 = 0.63, RMSEP = 1.83) Si-PLS (pH: R2 = 0.81, RMSEP = 0.49)	[70]
‘Keitt’, ‘Osteen’	200 (100 each)	HSI	450–980	MLP, SLP, QDA, RF, XGB; feature selectors: SLP4FS, SLDA, SQDA, PCA4FS, etc.	Early anthracnose detection	MLP (Keitt): Accuracy = 96.1%, Recall = 96.1%, MCC = 0.953 MLP (Osteen): Accuracy = 97.5%, Recall = 97.6%, MCC = 0.971 (within 48 h)	[16]
Kent	91 (56 calibration, 35 prediction)	NIRS	1000–2500	PLSR + spectra preprocessing (MSC, SNV, OSC, etc.)	TA, SSC	TA: R2pred = 0.72, RMSEP = 52 mg/100 g FM, RPD = 1.9 SSC: R2pred = 0.76, RMSEP = 0.6 Brix, RPD = 1.8	[26]
Hard green and ripe mangoes	11,691 samples (9711 hard green + 1980 ripe)	NIRS	462–1032	Gaussian process regression (GPR), PLSR	Dry matter content (DMC)	GPR (ID set): R2 = 0.91, RMSE = 0.69 PLSR (ID set): R2 = 0.88, RMSE = 0.80 GPR + confidence (OOD RMSE reduced by ~62%)	[69]
Nam Dok Mai Si Thong (Thailand)	600 spectra (30 fruits × five time points × two sides × two trials)	HSI	400–1000	PCA + SVM (Gaussian kernel)	Anthracnose infection severity	SVM accuracy = 99.6%; d0, d2, d6, d8 TPR = 100%, d4 TPR = 98%	[76]
Succarri	75	HSI	302–1148	ANN, RF, DT with new and published SRIs	SPAD, TSS, firmness	SPAD: R2 = 0.98–0.99 (test); TSS: R2 = 0.88–0.93; Firmness: R2 = 0.98–0.99; best results by RF and DT, with MSE as low as 0.03	[30]
Cengkir, Kweni, Kent, Palmer	244 (186 + 58 samples)	NIRS	999.9–2500.2	DT, LR, SVC, RF, ETC, stacking (RF meta) + PCA + SMOTE + SG, MSC, SNV preprocessing	Vit-C, titratable acidity (TA), mango varieties	Vit-C: 95.0%, TA: 83.0%, Mango Varieties: 100.0%, 5-fold average: 98.0% accuracy using stacking classifier	[77]
Mango (flesh only, peeled, flat cut)	Five real mangoes (for validation) + 40 simulated samples	HSI	400–1000	Step-by-step inverse algorithm + Monte Carlo modelling + M − 1-step baseline method	Optical properties: absorption (ma), reduced scattering (m′s)	Step-by-step: ma = 9.2%, m′s = 5.7% MAPE (real mangoes); M − 1-step: ma = 3.8%, m′s = 3.7% (simulation, 0.1 mm spatial resolution)	[78]
Namdokmai Sithong	64 mangoes, 1112 grid areas (792 intact, 230 IBD, 90 BSV)	NIRS	800–2500	LDA (SLDA), ANN (non-linear classifier)	Internal breakdown (IBD) and black-streaked vascular tissue (BSV)	LDA: 86.25% (SNV preprocessed) ANN: 91.37% (MSC preprocessed)	[79]
Samar Bahisht Chaunsa and Sufaid Chaunsa	240 (120 per cultivar, two seasons)	NIRS	729–975 nm (selected)	PLSR, MLR, ANN, SVM (for regression); KNN, SVM, LDA, ANN, tree, ensemble (for classification)	Dry matter (DM)/maturity (binary)	Direct Classification: 88.2% (KNN/SVM/Ensemble) Indirect Estimation: 55.9% (MLR)	[80]
11 mango varieties (e.g., Aeromanis, Jaffro, Irwin, Kent, etc.)	220 fruits (20 per variety × four slices)	HSI	400–1700 (Vis-NIR + NIR)	ANN, KNN, LDA + covering array feature selection (CAFS)	Varietal identification	NNFC (ANN): Accuracy = 98.2% (full Vis-NIR), 97.2% (optimised); KNN: up to 91.7%; LDA: 83.1% (NIR), 87.4% (Vis-NIR)	[81]
Kent	90 (from Spain, Brazil, Peru)	NIRS	1000–2500	PLSR, SVMR, ANN + SNV preprocessing	Total acidity (TA)	ANN: R2cal = 0.985, R2pred = 0.943, RMSEC = 25.29, RMSEP = 28.42 mg/100 g, RPD = 4.02	[82]
Tainong and Jin Huang Awn	95	NIRS	1300–2300	BP-PLS and SA-BP-PLS (Simulated Annealing optimised NN)	Brix (sugar)	SA-BP-PLS: R2 = 0.9854, RMSE = 0.0431 BP-PLS: R2 = 0.906, RMSE = 0.219	[83]
Kent	60 mangoes, 2880 ROIs (252,801 spectra)	HSI	450–980	QDA, LDA, PLSDA + dimensionality reduction (Pearson, PCA, Tukey)	Anthracnose disease stage	QDA (full spectrum): 90.9% accuracy, R2 = 0.985, RMSE = 0.043; with only 27 bands: 87.6%; with 20 bands: 79.8%	[84]
Keitt	576 mangoes (946 spectra)	VIS-NIRS	400–1000	ANN, PLS (regression + classification)	Internal browning	ANN classification: 83.1% (test), 87.1% (calibration), 82.3% (extra test); ANN regression: R2 = 0.57 (vs. PLSR R2 = 0.53)	[85]
Keitt	141 mangoes (two harvests, 282 readings)	VIS-NIR	550–650 (selected range)	Logistic regression, LDA, SVM, functional data model, random forest	Internal physiological disorders (jelly seed and black flesh)	LDA (after storage): Accuracy = 76%, Sensitivity = 78%, Specificity = 73% Logistic (at harvest): Accuracy = 65%, Sensitivity = 78%, Specificity = 49%	[86]
Keitt	120 (90 calibration, 30 prediction)	HSI	475–1100	SVR, ELM, BPNN + feature reduction (UVE, SPA, RF, CARS, (CARS + RF)-SPA, etc.)	Dry matter (DM)	BPNN with (CARS + RF)-SPA: R2C = 0.971, R2p = 0.966, RMSEC = 0.142, RMSEP = 0.153	[71]
Keitt	120 (T1, T2, T3 stages)	NIR	900–2500	LS-SVM + filters (FIR, ME, GS) + variable selection (SPA, CARS)	FI, DMC, SSC, TA	Best RPDs: FI = 3.05 (PE-GS), DMC = 2.31 (EPE-GS), SSC = 2.61 (PVC-FIR), TA = 2.94 (PVC-ME)	[87]
Cat Hoa Loc (Vietnam)	106 total (67 training, 39 testing)	Vis-NIRS	410–940	RPR (RF → PLS → RF), compared with SVM and multinomial logistic regression	Sweetness (Brix-based grading: I (>24° Brix), II (20–24), III (<20))	RF classifier: Training acc. = 100%, Testing acc. = 82.1%, DI = 0.287 SVM: 66.7%, MLR: 61.5%	[88]
Calypso and Kensington Pride	240 fruits (120 per cultivar)	HSI	400–1000	PLSR, ANN, SVMR (with OSC, SG1, SNV, PCA preprocessing)	Brix, acidity, Brix/acid ratio, 13 nutrients	PLSR (Brix): R2 = 0.89 (skin), 0.74 (flesh); SVMR (Brix): R2 = 0.77 (skin), 0.73 (flesh), RPD > 2.4; ANN (Brix): R2 = 0.78–0.90; acidity and Brix/acid also well predicted	[89]
Mango (three ripeness classes: unripe, ripe, overripe)	80 mangoes (70% train, 15% val, 15% test)	HSI	400–1000	Shared subspace tensor classification (SSTC) via partial Tucker decomposition + RF, XGBoost, SVM, logistic regression	Ripeness classification	SSTC + RF: 92% accuracy SSTC + XGBoost: 84% SSTC + Logistic/SVM: 75% HSCNN: 75% Flattened + RF: 58%	[90]
Mango mesocarp	11,690 samples	NIRS	309–1149	St-SNE with multi-attribute info (region, maturity type, cultivar); compared with PCA, LPP, t-SNE, UMAP, Fisher t-SNE	Sample similarity (clustering)	St-SNE: Highest classification accuracy (e.g., up to 89.47% on cultivar), best trustworthiness and neighbourhood preservation	[91]
Keitt	78 mangoes, 156 samples (two sides each)	HSI	400–890	Support vector regression (ε-SVM), genetic algorithm (GA), brute force (BF), RGB-SVM, linear regression	Dry matter (ripeness proxy)	Full HSI: R2 = 0.74 Best 4-filter multispectral GA: R2 = 0.69 Best RGB + filter: R2 = 0.63 (real), 0.61 (simulated) Best COTS 4-filter GA: R2 = 0.66	[92]
Multiple cultivars (Tommy Atkins, Palmer, Keitt, Kensington Pride, etc.)	30–1200 per study (varies)	VIS-NIRS/FT-NIRS	306–2500	PLS, MLR, PCR, ANN, SVM (with preprocessing: SG, SNV, MSC, EMSC, 1st/2nd derivative, PCA, etc.)	SSC, dry matter, firmness, TA, pH	R2 up to 0.98 (Harumanis SSC), 0.95 (Tommy Atkins Acidity), 0.93 (Sunshine pH), 0.92 (Tommy Atkins SSC), 0.53–0.57 for internal browning (ANN > PLS)	[93]
Arumanis	175 mangoes (696 spectra across five maturity classes)	NIRS	1350–2500	LDA, SVM, MLP, DT, KNN; indirect: PLS + fuzzy logic with TA, SSC, starch, firmness	Maturity classification across five levels (80–100%)	Direct (LDA + SAVGOL): 91.43% Indirect (PLS + Fuzzy Logic): 95.7% accuracy using TA, SSC, firmness, starch as inputs	[94]
Nam Dokmai	173 (129 calibration, 44 validation)	NIRS	860–1760	GA-PLS, SPA-PLS, full-spectrum PLS; preprocessing: SNV, first/second derivative (D1, D2)	Soluble solids content (SSC)	Best model (GA-PLS + D2): R2 = 0.72, RMSEP = 0.74° Brix, RPD = 2.0 Full PLS + D2: R2 = 0.74, RMSEP = 0.72° Brix, RPD = 2.0	[95]

. A study has demonstrated that Gaussian Process Regression (GPR) yielded high accuracy with R² = 0.91 and RMSE = 0.69 for predicting DMC in large datasets of hard green and ripe mangoes [69]. Similarly, hybrid models such as ANN + GPR + LPLS-S achieved the lowest RMSEP of 0.839%, indicating robust performance across multiple cultivars [70]. Further enhancement was observed with feature optimisation approaches; for example, BPNN combined with CARS + RF-SPA achieved R²p = 0.966 and RMSEP = 0.153 [71].

Sugar content and acidity have also been successfully predicted using ML methods. For Kent mangoes, a PLSR model reached R²pred = 0.76 for SSC and 0.72 for TA, with RPD values near 2.0, suggesting moderate predictive reliability [26]. Calypso and Kensington Pride cultivars demonstrated R² values of up to 0.89 for Brix content on skin using PLSR, while ANN models ranged between 0.78 and 0.90 [89]. Disease detection, particularly anthracnose, achieved remarkable performance using neural networks. A Multilayer Perceptron (MLP) model reached 96.1% accuracy for the ‘Keitt’ cultivar and 97.5% for ‘Osteen’ within 48 h post-infection [16]. Another study using SVM with PCA features reported 99.6% accuracy in early-stage anthracnose detection for Nam Dok Mai Sithong mangoes [72].

Internal disorders, such as internal breakdown (IBD), black-streaked vascular tissue (BSV), and browning, have been targeted using classification models like ANN and LDA. For instance, ANN achieved 91.37% accuracy in detecting IBD and BSV using MSC preprocessed data [79], while LDA showed 76% accuracy in detecting jelly seed and black flesh disorders post-storage [86]. ANN also outperformed PLSR in the regression-based detection of internal browning, achieving R² = 0.57 [85]. Ripeness and maturity stages were effectively predicted using advanced models. SSTC combined with random forest reached 92% classification accuracy across three ripeness stages [90]. Arumanis mangoes were classified across five maturity levels with 91.43% accuracy using LDA and up to 95.7% using an indirect fuzzy logic system based on SSC, TA, starch, and firmness [94].

Varietal and fruit classification studies demonstrated the robustness of ensemble and hybrid approaches. For instance, an ANN model using full Vis-NIR data achieved 98.2% accuracy in identifying 11 mango varieties [81], while a stacking classifier combining DT, LR, SVC, RF, and ETC yielded 100% accuracy for varietal classification among cultivars like Cengkir, Kweni, Kent, and Palmer [77]. Across these studies, traditional algorithms like PLSR, SVM, and LDA were frequently used, but enhanced performance was observed with optimisation techniques such as Genetic Algorithms, Simulated Annealing, and novel feature selection methods like CARS, SPA, and UVE. Dimensionality reduction tools like PCA, CAFS, and St-SNE were also essential for model generalisation.

In conclusion, ML approaches offer powerful solutions for non-destructive mango quality assessment. High prediction accuracy, especially from ensemble and feature-optimised models, affirm their potential in real-time agricultural applications. However, challenges remain in terms of their generalisability across different mango cultivars and environmental conditions, underlining the need for further validation studies.

4.3.3. Deep Learning Approaches

Table 3. Summary of deep learning models.

Mango cv	Sample Info	NIRS/ HSI	Spectral Range (nm)	Algorithm Used	Quality Trait	Accuracy Achieved	Ref.
Multiple cultivars (four seasons)	4676 samples	NIRS	684–990	CNN (Fine-tuning), Global Model (CNN/PLS), Recalibration (PLS), S/B Correction	Dry Matter Content (DMC)	CNN Fine-tuning: RMSE = 0.642, R2 = 0.907	[96]
Multiple cultivars (four seasons)	11,691 spectra (4675 fruits)	NIRS	684–990	1D-CNN + Chemometrics (Outlier Removal + SNV + Derivatives), PLS	Dry Matter Content (DMC)	RMSEP: 0.75% (1D-CNN with outlier removal and augmented data), 0.79% (1D-CNN without outlier removal)	[28]
Multiple cultivars (four seasons)	11,691 spectra (4675 fruits)	NIRS	684–990	Shallow CNN, Deep CNN, PLSR with MVN and Bjerrum-style Augmentation	Dry Matter Content (DMC)	Best DMC RMSE: 1.16% (PLSR MVN augmented); 1.20% (Shallow CNN MVN augmented)	[97]
Multiple cultivars (four seasons)	11,691 spectra (4675 fruits)	NIRS	742–990	PLSR vs. 1D-CNN (Raw and Preprocessed Spectra)	Dry Matter Content (DMC)	DL (raw absorbance): RMSEP = 0.76%, PLSR (raw absorbance): RMSEP = 0.87%	[25]
Old (2015–2018) + New (2020 Brazil)	11,691 spectra from four seasons (2015–2018) for training; 510 samples (2020) for test	NIRS	684–990	Deep Learning (1D-CNN), Transfer Learning (TL), PLS	Dry Matter Content (DMC)	DL (after TL): RMSEP = 0.518% PLS (best): RMSEP = 0.598%	[6]
Hainan mangoes (bruise detection)	92 samples × four storage timepoints	HSI	398.9–1015.6	CNN + Feature Fusion (Spectral + Texture GLCM), CARS/UVE Selection vs. RF, PLS-DA, XGBoost	Storage Time Post Mild Bruise	CNN + CARS + Feature Fusion 2: 93.48% accuracy	[98]
Multiple cultivars (four seasons)	11,834 spectra (4685 fruits)	NIRS	684–990 (from 350–1100)	CNN, ANN, PLS (144 Experiments)	Dry Matter Content (DMC)	Best CNN RMSEP: 0.77% FW (season four test); RMSEP: 1.18% on unseen season five	[7]
Bundaberg (Calypso, cultivar B74)	494 trees, multiple blocks (ground-based)	HSI	390–890	CNN (Per-pixel Spectral Classifier); CHC Optimiser for Fruit Segmentation + Yield Estimation	Fruit Count (Yield Estimation)	R2 = 0.79 vs. manual count (18-tree test), R2 = 0.83 vs. RGB count (216-tree test)	[99]
Bundaberg (Calypso, cultivar B74)	494 trees (Block A); +121 (Block B); +266 (Block C); fruit-on-tree (n = 662) and fruit-in-tray (n = 468)	HSI	411.3–867.0	CNN-COMB, CNN-PLS, PLS Baseline	Maturity (Orchard-Scale DMC Mapping)	Fruit-on-tree (CNN): R2 = 0.64, RMSE = 1.08% w/w Repeatability RMSE: ≤0.29% w/w	[100]
Keitt mangoes (three ripening stages)	230 packaged mangoes, 2760 spectra	NIRS	900–2500	PLSR, PCR + FNN Correction + GS Filtering	FI, DMC, SSC, TA (Packaged Mango)	Best PMs-FNN-GS: FI: R2 = 0.847, RMSEP = 10.705 N DMC: R2 = 0.932, RMSEP = 0.320% SSC: R2 = 0.821, RMSEP = 1.211% TA: R2 = 0.907, RMSEP = 0.032%	[101]

presents the analysis conducted on using a DL to examine various properties of mangoes. DL has emerged as a powerful and preferred technique in mango quality assessment, particularly when combined with NIRS and HSI. Its ability to automatically extract hierarchical and abstract features from spectral data makes it especially effective in handling the complexity and high dimensionality of such datasets [6]. For instance, one study [96] introduced CNN fine-tuning as an effective method for predicting DMC across multiple seasons, achieving an RMSE of 0.642 and R² of 0.907. Similarly, another study [28] incorporated spectral preprocessing and CNNs to improve prediction robustness, yielding an RMSEP of 0.75%. In the same vein, the authors of [97] demonstrated that MVN augmentation enhanced CNN model stability, achieving an RMSE of 1.20%. Supporting these findings, the research in [25] confirmed that CNNs trained on raw spectra outperformed PLSR models applied to preprocessed data, with an RMSEP of 0.76%. The authors of [6] further emphasised model generalisation across devices and seasons, reporting the best performance (RMSEP = 0.518%) through transfer learning using CNNs.

In terms of broader applications, the study in [96] addressed early bruise detection in Hainan mangoes using HSI and CNNs, achieving a classification accuracy of 93.48%. The authors of [7] benchmarked CNN, ANN, and PLS across 144 configurations, confirming CNN’s superior performance (RMSEP = 0.77%). Yield estimation was explored in [99], where CNN-based spectral classifiers combined with morphology-based fruit counting reached R² = 0.83. The work in [100] demonstrated orchard-scale DMC mapping using a UGV with HSI sensors and CNN-COMB model, achieving R² = 0.64 and RMSE = 1.08% with high repeatability. Finally, the authors of [101] tackled spectral distortion from paper packaging in pre-harvest mangoes and employed FNN + GS filtering to restore accurate predictions for FI (R² = 0.847), DMC (0.932), SSC (0.821), and TA (0.907). Collectively, these studies validate deep learning’s potential for high-throughput non-destructive mango quality monitoring in both controlled and field environments.

4.3.4. Transformer-Inspired Approaches

Transformer-inspired models, particularly those integrating attention mechanisms, are gaining momentum in non-destructive mango quality assessment (

Table 4. Summary of transformer-inspired models.

Mango cv	Sample Info	NIRS/ HSI	Spectral Range (nm)	Algorithm Used	Quality Trait	Accuracy Achieved	Ref.
Calypso™, KP, Honey Gold, Keitt, R2E2, Lady Grace, Lady Jane, 1201, 1243, 4069	>10,000	NIRS	300–1100	Two-stream DL (1D-CNN + BiGRU + XGBoost)	Variety classification	95%	[102]
Kweni, Cengkir, Palmer, Kent	186 (original) + 2000 (augmented)	NIRS	1000–2500	MCNN (CNN + Channel Attention)	Variety classification	98.67%	[103]

). These models enhance deep learning architectures by enabling selective focus on the most informative spectral features. For instance, in [102], a novel two-stream deep learning model for mango variety classification using NIRS, combining 1D-CNN and Bidirectional Gated Recurrent Units (BiGRU), was presented. The 1D-CNN extracts local spectral features, while the BiGRU capture long-range dependencies and positional information across spectral sequences. To further enhance feature representation, the model incorporates a lightweight attention mechanism and a Feature Pyramid Network (FPN) with multi-scale convolution kernels. These modules are fused using a feature alignment and interpolation strategy, followed by either a Fully Connected Network (FCN) or XGBoost for classification. Among all compared models, including traditional machine learning algorithms and deep learning approaches, the proposed two-stream model achieved the highest classification accuracy for mango varieties—85% with FCN and 95% with XGBoost. Ablation studies confirmed the effectiveness of each module, particularly the attention mechanism, BiGRU, and FPN, in improving classification performance on complex high-dimensional NIRS data.

Similarly, ref. [103] employed a CNN to classify mango varieties using NIRS data but enhanced the model with a channel attention mechanism to focus on the most informative spectral features. This model, referred to as MCNN, integrates parallel convolutional layers, residual connections, and an SE (Squeeze-and-Excitation) block to improve feature extraction and classification performance. By dynamically adjusting channel weights, the model suppresses less relevant features and emphasises key spectral information, leading to a highly accurate non-destructive classification system. Their approach achieved an accuracy of 98.67%, outperforming traditional deep learning and machine learning models and confirming the effectiveness of attention-based deep learning for mango variety discrimination using NIRS.

4.4. Quality Traits Used for Mango Quality Assessment

NIRS and HSI have become essential tools in the non-destructive evaluation of internal quality traits in mangoes. These techniques provide rapid and reliable assessments of key parameters such as firmness, SSC, DMC, and TA, all of which are critical indicators of mango ripeness, sweetness, and market quality. A comparative study found that NIRS outperformed HSI in predicting most internal traits such as firmness, TSS, and TA, whereas HSI showed better performance for β-carotene and pH due to its inclusion of the visible wavelength range [22].

In another advancement, a ground-based HSI system was developed for mango yield estimation. The system achieved R² values of up to 0.75 when validated with field counts and 0.83 with RGB counts. Although RGB provided slightly better accuracy, HSI presented itself as a reliable alternative, particularly valuable when simultaneously used for assessing traits like disease presence or fruit maturity [99].

Additionally, ref. [39] proposed a flexible non-destructive method for evaluating overall fruit quality through the development of two novel Fruit Quality Indices, FQI1 and FQI2. These indices consolidate multiple physicochemical attributes, including TSS, TA, firmness, pH, DM, Moisture Content (MC), and lipid content, into a single score ranging from 0 to 100. FQI1 is calculated using the geometric mean of scaled parameters, while FQI2 employs a weighted arithmetic mean with weights determined via PCA. The study further combined NIR spectroscopy with PLS regression to predict these indices in mango, tangerine, and avocado samples, offering a fast and robust alternative to traditional destructive fruit quality evaluation methods.

5. Results and Discussion

Based on research questions, the results have been analysed to address key objectives such as the effectiveness of NIRS and HSI in assessing mango quality, the performance of different analytical techniques, and the accuracy of the models used for classification and prediction. The analysis highlights the strengths and limitations of each approach in relation to the research goals.

5.1. Distribution Based on the Year of Publication

Figure 2 illustrates the distribution of algorithms used in 85 research papers from 2015 to 2025. The number of papers shows a general upward trend, peaking in 2024 with 16 publications, noting that the apparent decline in 2025 is due to incomplete indexing of the most recent year. The rise in in publications reflects advancements in deep learning and non-destructive techniques, with improved accessibility of NIRS and HSI instruments significantly boosting interest in applying these technologies to mango quality assessment [104].

5.2. Distribution of NIRS and HIS

Figure 3 represents the distribution of different spectral imaging techniques, NIRS, HSI, Vis-NIRS, and combined NIRS + HSI, used across 85 research papers. Most studies (63%) employed NIRS, indicating its popularity and effectiveness in various research applications, particularly in food and agricultural analysis. HSI accounted for 25% of the papers, reflecting its growing role in capturing spatial and spectral data simultaneously. This shows that while NIRS remains dominant, HSI is also becoming a prominent tool in non-destructive assessment and classification tasks.

Additionally, 11% of the studies used Vis-NIRS, combining the visible and near-infrared spectrum, which provides extended information for more comprehensive analysis. Only 1% of the papers utilised a combination of both NIRS and HSI techniques, suggesting that although hybrid approaches have potential, they are still underexplored in the current research. Overall, this distribution indicates that while NIRS continues to be the most preferred technique, there is an increasing interest in more complex or integrated spectral methods for enhanced performance and accuracy.

5.3. Distribution of Mango Quality Traits Assessed

Figure 4 illustrates the distribution of quality traits assessed using algorithms across 85 research papers. TSS is the most frequently studied trait, appearing in 28 papers, followed by TA with 22 papers and DMC in 20 papers. FI also received notable attention, with 11 papers focusing on this physical quality parameter, essential for evaluating the fruit texture and shelf life. Other traits such as anthracnose disease detection (five papers), variety classification (six papers), and vitamin C content estimation (six papers) were comparatively less studied.

5.4. Distribution Based on Analysis Method

Figure 5 depicts the distribution of algorithm types employed across 85 reviewed papers. Statistical methods are the most widely used, appearing in 43 studies, which highlights their foundational role in data analysis and pattern recognition, particularly in early and classical approaches. ML algorithms are also significantly represented, with 30 papers indicating a strong shift towards more adaptive and data-driven modelling techniques in recent research. These methods offer improved predictive performance and have become increasingly popular with the rise of computational resources and structured datasets.

DL approaches were used in 10 papers, showcasing their growing application, especially for tasks involving complex data, such as images and spectral information. Notably, only two papers implemented deep learning models integrated with attention mechanisms, indicating that such advanced architectures are still in the exploratory phase within this field. This distribution reflects a progressive trend from traditional statistical models toward machine learning and deep learning approaches, with attention-based models poised to gain traction as the field evolves and more data become available.

5.5. Summary of Survey

The answers to the research questions are addressed as follows.

• RQ1: How are NIRS and HSI Data Collected for Mango Quality Assessment?

Most datasets were collected using a variety of commercial spectrometers available on the market. Some studies aimed to develop and use low-cost spectrometers to make technology more affordable. Data collection environments included laboratory settings, in-field scenarios using Unmanned Ground Vehicles (UGVs), and manual acquisition using handheld spectrometers.

• RQ2: What preprocessing and dimensionality reduction techniques are used for NIRS and HSI data analysis?

A variety of preprocessing and dimensionality reduction techniques are employed to enhance the performance of NIRS and HSI models. Common spectral preprocessing methods include Standard Normal Variate (SNV), first and second derivatives, MSC, and BLC, which are used to reduce noise, correct for scatter effects, and highlight important spectral features. Some studies also employ stacking of multiple preprocessing methods to augment spectral data and improve model robustness, especially in deep learning applications such as 1D-CNN.

In addition, feature selection techniques, such as Fisher’s criterion have been applied to identify optimal wavelength ranges, such as 702.72–752.34 nm, for effective classification tasks. SRIs are also developed to reduce dimensionality and extract meaningful features correlated with key quality traits. Notably, some studies suggest that using raw spectral data may outperform preprocessed data, particularly when employing deep learning models, as raw spectra retain both absorption and scattering information, which are critical for accurate predictions.

• RQ3: How effective are attention-augmented deep learning and transformer-based models for non-destructive mango classification and grading compared to traditional statistical and machine learning methods?

Statistical models such as PLSR remain widely popular for analysing spectral data due to their simplicity and interpretability. ML approaches, particularly stacked models, have demonstrated strong performance by combining the strengths of multiple algorithms. DL algorithms have set benchmark results in spectral data analysis by effectively capturing complex non-linear patterns in high-dimensional datasets. More recently, transformer-inspired architectures have emerged as powerful alternatives, addressing some limitations of traditional DL methods by focusing attention on the most relevant features in the spectral data. Recent studies have shown that transformer-inspired architectures achieve excellent results in spectral data analysis, demonstrating their ability to effectively capture relevant features and improve model performance.

• RQ4: What quality traits are assessed in mangoes using NIRS and HSI data?

NIRS and HSI are widely used for the non-destructive assessment of key internal quality traits in mangoes. These techniques effectively evaluate attributes such as the firmness, SSC/TSS, DMC, TA, pH, MC, and β-carotene levels. These traits are essential indicators of mango ripeness, sweetness, and overall market quality. NIRS generally outperforms HSI in predicting internal traits like firmness, SSC, and TA, while HSI shows superior performance for attributes influenced by visible wavelengths, such as β-carotene and pH. Additionally, composite indices like Fruit Quality Indices (FQI1 and FQI2) have been developed by combining multiple physicochemical parameters and can also be accurately predicted using NIR and PLS regression models, further enhancing the efficiency of quality assessment.

• RQ5: What are the key challenges and future opportunities in applying advanced machine learning and deep learning techniques for mango grading using NIRS and HSI data?

The research has revealed several challenges and opportunities, as outlined below. The review revealed several technical and practical challenges:

• Limited dataset sizes and lack of public databases, which constrain the training of robust deep learning models.
• High dimensionality of hyperspectral data, which requires extensive preprocessing and advanced modelling to prevent overfitting.
• Environmental variability, lighting, temperature, and seasonal differences that can affect spectral measurements and reduce model transferability.
• Lack of standardisation in spectral acquisition, preprocessing, and validation methods across studies.

Few papers incorporated attention mechanisms inspired by transformer models, and no paper implemented full transformer architectures like ViT, indicating a clear research gap.

Future opportunities include the following:

• Adoption of transformer-based models for spectral feature extraction and sequence learning.
• Development of lightweight real-time AI models integrated with edge devices for orchard or supply-chain deployment.
• Use of multi-modal sensor fusion, combining NIRS, HSI, RGB, and thermal data, for comprehensive fruit quality profiling.
• Implementation of transfer learning and domain adaptation to improve model robustness across varying field conditions.

By addressing these gaps, researchers can enhance the scalability and practicality of intelligent mango quality assessment systems.

6. Conclusions

This comprehensive review of 85 research articles highlights the significant advancements and ongoing challenges in applying NIRS and HSI for non-destructive mango quality assessment. The studies reveal that most datasets were collected using a range of commercial and custom-built spectrometers across various settings, including laboratories, field-based platforms like UGVs, and handheld systems.

A wide array of spectral preprocessing and dimensionality reduction techniques, such as SNV, MSC, derivatives, and feature selection methods like Fisher’s criterion, were applied to enhance model robustness. DL methods, particularly 1D-CNN and stacked ML models, have demonstrated strong predictive capabilities. More recently, transformer-inspired architecture has emerged as a promising direction, addressing some limitations of traditional DL by focusing on the most relevant spectral features.

Key quality traits assessed using NIRS and HSI include firmness, TSS/SSC, DMC, TA, pH, β-carotene, and moisture content. These attributes are critical for evaluating mango ripeness and consumer acceptability. Composite indices such as FQI1 and FQI2 have also been developed to integrate multiple traits into a unified quality metric.

Despite these advancements, the research identifies notable challenges, including limited dataset sizes, the high dimensionality of spectral data, environmental variability, and lack of standardisation across studies. Importantly, attention-based models remain underutilised, and full transformer architectures like ViTs have yet to be applied to mango spectral data.

Looking ahead, future research opportunities lie in adopting full transformer-based models, developing real-time and edge-deployable AI systems, integrating multi-modal sensor data, and applying transfer learning and domain adaptation techniques. Addressing these areas can significantly enhance the scalability, accuracy, and practical implementation of intelligent mango grading systems using NIRS and HSI technologies.