Spectral-Spatial Fusion for Soybean Quality Evaluation Using Hyperspectral Imaging

Abstract

Accurate postharvest quality evaluation of soybeans is essential for preserving product value and meeting industry standards. Traditional inspection methods are often inconsistent, labor-intensive, and unsuitable for high-throughput operations. This study presents a non-destructive soybean classification approach using a simplified reflectance-mode hyperspectral imaging system equipped with a single light source, eliminating the complexity and maintenance demands of dual-light configurations used in prior studies. A spectral–spatial data fusion strategy was developed to classify harvested soybeans into four categories: normal, split, diseased, and foreign materials such as stems and pods. The dataset consisted of 1140 soybean samples distributed across these four categories, with spectral reflectance features and spatial texture attributes extracted from each sample. These features were combined to form a unified feature representation for use in classification. Among multiple machine learning classifiers evaluated, Linear Discriminant Analysis (LDA) achieved the highest performance, with approximately 99% accuracy, 99.05% precision, 99.03% recall and 99.03% F1-score. When evaluated independently, spectral features alone resulted in 98.93% accuracy, while spatial features achieved 78.81%, highlighting the benefit of the fusion strategy. Overall, this study demonstrates that a single-illumination HSI system, combined with spectral–spatial fusion and machine learning, offers a practical and potentially scalable approach for non-destructive soybean quality evaluation, with applicability in automated industrial processing environments.

1. Introduction

Soybean (Glycine max) is a globally important crop, extensively used for food, feed, and industrial purposes due to its high protein and oil content [1]. In 2024, the United States produced approximately 4.37 billion bushels of soybeans [2], highlighting the significance of efficient postharvest quality control to maintain product value and marketability. However, mechanical harvesting can introduce defects such as split, cracked, or diseased seeds and contamination by foreign materials like stems and pods. These impair product quality and demand precise postharvest inspection systems.

Traditional quality evaluation methods, such as manual sorting or chemical analysis, are destructive, labor-intensive, and often inconsistent. They are also inadequate for large-scale industrial operations that require fast, objective, and non-destructive solutions [3,4]. Recent advances in machine vision systems have facilitated automated classification using hyperspectral imaging (HSI) for various agricultural products, including soybean [4], and maize [5], establishing HSI as a powerful technology for non-destructive quality assessment. HSI captures both spectral information, the unique light reflectance profile at different wavelengths, and spatial information, such as shape, surface texture, and structural defects [6]. While spectral data is valuable for detecting internal or chemical changes, spatial features help identify external damage or contamination [7,8]. However, using either modality alone can limit classification accuracy, especially when different objects exhibit overlapping spectral or visual characteristics. To overcome these limitations, spectral-spatial fusion combines both feature types, enabling classifiers to exploit the complementary strengths of each domain. This integrated approach improves class separability, enhances robustness, and reduces errors in distinguishing visually similar categories like normal vs. split soybeans.

Complementing these imaging advances, machine learning techniques have emerged as powerful tools for interpreting high-dimensional HSI data. Algorithms such as Support Vector Machines (SVM), Random Forest (RF), Gradient Boosting, and deep learning models like Convolutional Neural Networks (CNNs) have been used to classify soybean seed defects, estimate nutrient content, and predict yield with high accuracy [9,10,11,12,13]. For example, a hybrid SVM-CNN model achieved 95.89% F1-score in classifying soybean defects using the SOYPR dataset [9], while deep learning frameworks such as SNet, SSDINet, and MobileNetV2-based systems reached accuracies exceeding 96% across multiple defect categories [11,14,15,16]. Gulzar et al. [17,18] applied modified Inception and MobileNetV2 architectures for seed classification using transfer learning and feature augmentation. However, many of these studies are limited in scope, often addressing only specific defect types and struggling to differentiate visually similar classes such as normal and split soybeans. Other approaches relying on morphological or texture-based features may suffer reduced performance in the presence of noise, overlapping objects, or high spectral redundancy [11,12]. Moreover, models trained solely on spectral or spatial features often overfit and fail to generalize across diverse environmental or postharvest conditions [4,19].

Recent studies have also explored integrating spectral and spatial data from hyperspectral images for soybean classification. For example, Tan et al. [20] combined texture features and spectral information in a Takagi-Sugeno fuzzy neural network, achieving 92% accuracy in classifying soybean varieties. Similarly, an ensemble learning model achieved 99.8% accuracy for classifying Proso Millet seed varieties using hyperspectral data [21]. However, many of these studies are limited to controlled experimental conditions or a narrow range of seed types, which can affect model stability and generalizability when applied to broader field conditions.

In previous work, Momin et al. [22] developed a proof-of-concept machine vision system using basic image processing techniques to evaluate soybean quality by detecting materials other than grain (MOGs). The system utilized both front-lit and back-lit imaging and achieved high detection rates, 98% for defective beans and stem/pods, 96% for split beans, and 75% for contaminated beans. However, the reliance on a glass plate for backlighting introduced mechanical challenges. Accumulation of dust and moisture on the glass surface interfered with light transmission, hindering accurate classification of split or broken soybeans. These maintenance and reliability issues limited the system’s scalability and practicality for field deployment or continuous operation.

To address these challenges, particularly the need for simplified lighting in postharvest image acquisition, the present study proposes a non-destructive soybean classification framework using a reflectance-mode HSI system with a single light source. This simplification was made possible by leveraging the capabilities of a hyperspectral camera, which captures rich spectral information across the visible and near-infrared regions using only one illumination source. Each soybean category exhibits distinct spectral signatures, enabling reliable classification without the need for multiple lighting angles. Moreover, spectral reflectance data are directly correlated with the chemical and physical properties of biological materials. Since both soybean and non-soybean objects in our dataset are plant-based, meaningful spectral differences were observed, supporting robust classification. In contrast to earlier approaches that required dual lighting for conventional image processing, the current study integrates spectral and spatial information to classify harvested soybeans using machine learning. This eliminates the need for complex lighting hardware and minimizes mechanical maintenance.

The objectives of this study are to (1) acquire and analyze spectral reflectance data of harvested soybeans across visible and near-infrared wavelengths to characterize different object categories, (2) extract spatial texture features using the Gray Level Co-occurrence Matrix (GLCM) to capture surface patterns associated with different soybean attributes, and (3) implement a spectral–spatial fusion strategy and evaluate its classification performance using various machine learning models, comparing it against spatial-only and spectral-only classifiers. Overall, this study proposes a simplified and potentially scalable imaging solution that reduces mechanical complexity while enabling accurate, non-destructive soybean quality evaluation in postharvest processing environments. The main contributions of this study are as follows:

• A reflectance-mode HSI system with a single light source is introduced, eliminating the need for complex dual-light configurations typically used in earlier soybean classification systems.
• A spectral–spatial data fusion framework is proposed that improves classification robustness and accuracy compared to models using only spectral or spatial features.
• A broad evaluation of eight machine learning classifiers shows that the proposed approach achieves near state-of-the-art classification performance across four soybean categories using a simplified and potentially scalable setup suitable for integration into automated industrial environments.

2. Materials and Methods

2.1. Dataset Preparation of Soybean

Soybean samples were collected during the 2024 harvesting season from an experimental farm located in Cookeville, Tennessee, USA. The soybean cultivar used in this study was Asgrow AG45XF0, a widely cultivated variety in the southeastern United States. Harvesting was conducted using a John Deere S690 combine (John Deere, Moline, IL, USA). Due to the mechanical nature of the harvesting process, the postharvest soybeans often exhibited physical damage and were susceptible to contamination by foreign materials. After collection, all samples were stored at room temperature (20–25 °C) for subsequent analysis.

For classification purposes, the samples were divided into four categories: normal soybeans, split or broken soybeans, diseased soybeans, and foreign materials. Normal soybeans were defined as intact seeds without visible damage, while split or broken soybeans showed cracks, damaged seed coats, or partial crushing. Diseased soybeans were identified based on signs of contamination with soil, fungi, mold, or vegetative matter, indicating physiological degradation. The final category included non-soybean materials, also referred to as foreign materials, consisting of stems, pods, and other plant debris commonly mixed during mechanical harvesting. Representative samples from each category are illustrated in Figure 1. A total of 1140 samples were collected and initially labeled according to the classification criteria defined by Momin et al. [22] The categorization was then verified by an experienced local soybean grower with expertise in postharvest quality traits. Following verification, the samples were sealed in numbered airtight bags for laboratory analysis. The complete dataset was divided into training and testing subsets for model development and validation, as summarized in

Table 1. Distribution of samples across training and testing datasets.

Object Type	Training Dataset	Test Dataset
Normal soybean	230	53
Split soybean	250	50
Diseased soybean	240	50
Forgien materials	214	53
Total	934	206

. Minor variations in the train-test ratio across object types (e.g., 230/53 vs. 214/53) resulted from practical constraints during sample acquisition rather than deliberate imbalance. Despite these slight variations, the dataset maintains a well-balanced distribution across categories, ensuring robust model training and fair evaluation without introducing significant bias.

2.2. Hyperspectral Image Acquisition

Hyperspectral images were acquired using a push-broom, reflectance-based visible–NIR hyperspectral imaging system (PIKA L-675, Resonon, Bozeman, MT, USA), as illustrated in Figure 2. The system comprised a C-mount objective lens (CITRINE 1.4/23 mm, Schneider Optics, Bad Kreuznach, Germany), a hyperspectral camera, and a motorized translation stage for sample movement. Similar camera systems have been widely used in recent years in non-destructive evaluation of biological and agricultural samples [23,24,25]. Spectral data were captured across a wavelength range of 379.16 to 1020.87 nm, with a spectral interval of 2.02 nm, resulting in 300 bands covering both visible and near-infrared regions. A halogen fiber line light, equipped with a polarizing film, was used for illumination to minimize surface glare. The light source was preheated for 20 min to ensure stable output before image acquisition.

The entire system was enclosed to eliminate ambient light interference, and image acquisition was controlled using Spectronon software (version 3.5.5). An exposure time of 121.129 ms was used to optimize illumination without overexposure. Each sample was scanned at a speed of 3.45 mm/s, producing a hyperspectral data cube with dimensions of 700 lines × 900 spatial pixels per line × 300 spectral bands. The spectral axis recorded wavelength-dependent reflectance values, while the two spatial axes captured the physical structure and arrangement of the soybean samples.

2.3. Proposed Method

The proposed classification framework integrates both spectral and spatial features extracted from hyperspectral imagery, as shown in Figure 3. The process begins with data acquisition and continues with preprocessing steps, including single-band extraction, Gaussian smoothing, intensity thresholding, and morphological erosion. These steps were applied to remove background noise and segment individual soybean objects from the scene. Once the objects were isolated, features were extracted from both spectral and spatial domains. Spectral features consisted of the mean reflectance values across 300 wavelengths within the hyperspectral cube, which were preprocessed (as described in Section 2.4.2) to reduce spectral noise and enhance signal consistency. In parallel, spatial features were computed using the Gray Level Co-occurrence Matrix (GLCM) method, which captures texture characteristics from grayscale images reconstructed from the hyperspectral data. After feature extraction, a feature-level fusion step was performed to combine spectral and spatial features for classification. The resulting fusion vector F_fusion was constructed as follows:

F_fusion = [F_spectral ∥ F_spatial]

(1)

where F_spectral is the 300-dimensional preprocessed spectral feature vector, F_spatial is the 14-dimensional spatial texture feature vector, and ∥ denotes vector concatenation. This fusion strategy was not a simple raw concatenation but involved a deliberate integration of noise-reduced spectral data with structured spatial texture features to ensure complementary feature interactions and improve classification robustness. This enriched dataset was used to train classification models. To evaluate model performance, a k-fold cross-validation approach was employed, and metrics such as accuracy, precision, and F1-score were calculated to assess classification reliability and generalizability.

2.4. Data Processing

2.4.1. Spatial Data Processing

Spatial segmentation began with the extraction of a single NIR band at 1000.44 nm, which provided optimal contrast between soybean objects and the background, as shown in Figure 4a,b. This wavelength was selected based on its high reflectance in the dataset, which provided strong object–background contrast and enabled accurate and efficient segmentation. The processing pipeline included Gaussian blurring to reduce high-frequency noise while preserving edge detail, as shown in Figure 4c. Binary thresholding was then applied to generate a mask distinguishing foreground objects from the background, as visualized in Figure 4d. To improve object coherence, morphological erosion was subsequently used to refine object boundaries, followed by an area-based filtering step to remove noise-like elements smaller than 500 pixels, as illustrated in Figure 4e as the final object mask. To support spatial texture analysis, RGB images were reconstructed from selected spectral bands corresponding to red (640.2 nm), green (549.6 nm), and blue (460.6 nm) wavelengths. These images were then converted to grayscale for feature extraction using texture-based methods. The complete preprocessing sequence facilitated reliable object segmentation, enabling robust extraction of both spatial and spectral features.

2.4.2. Spectral Data Processing

Hyperspectral data, comprising up to 300 wavelengths, often contain noise introduced by environmental factors and instrumentation errors [26,27]. This noise can obscure true spectral characteristics and reduce the effectiveness of machine learning models. To mitigate these effects and enhance data quality, spectral preprocessing was performed prior to classification. Three commonly used techniques were evaluated in this study: Multiplicative Scatter Correction (MSC) [28], Standard Normal Variate (SNV) [29], and Savitzky–Golay (SG) smoothing [30]. While MSC and SNV are widely applied to correct scattering effects caused by variations in sample geometry, surface texture, and illumination conditions [31,32], SG smoothing demonstrated the most consistent improvements in classification performance and was therefore selected for further use. All spectral reflectance values were normalized to a range of 0 to 1 to standardize comparisons across samples. SG smoothing was then applied using a window length of 7 and a third-order polynomial, with a derivative order of zero to ensure smoothing without differentiation. This approach effectively reduced high-frequency noise while preserving key spectral features such as peak intensity and curve shape.

2.5. Feature Extraction

2.5.1. Spatial Feature Extraction

GLCM is a statistical method commonly used in image processing to quantify textural patterns by analyzing the spatial relationships between pixel intensities within an image. In this study, RGB images reconstructed from selected hyperspectral bands were converted to grayscale, and GLCM features were extracted to describe the texture characteristics of segmented soybean objects. A total of 14 spatial features were computed based on the method proposed by Haralick et al. [33], including angular second moment, contrast, correlation, entropy, inverse difference moment, difference entropy, difference variance, sum average, sum entropy, sum variance, sum of squares: variance, maximal correlation coefficient, and two measures of information correlation. These features capture various aspects of texture, such as uniformity, intensity variation, randomness, and spatial dependence.

Table 2. GLCM texture features extracted from representative samples across four soybean classification categories.

GLCM Features	Normal Soybean	Split Soybean	Diseased Soybean	Foreign Materials
Angular second moment	0.001166411	0.001468802	0.00060536	0.000800113
Contrast	9.9873399	6.42387764	41.7277266	54.2942626
Correlation	0.995690085	0.996226727	0.984167122	0.977210946
Difference entropy	2.690342632	2.467046877	3.592117992	3.704523883
Difference variance	0.00083119	0.001122403	0.000494607	0.000451044
Entropy	10.36128538	9.940645824	11.28466308	11.20165864
Information measure of correlation 1	−0.541876642	−0.55026881	−0.40437518	−0.37690467
Information measure of correlation 2	0.99975887	0.999714277	0.998191315	0.996879067
Inverse difference moment	0.400746661	0.41733519	0.246150516	0.262414697
Maximal correlation coefficient	3.253406957	3.026776646	3.946432939	4.852678038
Sum average	235.7983282	229.8465888	240.6570153	119.0831582
Sum entropy	8.094178612	7.847192389	8.051474218	7.892176898
Sum of squares: variance	1159.904379	852.2120745	1319.660286	1191.164885
Sum variance	4629.630174	3402.42442	5236.913416	4710.365278

presents the GLCM feature values for representative samples from each soybean category. The results show that normal and split soybeans exhibited similar values for features such as angular second moment, contrast, and inverse difference moment, reflecting their visual similarity under standard illumination. In contrast, diseased soybeans and foreign materials showed higher entropy and difference entropy values, indicating more irregular and unpredictable texture patterns. Correlation, which measures the linear relationship between neighboring pixels, was slightly higher for normal and split soybeans compared to diseased and foreign materials. The information measures of correlation also demonstrated strong texture dependency across all categories. Notably, split soybeans showed higher difference variance due to the inclusion of samples in multiple orientations, which introduced variability in local intensity differences. Features such as sum entropy, sum of squares: variance, and sum variance were lower for split soybeans, likely due to reduced surface randomness resulting from compromised seed coats. Foreign materials, which typically have less reflective surfaces, exhibited lower values in brightness-related features such as sum average and maximal correlation coefficient.

2.5.2. Spectral Feature Extraction

To enable effective classification based on spectral features, it is essential to identify distinct spectral signatures and characteristic reflectance patterns across different soybean categories. Figure 5 illustrates representative mean spectral reflectance curves for normal soybeans, diseased soybeans, split soybeans, and foreign materials, across the 379.16 to 1020.87 nm wavelength range. Since these categories exhibit different reflectance behaviors at various wavelengths, all 300 spectral bands were retained to preserve the complete spectral profile. Reducing the number of bands could risk omitting subtle but class-relevant variations, particularly for visually or chemically similar categories.

The spectra revealed that normal and split soybeans shared similar overall curve shapes, characterized by reflectance features in the visible region and increased reflectance in the NIR region. Despite this similarity, noticeable differences were observed in their reflectance intensities. Split soybeans generally exhibited reduced reflectance beyond 700 nm, likely due to compromised seed coats exposing internal structures and altering light scattering and absorption behavior. In contrast, normal soybeans, with intact surfaces, reflected more light in the NIR range, yielding higher intensity values, typically between 0 and 20,000. Split soybeans had different orientation such as, the broken soybean is directly exposed under light or the soybean coat which was compromised is directly under light. However, regardless of orientation, split soybean demonstrated lower intensities, generally ranging from 0 to 15,000. These intensity differences, while preserving the general spectral signature, offer a valuable discriminative feature for classification.

Diseased soybeans displayed significantly altered spectral characteristics. Their mean reflectance curves showed elevated intensities across most wavelengths, particularly in the visible and red-edge regions. These increases suggest structural degradation, chlorophyll loss, and physiological stress due to fungal or microbial infection [34]. The resulting spectral distortion reflects the disruption of normal absorption and scattering mechanisms associated with pathological tissue conditions.

Likewise, foreign materials exhibited spectral signatures that were clearly distinguishable from all soybean categories. Their unique absorption and reflectance profiles stem from differences in surface morphology and internal composition compared to soybean. This distinct spectral behavior provides a reliable foundation for classification models to accurately differentiate between soybean types and effectively exclude foreign materials like contaminants. Collectively, these spectral distinctions reinforce the value of HSI as a non-destructive and efficient tool for postharvest soybean quality assessment.

2.6. Machine Learning Algorithms

To evaluate the classification performance of the proposed spectral–spatial fusion approach, eight widely used machine learning classifiers were employed. These models were selected based on their strong performance in previous soybean-related studies and their ability to capture diverse data characteristics. The selection includes both linear models and ensemble-based methods, chosen to ensure a comprehensive assessment across varying model complexities and assumptions.

The classifiers used in this study include Linear Discriminant Analysis (LDA) [19], Quadratic Discriminant Analysis (QDA) [35], Logistic Regression (LR) [36], K-Nearest Neighbors (KNN) [37], Random Forest (RF), Extra Trees (ET), Gradient Boosting (GB), and Light Gradient Boosting Machine (LightGBM). Among these, ensemble methods such as RF, ET, GB, and LightGBM were included due to their ability to model non-linear relationships and reduce overfitting. For example, RF was utilized by Chen et al. [4] for impurity detection in machine harvested soybean; ET was applied by Dhal et al. [38] to predict water uptake by hydroponically grown soybeans; GB was used by Khalili et al. [39] to classify soybean charcoal rot disease; and LightGBM employed by Li et al. [40] to estimate soybean yield.

LDA was particularly effective in our experiments due to its strength in maximizing class separability in high-dimensional, linearly separable datasets, characteristic of hyperspectral data. KNN and LR were included as baseline models for comparative evaluation. Some classifiers performed better with high-dimensional spectral data (e.g., LDA), while others handled lower-dimensional spatial features more effectively. This diverse selection allowed us to thoroughly assess the impact of spectral, spatial, and fused features across a broad range of learning paradigms.

All models were implemented in Python 3.10 using the PyCaret library. Training and evaluation were conducted on a workstation equipped with an Intel Core i7-14700 processor and 64 GB of RAM. No GPU acceleration was used. Traditional machine learning models used in this study do not follow iterative, epoch-based training processes typical of deep learning frameworks and therefore do not produce training or validation loss curves. To ensure robust evaluation and mitigate overfitting, we employed 10-fold stratified cross-validation and reported standard performance metrics.

2.7. Evaluation Metrics

To comprehensively evaluate the performance of the classification models, several standard metrics were employed, including accuracy, precision, recall, F1-score, the confusion matrix, and the receiver operating characteristic (ROC) curve. These metrics were selected to assess both the overall reliability of the models and their ability to differentiate between high-quality soybeans and defective or foreign materials.

Accuracy reflects the overall proportion of correctly classified instances among all predictions and serves as a broad indicator of model effectiveness. In the context of postharvest quality assessment, it is particularly important for ensuring that normal soybeans are not misclassified or discarded. However, relying solely on accuracy can be misleading in imbalanced datasets. Therefore, class-specific metrics such as precision and recall are also critical. Precision quantifies the proportion of true positive predictions among all samples predicted as positive, offering insight into the model’s ability to avoid false positives, such as misclassifying defective soybeans as normal. Recall, or sensitivity, measures the proportion of actual positive cases correctly identified by the model, which is essential for detecting all instances of defective or diseased soybeans that should be removed from the processing stream.

The F1-score, defined as the harmonic mean of precision and recall, provides a balanced measure that accounts for both types of misclassifications. It is particularly useful when there is a trade-off between false positives and false negatives, as in the case of distinguishing between similar-looking categories such as normal and split soybeans. Additionally, the confusion matrix was used to provide a detailed breakdown of classification results, highlighting true positives and false negatives across all categories. High values along the diagonal indicate strong model performance, while off-diagonal elements reveal areas where misclassifications are occurring.

Finally, the ROC curve was used to evaluate the model’s diagnostic ability by plotting the true positive rate (TPR) against the false positive rate (FPR) across various threshold settings. The area under the curve (AUC) serves as a scalar summary of the model’s class separability, with values approaching 1.0 indicating excellent classification performance. Together, these metrics offer a robust framework for assessing model performance and ensuring its suitability for real-world, non-destructive soybean quality evaluation.

3. Results and Discussion

3.1. Model Performance

This section presents the classification performance of the top five machine learning models introduced in the Section 2.6, evaluated across three types of input features: spatial features derived from GLCM, spectral reflectance features obtained from hyperspectral imaging, and a combination of both through data fusion. Performance metrics including accuracy, precision, recall, and F1-score were reported using 10-fold stratified cross-validation. This approach helped ensure balanced evaluation, address class distribution differences within the dataset and identify the most effective feature representation and modeling approach for soybean quality classification.

3.1.1. Using Spatial Features

The performance of classifiers trained using only spatial texture features derived from GLCM is presented in

Table 3. Performance evaluation of classifiers using spatial features.

Classifier	Accuracy	Precision	Recall	F1-Score
GB	0.8811	0.8847	0.8811	0.8802
LightGBM	0.8747	0.8786	0.8747	0.8739
ET	0.8715	0.8738	0.8715	0.8708
RF	0.8544	0.8593	0.8544	0.8533
LDA	0.8339	0.8414	0.8339	0.8309

Bold values indicate the highest value for each evaluation metric.

. Among all models, the GB classifier achieved the highest overall performance, with an accuracy of 0.8811, precision of 0.8847, recall of 0.8811, and an F1-score of 0.8802. GB outperformed other models such as LightGBM, ET, RF, LDA, and QDA. The superior performance of GB can be attributed to its ability to capture subtle non-linear interactions within the compact set of 14 texture-based spatial features. It effectively boosted weak learners to differentiate among features such as angular second moment, entropy, contrast, and inverse difference moment, which were critical for class separation.

3.1.2. Using Spectral Features

For hyperspectral input, we used the mean reflectance value across 300 wavelengths (379.16 nm–1020.87 nm) for each object. This approach significantly reduced data volume while preserving key spectral characteristics. Various preprocessing techniques were applied to the spectral data, including MSC, SNV, and SG smoothing. As shown in Figure 6, the LDA classifier resulted the highest accuracy (0.9893) when SG smoothing was applied. SG smoothing was most effective at reducing high-frequency noise while preserving critical features necessary for class separation. In contrast, MSC and SNV, although useful for correcting scatter and baseline shifts, sometimes distorted subtle spectral differences between categories.

Based on this result, SG-preprocessed spectral data were used for further analysis. Classifier performance using SG-smoothed features is summarized in

Table 4. Performance evaluation of classifiers using spectral reflectance features.

Classifier	Accuracy	Precision	Recall	F1-Score
LDA	0.9893	0.9900	0.9893	0.9893
LR	0.9314	0.9378	0.9314	0.9311
QDA	0.9314	0.9415	0.9314	0.9285
ET	0.9229	0.9263	0.9229	0.9227
KNN	0.9165	0.9221	0.9165	0.9162

Bold values indicate the highest value for each evaluation metric.

. LDA again outperformed all other models, including LR, QDA, ET, KNN, and LightGBM, likely due to its suitability for high-dimensional data and its ability to project features in a maximally separable space.

3.1.3. Data Fusion

To assess the benefit of combining spatial and spectral features, we fused 14 GLCM-based spatial features with 300 SG-preprocessed spectral features. As shown in Figure 7, LDA achieved the best performance, with all four metrics (accuracy, precision, recall, F1-score) reaching approximately 0.99. Other classifiers, including RF, ET, GB, and LightGBM, yielded strong results in the 0.94–0.96 range but fell short of LDA’s consistency and precision. The effectiveness of LDA under data fusion is attributed to its capacity to leverage complementary information from both spectral and spatial domains, especially in high-dimensional but linearly separable spaces.

Although the improvement in classification accuracy over the spectral-only model is less than 1% and not statistically significant, the fusion-based model offers practical advantages in prediction robustness and generalizability. While the spectral-only model benefits from its use of 300 wavelength-specific features, it lacks spatial and textural context, which may limit its capacity to capture structural object characteristics. In contrast, the spatial-only model demonstrated notably lower performance. The fusion-based approach effectively combines both feature types, resulting in more reliable and interpretable predictions.

3.1.4. Model Performance Comparison

Figure 8 presents confusion matrices comparing the top-performing classifiers across feature types. The GB classifier, trained with only spatial features (Figure 8a), shows noticeable misclassifications, particularly between normal and split soybeans, highlighting the limitations of texture-only data. In contrast, LDA classifiers using only spectral features (Figure 8b) and the combined spectral–spatial feature set (Figure 8c) demonstrated excellent classification accuracy across all categories.

The fusion-based model misclassified two samples, reinforcing that combining spectral reflectance and texture features improves class separability and reduces error rates. In one case, the model likely misinterpreted the hilum region of a normal soybean as a darkened spot, resulting in its misclassification as a diseased soybean. In the other case, the irregular shape and size of a split soybean closely resembled foreign materials, contributing to the incorrect prediction. These instances highlight the inherent challenge of classifying visually ambiguous samples where feature overlap may occur despite spectral–spatial fusion.

Additionally, the ROC curves presented in Figure 9 further validate the advantage of integrating spectral and spatial features for soybean classification. The LDA models (Figure 9b,c) achieved outstanding area under the curve (AUC) scores of 0.99 to 1.00 across all categories, indicating near-perfect sensitivity and specificity. In contrast, the GB model (Figure 9a), which relied solely on spatial features, showed slightly lower AUC values, particularly for normal soybeans (0.94) and split soybeans (0.93), suggesting greater uncertainty in distinguishing these classes. The enhanced performance of the LDA classifiers highlights the effectiveness of combining hyperspectral reflectance data with spatial texture patterns. This superior result is likely attributed to LDA’s strength in maximizing class separability within high-dimensional, linearly separable feature spaces created through data fusion.

To contextualize the performance of our approach, we compared it with previous deep learning-based studies, as summarized in

Table 5. Performance comparison between the proposed method and deep learning-based classification methods reported in previous studies.

Work	Model Architecture	Classes of Objects	Accuracy
Proposed Method	Data-fusion (Spectral + Spatial)	4	99.03%
Kaler et al. [41]	CNN	2	97.72%
Zheng et al. [42]	ShuffleNet	4	98.36%
N. Zhang et al. [43]	DenseNet	4	98.48%
Gulzar et al. [18]	InceptionV3	4	98.73%

. Kaler et al. [41] used a CNN to classify two soybean seed categories, achieving 97.72% accuracy. Zheng et al. [42] and Zhang et al. [43] applied ShuffleNet and DenseNet, respectively, to classify four categories with accuracies of 98.36% and 98.48%. Gulzar et al. [18] used an InceptionV3 architecture for five classes, achieving 98.73% accuracy. In contrast, our proposed method employed a data-fusion model that integrates spectral data with spatial attributes to classify four categories of harvested soybean objects, achieving an accuracy of 99.03%. This result is comparable to or better than prior studies, demonstrating the effectiveness of the proposed approach in postharvest soybean quality evaluation.

3.2. Visualization of Model Prediction

To better illustrate the practical implications of the classification models, this section presents visual examples of prediction outcomes on previously unseen hold-out samples. Three scenarios are compared: models trained using only spatial features, only spectral features, and a combination of both through data fusion. These visualizations demonstrate the effectiveness and limitations of each feature set in accurately identifying soybean categories.

3.2.1. Inference Using Spatial Features

Figure 10 shows the prediction results from the GB model trained solely on spatial texture features. The model effectively captured the general structure and spatial positioning of the objects, with the output closely matching the ground truth, except for two misclassified instances. This result highlights the ability of spatial features to represent geometric and textural variations. However, it also reveals their limitations in capturing subtle spectral differences, particularly when objects have similar surface textures but differ in internal composition.

3.2.2. Inference Using Spectral Features

Figure 11 depicts the model output from the LDA classifier trained using only spectral features. The model successfully identified most objects with high fidelity to the ground truth. Spectral features, which represent reflectance patterns across wavelengths, proved effective in differentiating soybean categories. However, one misclassification was observed, suggesting that spectral similarity between certain classes, such as normal and split soybeans, can lead to ambiguity when reflectance characteristics overlap. This inference, performed on a completely independent hold-out set, affirms the robustness and generalizability of the spectral feature-based model.

3.2.3. Inference Using Data Fusion

Figure 12 illustrates the prediction from the LDA classifier trained on the fused dataset comprising both spectral and spatial features. The model produced a highly accurate classification, closely aligned with the ground truth. The fused model effectively captured both spectral reflectance patterns and spatial structural features, leading to more reliable and consistent classification across all object categories. These results confirm the value of data fusion in enhancing classification accuracy and generalization, even when evaluated on unseen samples.

3.3. Limitations

This study has several limitations that warrant consideration. First, classification was performed on isolated soybean objects obtained through controlled image segmentation. While this approach ensured consistency and precise feature extraction, it did not incorporate scenarios involving occlusion, such as overlapping or partially visible beans, conditions common in real-world processing environments. The choice to focus on isolated objects was intentional, allowing for a foundational analysis of spectral–spatial feature fusion without the confounding effects of visual obstruction. Addressing occlusion is a meaningful next step; however, it introduces challenges in segmentation accuracy, object boundary preservation, and feature distortion. Second, lighting variability was minimized by using a halogen fiber line light combined with a polarizing film (Section 2.2), effectively reducing glare and halation. However, performance under variable lighting conditions, typical of industrial settings, remains to be evaluated. Third, we did not perform explicit runtime benchmarking, though hardware specifications were provided (Section 2.6). The system was intentionally designed to reduce computational overhead by avoiding time-intensive operations like pixel-wise segmentation. Lastly, the study focused on a single soybean cultivar (Asgrow AG45XF0), limiting the genetic and phenotypic diversity of the dataset.

4. Conclusions

This study presents a non-destructive approach for postharvest soybean classification using hyperspectral imaging and spectral–spatial data fusion. Analysis across a broad spectral range (379.16–1020.87 nm) revealed distinct reflectance patterns for normal, split, diseased, and foreign materials. The proposed methodology employed a push-broom, reflectance-based hyperspectral imaging system with single-source illumination, which streamlined data acquisition while avoiding the mechanical complexity and maintenance issues of dual-lighting setups. This simplified configuration also minimized the need for extensive image preprocessing, enhancing the system’s practicality for real-world postharvest applications. By integrating 300 spectral features with 14 spatial texture attributes extracted via GLCM, classification performance was significantly enhanced. Among tested classifiers, LDA achieved the highest accuracy, reaching approximately 99%. These findings demonstrate the potential of combining spectral and spatial information to support reliable, scalable, and practical soybean quality assessment in postharvest environments. Future work will focus on extending the system to include realistic field conditions, such as overlapping samples, and exploring advanced segmentation techniques or deep learning-based occlusion-handling methods to enhance model generalizability and robustness. Multiple soybean cultivars will also be incorporated to increase sample diversity and improve applicability across a wider range of postharvest scenarios.