1. Introduction
Litchi (Litchi chinensis) is a tropical and subtropical fruit of significant economic importance, renowned for its distinct flavor and rich nutritional profile. However, its quality is frequently compromised by various factors, particularly pest infestations. Among these, the litchi stem borer (Tetranychus cinnabarinus) is one of the most prevalent and damaging pests, causing both external and internal damage to the fruit and thereby reducing its market value. Traditional pest detection methods, such as visual inspection and destructive testing, are not only time-consuming but also prone to human error and may damage the fruit, making them ill suited for the modern agricultural need for efficient, accurate, and non-destructive detection techniques. Therefore, the development of effective, non-destructive, and precise pest detection methods is essential in ensuring the quality of litchis and enhancing competitiveness in the industry.
Recent advancements in spectroscopy and imaging technologies have led to the proposal of various non-destructive detection methods, which have yielded promising results in agricultural product quality assessment [1]. In particular, visible/near-infrared (Vis/NIR) spectroscopy [2], X-ray imaging [3], and hyperspectral imaging technologies [4] have shown considerable potential in detecting the physical, chemical, and structural properties of fruits. Vis/NIR spectroscopy is particularly effective in capturing the chemical composition of fruits and is widely employed to assess attributes such as ripeness, sugar content, and acidity. Hyperspectral imaging offers a higher spatial resolution, allowing for the detection of subtle internal changes in fruits, while X-ray imaging enables the visualization of the internal structure and potential defects in fruits. Despite the advantages of these individual technologies, single-sensor approaches are often insufficient in detecting pests in complex fruits, particularly those with intricate skin structures.
The existing research on fruit quality detection primarily focuses on the use of single-sensor technologies. For instance, a study on a visible/near-infrared (Vis/NIR) spectroscopy system for detecting cold damage in bananas demonstrated the flexibility and efficiency of this technology in agricultural monitoring [5]. This study shares similarities with our work, particularly in utilizing Vis/NIR spectroscopy for fruit quality assessment. However, our research extends this approach by exploring its application in detecting fruit damage and diseases, emphasizing both its advantages and limitations.
Furthermore, the study “Source tracing and identification of garlic based on multi-source heterogeneous spectral information fusion” highlights the potential of multi-source spectral information fusion in tracing the origin of garlic [6]. This approach aligns closely with our strategy of data fusion for detecting litchi stem borer damage. By integrating spectral data from different sources, the garlic origin tracing system achieved high efficiency and accuracy. Similarly, we adopted this fusion strategy to improve the precision and reliability of our litchi pest detection system. In the following sections, we will briefly discuss how this multi-source data fusion method enhances the overall system performance and provides the theoretical foundation for our methodology.
Furthermore, a comparison of the physical and chemical properties of pest-free and pest-infested litchis reveals notable differences. Pest-free litchis exhibit higher hardness, soluble sugar content [7], vitamin C [8], polyphenols [9], anthocyanins [10], and ORAC values [11] than pest-infested litchis. These biological differences provide a solid theoretical foundation for the non-destructive detection of internal pest damage in litchis.
This study presents an effective non-destructive detection approach for litchi, a fruit with a complex skin structure, addressing a gap in existing methods that predominantly focus on fruits with simpler or more uniform skins. Traditional detection techniques are often less applicable to such complex fruits. By leveraging multi-source information fusion, this research demonstrates that integrating spectral and imaging data from Vis/NIR spectroscopy with hyperspectral imaging and X-ray imaging can allow the effective assessment of internal quality and distinction between healthy and damaged litchis.
The primary aim of this study is to develop an efficient and accurate non-destructive detection method for litchi stem borer damage through multi-sensor data fusion [12]. By combining spectral and imaging data from Vis/NIR spectroscopy, hyperspectral imaging, and X-ray imaging, this study presents a comprehensive approach to assessing the internal quality of litchi and effectively distinguishing between healthy and damaged fruits. To optimize data utilization, we employ a feature-level fusion strategy, where data from different sensors are preprocessed, standardized, and combined into a unified feature set. This approach fully capitalizes on the complementary strengths of each sensor, thereby enhancing the detection accuracy and reliability.
In future studies, we will further investigate the impact of various data processing strategies on model performance, including preprocessing methods, feature extraction techniques, and model optimization configurations. By combining multi-source information fusion with advanced data processing strategies, this research provides a robust and reliable non-destructive detection solution for litchi pest monitoring and offers valuable insights for use in pest detection in other fruits.
2. Materials and Methods
2.1. Litchi Samples
The litchi variety used in this study was ‘Feizixiao’. A total of 420 litchis were harvested from Zengcheng District, Guangzhou City, Guangdong Province, including 93 pest-infested fruits and 327 healthy ones. On the same day, the samples were transported to the laboratory for data collection, which included visible/near-infrared (Vis/NIR) spectra, hyperspectral images, and X-ray images. Following data acquisition, the litchis were carefully cut open to observe and document their internal conditions.
2.2. Data Acquisition
2.2.1. Visible/Near-Infrared Spectroscopy
The Vis/NIR spectroscopy non-destructive testing platform used in this study was custom-built and primarily consisted of an Ocean Optics QE Pro Vis/NIR spectrometer (Ocean Insight, Orlando, FL, USA) with a spectral range of 400–1100 nm, a litchi holder, a 10 W halogen lamp, and a computer. To ensure the stability and consistency of the spectral data, an opaque holder was designed to securely position the litchi and minimize interference from external light sources.
After several adjustments and tests, the optimal parameters for collecting the Vis/NIR spectra of litchis were determined as follows: an integration time of 2000 ms, a 5 mm distance between the receiving fiber and the litchi holder, and a 5 mm distance between the litchi and the light source. During the experiment, the litchis were positioned with their stems facing upward and centrally aligned on the holder, ensuring that the light source, fiber optic entrance, litchi, light exit, and receiving fiber were all vertically aligned. This configuration facilitated effective spectral signal collection and ensured the reproducibility of the results (Figure 1).
2.2.2. Hyperspectral Imaging
The hyperspectral imaging platform used in this study was developed in-house and consists of an FX17E hyperspectral imager (Specim, Oulu, Finland) with a spectral range of 400–1000 nm and a spectral resolution of 8 nm. The system is equipped with a matte black motorized conveyor belt, an illumination system comprising eight 50 W halogen lamps, a black diffuse reflection wall, and a computer for data processing. Hyperspectral images were captured using Spectronon Pro (2.102) software, with the imager preheated for 15 min prior to data collection to ensure stable image quality.
To prevent image distortion and blurriness, optimal data acquisition parameters were determined through preliminary testing: a 45 mm distance between the spectrometer and the sample, an exposure time of 64.742 ms, a frame rate of 15 Hz, and a scanning speed of 800 pps. The resolution for each litchi image was set to 200 × 200 pixels. A region of interest (ROI) of 120 × 120 pixels, centered around the litchi stem, was selected for stem borer detection, based on previous research findings. Black-and-white reference calibration was performed before each data acquisition session to ensure accuracy in the spectral data (Figure 2).
2.2.3. X-Ray Imaging
The X-ray imaging platform developed in this study comprises a RAYON4343 X-ray generator and detector (Rayon Testing Equipment Co., Ltd., Dongguan, China) with a source voltage of 100 kV, a radiation power of 15 W, and a focal spot size of 5 µm. The system is equipped with a computer-controlled conveyor belt coated with X-ray-reflective material and a computer for data processing.
To optimize image clarity and detection performance, the ideal data acquisition parameters for X-ray imaging were established through multiple preliminary experiments. The X-ray tube voltage was set to 50 kV, the current to 192 µA, and the conveyor belt speed to 2.72 m/s. The litchi was placed under the X-ray tube for 2.8 s to ensure complete image acquisition, resulting in an image resolution of 3000 × 3000 pixels.
During the experiments, the litchis were placed on low-density foam trays to minimize movement and external interference. The X-ray generator and image collector were aligned vertically with the litchi stem and core, respectively, to ensure the accurate capture of key features during imaging (Figure 3).
2.3. Data Analysis and Modeling
2.3.1. Spectral Preprocessing Methods
To mitigate the effects of the sample background, high-frequency noise, baseline drift, and light scattering on the stability of the spectral data, systematic preprocessing was performed on the collected spectra. Specifically, multiplicative scatter correction (MSC) [13] and standard normal variate (SNV) [14] transformations were applied to eliminate interference caused by scattering effects and baseline drift. Furthermore, to further reduce noise and enhance the signal-to-noise ratio, the spectral data were smoothed twice using a five-point Savitzky–Golay (SG) [15] filter and a moving window smoothing method. These preprocessing steps effectively improve the quality of the spectral data, ensuring stability and accuracy in the subsequent modeling process.
2.3.2. Feature Selection Methods
To manage the large volume of spectral data and mitigate the severe collinearity among variables, while enhancing both computational efficiency and model accuracy, this study applies two feature selection algorithms to extract and compare feature variables, with the goal of identifying the optimal selection method.
First, the Successive Projections Algorithm (SPA) [16] is employed as a forward selection method designed to reduce variable collinearity by identifying redundant variables within the spectral data. This method iteratively selects variables, forming a subset that minimizes redundancy. In this study, the number of selected variables was predetermined to range from 5 to 30, with a final selection of 20 feature wavelengths that yielded the lowest root mean square error (RMSE) for modeling.
Second, the Competitive Adaptive Reweighted Sampling (CARS) [17] algorithm evaluates the importance of variables based on their regression coefficients (RCs). This method combines adaptive reweighted sampling with an exponential decay function, prioritizing features with higher absolute regression coefficients in the partial least squares regression (PLSR) model. The optimal subset of variables was selected through cross-validation, using the Root Mean Square Error of Cross-Validation (RMSECV) as the selection criterion. In this study, 5000 Monte Carlo samplings were performed with a training-to-test set ratio of 2:1, leading to the selection of 20 feature wavelengths as the optimal variables for the CARS algorithm.
Third, Uninformative Variable Elimination (UVE) [18] is a method designed to remove variables that contribute minimally to model accuracy, based on their statistical significance in the regression model. UVE assesses each variable’s relevance by applying stability criteria, comparing its importance to that of randomly generated noise variables. Variables with importance values below a predefined threshold are considered uninformative and are subsequently eliminated. In this study, UVE was employed to identify and remove irrelevant spectral variables, thereby enhancing model performance by reducing data complexity and mitigating the risk of overfitting.
Through a comparison of the three methods, SPA, CARS, and UVE, this study investigates the most effective feature selection approach to optimize model performance and reduce computational complexity.
2.3.3. Model Construction
The spectral data, after feature extraction, were divided into training and test sets, and pattern recognition was performed using partial least squares regression (PLSR) and support vector regression (SVR) [19]. After several trials, the optimal number of latent variables for PLSR was determined to be 7, and the radial basis function (RBF) was selected as the kernel type for SVR. During model training, hyperparameter optimization was conducted using cross-validation combined with grid search on the training set, with the minimum root mean square error (RMSE) as the criterion to identify the optimal values for the penalty parameter C and kernel parameter γ. The expected output values for pest-free and infested litchis were assigned as 0 and 1, respectively.
Model performance was evaluated using the coefficient of determination (R2), RMSE, and accuracy. Since both PLSR and SVR generate continuous outputs, the classification accuracy was calculated by rounding the predicted values: outputs less than or equal to 0 were classified as pest-free, and those greater than or equal to 1 were classified as infested. This method effectively transforms the regression outputs into binary classification results, providing a robust measure of detection performance for non-destructive testing.
2.3.4. Multi-Sensor Data Fusion Method
To enhance the accuracy of detecting litchi stem borers, this study proposes a feature-level multi-sensor data fusion method that effectively integrates hyperspectral imaging, visible/near-infrared spectroscopy, and X-ray imaging techniques [20]. Initially, data from each modality are independently collected and preprocessed. The visible/near-infrared spectroscopy data undergo Savitzky–Golay (SG) smoothing and standard normal variate (SNV) correction, with feature selection performed using the Competitive Adaptive Reweighted Sampling (CARS) algorithm to extract key features related to sugar and vitamin C content. The hyperspectral imaging data are preprocessed using SG smoothing and multiplicative scatter correction (MSC), with spectral features linked to litchi hardness and sugar content extracted. For X-ray imaging, key structural features of the fruit stem and seed are extracted through noise reduction, sharpening, and brightness adjustment.
Through multi-sensor fusion, chemical features from hyperspectral and visible/near-infrared spectroscopy are combined with structural features from X-ray imaging to create a unified feature vector. Recursive Feature Elimination with Cross-Validation (RFECV) is then employed to select the most significant features, resulting in a consolidated feature set. RFECV iteratively eliminates features with minimal contribution to model prediction, retaining only the most influential ones. Cross-validation is used at each iteration to evaluate model performance, ensuring that the final feature set optimizes predictive accuracy while minimizing overfitting. RFECV is particularly suitable for multi-sensor data fusion, where features may be redundant or highly correlated. It effectively removes irrelevant or duplicated features, focusing on those with the highest predictive power, thus improving the accuracy and stability of the model in detecting litchi stem borers [21,22].
The model performance was evaluated using k-fold cross-validation with k values ranging from 3 to 10. In this method, the i-th subset served as the test set, while the remaining subsets were used for training. This approach ensured a robust evaluation of the model’s ability to generalize across different data partitions, reducing any bias from a specific split. Cross-validation enhanced the statistical rigor and reliability of the evaluation. Additionally, hyperparameter optimization [23] was performed to select the optimal configuration, further improving the model’s detection accuracy and ensuring better generalization across diverse datasets.
To reduce feature dimensionality and improve model efficiency, principal component analysis (PCA) was applied. PCA reduces redundant information in the feature space, enhancing computational efficiency and mitigating the risk of overfitting. In multi-sensor data fusion, each sensor may generate high-dimensional, correlated features. PCA compresses high-dimensional data into a lower-dimensional space while retaining key variance, thus accelerating computation, reducing noise interference, preventing overfitting, and improving the accuracy and stability of the litchi stem borer detection model [24]. By combining PCA with Recursive Feature Elimination with Cross-Validation (RFECV), we further optimized the feature set, ensuring that the selected features had high predictive power and enhancing model stability.
Ultimately, by integrating Recursive Feature Elimination with Cross-Validation (RFECV) for feature selection, principal component analysis (PCA) for dimensionality reduction, k-fold cross-validation for performance evaluation, and hyperparameter optimization, the model’s robustness and statistical integrity were ensured. The optimization of feature selection and hyperparameters, coupled with rigorous validation, significantly enhanced the model’s generalization ability and detection accuracy for litchi stem borers [25,26]. This comprehensive approach provides strong support for the development of a reliable and efficient detection model, thereby facilitating effective fruit quality control and pest management (Figure 4).
3. Results
3.1. Visible/Near-Infrared Spectroscopy Detection
3.1.1. Visible/Near-Infrared Spectral Analysis of Internal Litchi Infestation
In this study, visible/near-infrared (Vis/NIR) spectroscopy was used to investigate the spectral differences between non-infested and infested litchi fruits. Significant variations in transmittance were observed at multiple wavelengths, particularly at 630 nm, 815 nm, 875 nm, and 1050 nm. These spectral differences are closely linked to changes in the biochemical composition of the fruit, including hardness, soluble sugars, vitamin C, polyphenols, anthocyanins, and ORAC values, which reflect the impact of pest infestation on the fruit’s internal structure and composition.
Specifically, the higher transmittance peak at 630 nm is associated with increased anthocyanin content and antioxidant capacity in non-infested litchis. Anthocyanins are key natural antioxidants that are directly related to the fruit’s antioxidant properties. In infested litchis, pest damage reduces anthocyanin levels, leading to lower transmittance at this wavelength range [27]. Furthermore, the anthocyanin content correlates with changes in the pigment composition of the fruit peel, further highlighting the differences in peel composition between infested and non-infested fruits.
At 815 nm, the higher spectral peak observed for non-infested litchis indicates greater fruit hardness, suggesting that their cell structure remains intact and unaffected by pest damage. Hardness is a crucial indicator of fruit maturity and health, directly influencing the integrity of the cell wall and the firmness of the flesh. In contrast, infested litchis exhibit reduced hardness due to pest-induced damage to the cell wall, which results in increased absorption in this spectral region [28].
Vitamin C, an essential nutrient in fruits, is closely linked to antioxidant capacity. In this study, it was found that the vitamin C content in infested litchis is lower. The variation in vitamin C levels is reflected in the absorption changes between 630 nm and 815 nm, which are directly associated with the physiological damage caused by pest infestation. Pest activity compromises the integrity of the fruit peel and inhibits vitamin C synthesis, leading to reduced antioxidant capacity in infested litchis. In contrast, non-infested litchis, which experience healthier growing conditions, maintain higher and more stable vitamin C levels, as indicated by more consistent spectral characteristics in these regions [29].
The absorption peak at 875 nm further emphasizes a reduction in soluble sugars and other key biochemical components in infested litchis. Pest infestation disrupts nutrient transport within the fruit, impairing the synthesis and accumulation of sugars and other essential compounds, leading to increased absorption at this wavelength [30]. Soluble sugars are a primary source of energy and flavor in fruits, and their reduction adversely affects both the flavor and overall quality of the fruit.
Regarding polyphenol analysis, spectral differences at 875 nm are also linked to variations in polyphenol content between infested and non-infested litchis. Polyphenols are important antioxidants that provide significant health benefits. Non-infested litchis have a higher polyphenol content, which is associated with their superior antioxidant capacity and resilience to environmental stress. In contrast, pest-induced physiological damage in infested litchis results in a decreased polyphenol content, undermining their antioxidant activity. These variations in polyphenol content are reflected in distinct absorption differences at the 875 nm wavelength, further highlighting the role of polyphenols in distinguishing between infested and non-infested fruits [31].
Additionally, spectral changes at 1050 nm are closely associated with the soluble sugar and water content in the fruit. Infested litchis, due to sugar loss, exhibit higher absorption at this wavelength. This region is particularly sensitive to variations in sugar and moisture content, with reductions in sugars and water loss accelerating fruit shrinkage and maturation, which intensifies the damage caused by pests. Therefore, spectral analysis at 1050 nm provides an effective means of detecting changes in sugar and moisture content in infested fruits.
To minimize the impact of scattering on the spectral data, data preprocessing techniques such as standard normal variate (SNV) and multiplicative scatter correction (MSC) were applied. After noise reduction, the spectral differences between infested and non-infested litchis became more distinct. This preprocessing step effectively mitigated environmental and surface scattering effects, enhancing the differentiation between infested and non-infested fruits and revealing more significant biochemical variations. These findings suggest that by analyzing specific spectral features, we can effectively distinguish between infested and non-infested litchis, providing robust support for non-destructive detection technologies.
3.1.2. Visible/Near-Infrared Spectroscopy Detection Model of Internal Litchi Infestation
In this study, we aimed to improve the accuracy of the non-destructive detection of litchi stem borers using visible/near-infrared (VIS/NIR) spectroscopy. To achieve this, we carefully preprocessed the litchi VIS/NIR spectral data to minimize interference and extract the most relevant wavelengths. Six preprocessing methods were applied: no preprocessing, Savitzky–Golay (SG) smoothing, standard normal variate (SNV), multiplicative scatter correction (MSC), and combinations of SG+SNV and SG+MSC. These methods were selected based on their theoretical advantages in reducing noise, correcting scattering effects, and mitigating baseline drift.
To identify the most representative feature wavelengths from the preprocessed spectral data, three feature selection methods were employed: Competitive Adaptive Reweighted Sampling (CARS), Successive Projections Algorithm (SPA), and Uninformative Variable Elimination (UVE). These methods utilize distinct strategies to identify and select features that significantly influence model performance, thus improving both the predictive power and interpretability of the models.
In the model-building phase, the performance of two regression models, partial least squares regression (PLSR) and support vector regression (SVR), was compared under different preprocessing and feature selection conditions. The experimental results showed that the model using spectral data preprocessed with SG smoothing and SNV correction, combined with SPA for feature selection and PLSR for modeling, achieved the best performance in detecting litchi stem borers. This model produced a coefficient of determination (R2) of 0.82, a root mean square error (RMSE) of 0.18, and an accuracy of 89.22%. These results were attributed to the synergistic advantages of SG smoothing in reducing noise, SNV in correcting scattering effects and baseline drift, SPA in optimizing feature selection, and PLSR in handling high-dimensional data and extracting relevant components.
Although VIS/NIR spectroscopy showed strong performance in detecting most infested fruits, it may lead to misclassification in fruits with shorter infestation periods, as changes in the chemical composition of the flesh and peel may not be sufficiently pronounced (Figure 5 and Table 1).
3.2. Hyperspectral Detection
3.2.1. Hyperspectral Analysis of Internal Litchi Infestation
In this study, hyperspectral imaging was employed to analyze litchi fruits, revealing spectral differences between infested and non-infested fruits within the 400–1000 nm range. However, no significant distinctions were observed in the raw spectra, which can be attributed to hyperspectral reflectance primarily reflecting the properties of the fruit skin and surface tissues, whereas visible/near-infrared spectroscopy is more sensitive to the internal characteristics of the fruit. Hyperspectral reflectance analysis provides valuable insights into the composition of the fruit skin, including polyphenols, anthocyanins, chlorophyll, and moisture content. These factors are closely related to biochemical changes such as fruit hardness, soluble sugars, vitamin C, and ORAC values, especially under pest infestation.
Polyphenols are key antioxidants in litchi fruit skin, typically influencing the spectra in the 400–500 nm range. The polyphenol content in infested fruits is lower, leading to a reduction in antioxidant activity, which results in a flattened absorption peak in this region. In contrast, non-infested fruits, with a higher polyphenol content, exhibit stronger antioxidant activity, reflected by a more pronounced absorption peak in this range [32].
In the 600–650 nm range, the absorption peak for infested fruits is higher than that for non-infested fruits, primarily due to electronic transitions in the C=C and C-O bonds of anthocyanins in the fruit skin. Anthocyanins are important antioxidants in litchi skin, and their content directly correlates with the fruit’s antioxidant capacity. Infested fruits exhibit reduced anthocyanin levels, indicating a decline in antioxidant activity, which is reflected in the spectral differences observed in this wavelength range [33,34]. Moreover, non-infested fruits, which contain higher levels of anthocyanins, typically show a better antioxidant performance and higher ORAC values, resulting in a more prominent absorption peak at this wavelength.
In the 700–800 nm range, the absorption peak for infested fruits is lower than that for non-infested fruits, likely due to a reduction in chlorophyll content caused by pest damage. Chlorophyll, a key component in photosynthesis, not only reflects damage to the fruit skin but also influences the fruit’s hardness and nutritional components, such as vitamin C synthesis. Non-infested fruits, which contain higher levels of chlorophyll, typically exhibit greater fruit hardness and higher vitamin C levels [35,36], both of which are important indicators of fruit quality and health. Pest-induced damage leads to a decrease in chlorophyll content, weakening the fruit skin structure and reducing hardness, which is reflected in the lower absorption peak in this wavelength range.
The absorption peak around 970 nm is likely associated with the stretching vibrations of OH bonds in water. Pest damage to the fruit skin exposes underlying tissues, resulting in an increase in moisture content. Infested fruits exhibit stronger absorption at this wavelength due to higher moisture levels, which also correlates with changes in soluble sugar content. Pest damage leads to a loss of sugars and moisture, accelerating the ripening process and increasing the absorption peak in this range. In contrast, non-infested fruits, with a more stable sugar and moisture content, exhibit relatively weaker absorption at this wavelength [37].
However, since hyperspectral imaging primarily captures the spectral characteristics of the fruit’s skin, the spectral differences observed in the 400–1000 nm range primarily reflect damage caused by stem borers to the skin. The differences in chlorophyll and moisture content between infested and non-infested fruits are relatively small, particularly when no visible pest entry points are evident on the surface. This limitation reduces the overall effectiveness of hyperspectral imaging in detecting litchi pest infestation. Therefore, preprocessing the hyperspectral data using standard normal variate (SNV) and multiplicative scatter correction (MSC), followed by further analysis using partial least squares regression (PLSR) and support vector regression (SVR), is essential in enhancing accuracy in distinguishing between infested and non-infested fruits. This approach improves detection by revealing subtle differences in biochemical indicators such as hardness, soluble sugars, vitamin C, polyphenols, anthocyanins, and ORAC values, thus optimizing the non-destructive detection of litchi pest infestation.
3.2.2. Hyperspectral Detection Model for Internal Litchi Infestation
In this study, we conducted a comprehensive analysis of the hyperspectral spectra of litchis to enhance the accuracy of detecting litchi stem borers using advanced data processing techniques. To minimize spectral interference and extract relevant features, several preprocessing methods were employed, including no preprocessing, Savitzky–Golay (SG) smoothing, standard normal variate (SNV), and multiplicative scatter correction (MSC), as well as combinations of SG+SNV and SG+MSC. The effectiveness of these preprocessing methods was validated through subsequent feature selection and modeling steps.
For feature wavelength selection, three algorithms were used: Competitive Adaptive Reweighted Sampling (CARS), Successive Projections Algorithm (SPA), and Uninformative Variable Elimination (UVE). These algorithms are designed to identify the spectral features most relevant to detecting litchi stem borers, reduce data dimensionality, and improve model performance. CARS enhances the weight of important features by competitive weighting, SPA reduces redundancy through stepwise projections, and UVE eliminates variables with minimal impact on the model.
Furthermore, we compared the performance of partial least squares regression (PLSR) and support vector regression (SVR) under various preprocessing and feature selection conditions. The results revealed that the PLSR model after SG smoothing and MSC preprocessing and combined with SPA for feature selection achieved the best performance on the test set, with a coefficient of determination (R2) of 0.69, a root mean square error (RMSE) of 0.23, and an accuracy of 81.74%. This outcome can be attributed to the synergistic effects of SG smoothing in noise reduction, MSC for scattering and baseline drift correction, and SPA in minimizing feature redundancy and multicollinearity. The PLSR model effectively handles high-dimensional data, extracting the latent components that are most relevant to detecting internal litchi infestations.
However, we acknowledge the limitations of hyperspectral imaging in detecting internal infestations. As hyperspectral imaging primarily captures the spectral characteristics of the skin, its detection capability is limited for internal damage or fruits with healed surfaces. While hyperspectral imaging is effective in identifying external pest damage, its accuracy in detecting deeper infestations requires further improvement (Figure 6 and Table 2).
3.3. X-Ray Imaging Detection
After being damaged and gnawed by stem borers, the fruit stems of infested litchis exhibit cavities and frass, leading to a reduction in density compared to the undamaged areas. X-ray imaging can effectively detect these density variations within the fruit samples and display them as differences in grayscale levels. Background-corrected grayscale values for the fruit stem, seed, and flesh were subjected to principal component analysis (PCA) for the preliminary visualization of the three grayscale datasets, as shown in Figure 7. The first principal component, which represents the grayscale values of the fruit stem, and the second principal component, representing the grayscale values of the seed, accounted for 73% and 27% of the variance, respectively, with a cumulative contribution of nearly 100%. These results indicate that pest-infested and pest-free samples can be distinguished, although the clustering is relatively close, suggesting that further refinement is needed to enhance the separation.
We employed two methods to analyze the X-ray images: (1) The neural network method utilizes a neural network model to classify X-ray images of litchi stems and seeds. Before classification, the images were cropped to focus on the stem regions, and adjustments were made to the brightness and contrast. This method achieved an accuracy of 73.65%. (2) In grayscale value extraction and modeling, grayscale values were extracted from the litchi stem, seed, flesh, and background. Support vector regression (SVR) and partial least squares regression (PLSR) were then applied for modeling. The results showed that the model based on grayscale value extraction combined with SVR outperformed others during testing, achieving a root mean square error (RMSE) of 0.24, a coefficient of determination (R2) of 0.69, and an accuracy of 76.25% [38].
Overall, the X-ray imaging technique demonstrated considerable potential for detecting internal pest damage in litchis, effectively leveraging density differences and grayscale analysis to improve the classification and detection accuracy (Figure 8 and Table 3).
3.4. Multi-Sensor Data Fusion Detection Method
Section 3.1 indicates that the detection model based on visible/near-infrared (VIS/NIR) spectroscopy achieved an identification rate of 89.22% on the test set. Although VIS/NIR spectroscopy performs well in detecting most infested fruits, it is prone to misclassification in cases of short-term infestations, where changes in chemical composition are minimal. The limitation of VIS/NIR spectroscopy lies in its heavy reliance on surface characteristics, which can hinder the accurate detection of early-stage or concealed infestations.
As discussed in Section 3.2, the detection model based on hyperspectral technology achieved an identification rate of 81.74% on the test set of litchi samples. However, hyperspectral imaging primarily detects external infestations with visible entry points, while its effectiveness in identifying internal infestations is limited. This is because stem borers may not penetrate the fruit skin, or the skin may have healed. Therefore, while hyperspectral technology is highly effective in evaluating external fruit quality, it is less reliable in detecting internal quality issues.
Section 3.3 reveals that the model based on X-ray imaging achieved an identification rate of 76.25% on the test set. While X-ray technology is effective in detecting moisture content in the flesh and texture changes in the seed, its ability to detect subtle pest damage is limited, particularly when structural changes in the fruit are minimal. X-ray imaging is constrained by its inability to capture the complex interactions between the fruit’s surface and interior.
To improve the accuracy of detecting litchi stem borers, we propose the integration of spectral and imaging features into machine learning methods to establish a multi-sensor data fusion detection model [39,40]. By combining data from various sensors, we can mitigate the limitations of individual methods and achieve more comprehensive and accurate pest detection.
In this study, we adopted a feature-level fusion approach to detecting litchi stem borers. Initially, data were extracted from various feature sets, including visible/near-infrared (VIS/NIR) spectroscopy, hyperspectral imaging, and X-ray imaging, followed by preprocessing and normalization. The normalized data were then integrated into a unified feature set to enhance the model’s representational capacity. To optimize model performance, Recursive Feature Elimination with Cross-Validation (RFECV) was applied to identify the most relevant features.
After feature selection, a Random Forest (RF) model was employed for training and classification. The use of RFECV within the Random Forest model is essential, as it recursively eliminates irrelevant features, preserving only the most informative ones. This process effectively mitigates overfitting, thereby enhancing the model’s performance and stability. Ultimately, through multi-sensor data fusion, we retained the most relevant features, further optimizing the accuracy of the Random Forest model.
To evaluate the model performance, k-fold cross-validation was applied with k values ranging from 3 to 10. In each fold, the dataset was divided into k subsets, with one subset designated as the test set and the remaining k − 1 subsets used for training. This approach enabled the evaluation of model performance across multiple data partitions, mitigating potential biases arising from any single data split. The optimal k value and corresponding hyperparameters, such as the number of trees, maximum depth, and minimum samples required for node splitting, were selected to fine-tune the Random Forest model, ensuring optimal performance in detecting litchi stem borers.
Principal component analysis (PCA) was applied in each fold to reduce the dimensionality of both the training and testing datasets, retaining 95% of the variance. This approach enhanced the computational efficiency and minimized the model complexity.
During training, the Random Forest model quantified the contribution of each feature to the decision tree splits, thereby identifying the most influential features for model prediction. This feature importance analysis is essential in understanding the role of each sensor (VIS/NIR, hyperspectral, and X-ray) in the model’s decision-making process. Specifically, we assessed feature importance by calculating the Gini impurity or information gain for each feature. Features with greater contributions to the decision-making process were ranked higher in terms of importance. Through this analysis, we identified key features, including both external fruit characteristics (such as color and texture) and internal indicators of pest damage (such as damage density and structural changes observed in hyperspectral and X-ray images). These features played a crucial role in classifying litchi stem borers, with certain features showing a stronger correlation with specific types of pest damage [41,42] (Table 4).
According to Table 5, when k = 8, the model achieved its highest identification rate of 92.39%. The optimal parameters for the Random Forest model were as follows: the maximum depth of the decision trees was set to 20, the minimum number of samples required at each internal node was 5, and the total number of trees was 200. Compared to using visible/near-infrared spectroscopy alone, the detection accuracy improved by 3.17%; compared to hyperspectral imaging alone, the accuracy increased by 10.65%; and compared to X-ray imaging alone, the accuracy improved by 16.14%. The following table compares the performance of the different detection methods.
This analysis not only improves the model’s interpretability but also offers valuable insights for use in practical applications. By identifying the most significant features, we can optimize sensor selection, ensuring that the most relevant sensors are prioritized in future data collection and experimental designs. Furthermore, feature importance analysis informs feature engineering by pinpointing which data sources—such as VIS/NIR, hyperspectral, and X-ray—are most effective for pest detection, thereby refining the data fusion process.
In conclusion, the application of the Random Forest model for feature selection and training, in combination with RFECV and k-fold cross-validation, not only enhanced the model’s accuracy and stability but also increased its practical applicability in detecting litchi stem borers. By optimizing feature selection, fine-tuning hyperparameters, and rigorously validating the model across multiple data partitions, we ensured its robustness and significantly improved its generalization ability.
4. Conclusions
This study investigates non-destructive methods for detecting internal pests in litchi using hyperspectral imaging, X-ray imaging, and visible/near-infrared (VIS/NIR) spectroscopy. Static laboratory measurements were conducted to analyze the structural, optical, and chemical properties of litchis, assessing the efficacy of these methods in pest detection. The key findings are as follows: (1) For VIS/NIR spectral data, the combination of SG smoothing, SNV correction, CARS feature selection, and PLSR modeling achieved the highest performance, with an accuracy of 89.22% in identifying litchi stem borers. (2) For hyperspectral data, the combination of SG smoothing, MSC, SPA feature selection, and PLSR modeling yielded the most accurate results in differentiating infested from non-infested litchis, with an accuracy of 81.74%. (3) X-ray imaging, when combined with grayscale value extraction and SVM modeling, achieved an accuracy of 76.25%. (4) The machine learning model based on multi-sensor data fusion demonstrated the highest accuracy in detecting litchi stem borers, with an accuracy of 92.39%.
These results demonstrate that multi-sensor data fusion, integrating spectral information and machine vision, is not only a viable approach but also significantly improves detection accuracy by 3.17%, 10.65%, and 16.14%, respectively, compared to the use of VIS/NIR spectroscopy, hyperspectral imaging, or X-ray imaging individually.
This study proposes an effective, non-destructive detection method for litchi stem borers by integrating hyperspectral imaging, X-ray imaging, and visible/near-infrared (VIS/NIR) spectroscopy through multi-sensor data fusion. In contrast to existing methods, the novelty of this research lies in addressing the challenges posed by litchi’s complex skin structure. While previous studies have focused primarily on fruits with simpler or more uniform skins, traditional detection methods are often less effective for fruits like litchi, which possess intricate and heterogeneous surface features. Our approach demonstrates the effectiveness of multi-source information fusion in detecting such fruits, thereby expanding the applicability of non-destructive detection technologies to a broader range of fruits and offering valuable insights for the detection of similar fruits in future research.
Future research will focus on optimizing algorithms to enhance the model’s adaptability under varying environmental conditions. We plan to conduct field experiments to assess the impact of environmental factors, such as temperature, humidity, and soil moisture, on the performance of different detection methods, while also working to improve the robustness of the multi-sensor fusion model. Controlled experiments will be conducted to investigate the influence of environmental variables on detection accuracy. Additionally, we aim to integrate other imaging technologies, such as magnetic resonance imaging (MRI) and thermal imaging, to further improve system performance. Furthermore, we plan to extend the application of this technology to a wider range of fruit types to evaluate its generalizability. With the integration of advanced techniques, such as deep learning, we expect to develop more efficient feature extraction and data processing methods, further enhancing the detection accuracy. These innovations will contribute to the advancement of intelligent and automated solutions for fruit and agricultural product quality control, driving progress in automated agricultural inspection technologies.
Resources, W.L.; Data curation, H.F.; Writing—original draft, Z.Z.; Writing—review & editing, S.X., H.L. and X.L. All authors have read and agreed to the published version of the manuscript.
Data are contained within the article.
We confirm that this manuscript has been read and approved by all named authors and that there are no other persons who satisfied the criteria for authorship but are not listed. The authors declare no conflicts of interest.
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Enlarge this image.Figure 1. Schematic diagram of the visible/near-infrared spectroscopy acquisition device.
Enlarge this image.Figure 5. (a) Raw visible/near-infrared spectrum, (b) visible/near-infrared spectrum after SG+SNV preprocessing.
Enlarge this image.Figure 6. (a) Raw hyperspectral spectrum, (b) hyperspectral spectrum after SG+MSC preprocessing.
Enlarge this image.Figure 7. PCA classification of grayscale values in X-ray imaging feature regions for stem-borer-infested and non-infested fruit.
Performance of prediction models based on visible/near-infrared spectroscopy with various preprocessing, feature extraction, and modeling methods.
| Methods | PLSR | SVM | ||||
|---|---|---|---|---|---|---|
| Test Set | Test Set | Test Set | Test Set | Test Set | Test Set | |
| NO | 0.67 | 0.24 | 70.24 | 0.34 | 0.24 | 75.04 |
| SNV | 0.79 | 0.19 | 77.43 | 0.45 | 0.19 | 75.88 |
| MSC | 0.78 | 0.19 | 76.33 | 0.42 | 0.19 | 74.65 |
| SG | 0.37 | 0.22 | 70.54 | 0.36 | 0.2 | 74.43 |
| SG+SNV | 0.79 | 0.18 | 78.62 | 0.49 | 0.18 | 76.53 |
| SG+MSC | 0.78 | 0.18 | 76.53 | 0.45 | 0.18 | 76.53 |
| SG+SNV+CARS | 0.82 | 0.17 | 89.22 | 0.52 | 0.18 | 85.52 |
| SG+SNV+UVE | 0.81 | 0.18 | 84.63 | 0.51 | 0.18 | 85.32 |
| SG+SNV+SPA | 0.81 | 0.18 | 85.79 | 0.47 | 0.19 | 83.83 |
Performance of hyperspectral-based prediction models with various preprocessing, feature extraction, and modeling methods.
| Methods | PLSR | SVM | ||||
|---|---|---|---|---|---|---|
| Test Set | Test Set | Test Set | Test Set | Test Set | Test Set | |
| NO | 0.66 | 0.24 | 79.04 | 0.16 | 0.25 | 79.76 |
| SNV | 0.66 | 0.24 | 81.14 | 0.13 | 0.25 | 80.95 |
| MSC | 0.67 | 0.24 | 80.84 | 0.11 | 0.25 | 80.95 |
| SG | 0.68 | 0.23 | 80.24 | 0.11 | 0.26 | 79.76 |
| SG+SNV | 0.67 | 0.24 | 80.24 | 0.11 | 0.25 | 79.76 |
| SG+MSC | 0.67 | 0.24 | 81.21 | 0.14 | 0.25 | 80.95 |
| SG+MSC+CARS | 0.67 | 0.24 | 77.25 | 0.11 | 0.26 | 80.95 |
| SG+MSC+UVE | 0.64 | 0.25 | 79.94 | 0.15 | 0.26 | 79.76 |
| SG+MSC+SPA | 0.69 | 0.23 | 81.74 | 0.11 | 0.25 | 79.76 |
Performance of X-ray imaging models established with different preprocessing and modeling methods.
| Methods | R2 | RMSE | Accuracy/% |
|---|---|---|---|
| Gray value extraction combined with PLSR | 0.67 | 0.24 | 74.84% |
| Gray value extraction combined with SVM | 0.69 | 0.22 | 76.25% |
| AlexNet architecture | / | / | 69.65% |
Model performance and parameter optimization for different k values in k-fold cross-validation.
| K | Average Accuracy/% | Optimal Parameters | ||
|---|---|---|---|---|
| Maximum Depth of the Decision Tree | Number of Internal Nodes | Number of Trees | ||
| 3 | 90.67 | 20 | 5 | 200 |
| 4 | 91.39 | Unlimited | 5 | 200 |
| 5 | 91.15 | Unlimited | 5 | 200 |
| 6 | 91.62 | 20 | 5 | 200 |
| 7 | 91.85 | 10 | 5 | 200 |
| 8 | 92.39 | 20 | 5 | 200 |
| 9 | 91.63 | 20 | 5 | 100 |
| 10 | 91.14 | Unlimited | 2 | 100 |
Comparison of results from various detection methods.
| Detection Methods | Preprocessing Methods | Feature Selection | Modeling Methods | Accuracy/% |
|---|---|---|---|---|
| Vis-NIR Spectroscopy | SG+SNV | CARS | PLSR | 89.22% |
| Hyperspectral Imaging | SG+MSC | SPA | PLSR | 81.74% |
| X-ray Imaging | Noise Reduction, Brightness Enhancement | Grayscale Value Extraction | SVM | 76.25% |
| Multi-Sensor Detection | Feature Fusion, Normalization | RFECV | RF | 92.39% |
References
1. Li, L.; Jia, X.; Fan, K. Recent advance in nondestructive imaging technology for detecting quality of fruits and vegetables: A review. Crit. Rev. Food Sci. Nutr.; 2024; pp. 1-19. [DOI: https://dx.doi.org/10.1080/10408398.2024.2404639] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39291966]
2. Xu, S.; Ren, J.; Lu, H.; Wang, X.; Sun, X.; Liang, X. Nondestructive detection and grading of flesh translucency in pineapples with visible and near-infrared spectroscopy. Postharvest Biol. Technol.; 2022; 192, 112029. [DOI: https://dx.doi.org/10.1016/j.postharvbio.2022.112029]
3. Xu, S.; Lu, H.; Fan, C.; Qiu, G.; Ference, C.; Liang, X.; Peng, J. Visible and near-infrared hyperspectral imaging as an intelligent tool for parasite detection in sashimi. LWT-Food Sci. Technol.; 2023; 181, 114747. [DOI: https://dx.doi.org/10.1016/j.lwt.2023.114747]
4. Tempelaere, A.; He, J.; Van Doorselaer, L.; Verboven, P.; Nicolai, B.; Giuffrida, M.V. Unsupervised anomaly detection for pome fruit quality inspection using X-ray radiography. Comput. Electron. Agric.; 2024; 226, 109364. [DOI: https://dx.doi.org/10.1016/j.compag.2024.109364]
5. Zhang, R.; Wang, M.; Liu, P.; Zhu, T.; Qu, X.; Chen, X.; Xiao, X. Flexible Vis/NIR sensing system for banana chilling injury. Postharvest Biol. Technol.; 2024; 207, 112623. [DOI: https://dx.doi.org/10.1016/j.postharvbio.2023.112623]
6. Han, H.; Sha, R.; Dai, J.; Wang, Z.; Mao, J.; Cai, M. Garlic Origin Traceability and Identification Based on Fusion of Multi-Source Heterogeneous Spectral Information. Foods; 2024; 13, 1016. [DOI: https://dx.doi.org/10.3390/foods13071016]
7. Gu, S.; Jing, M.; Li, D.; Ma, Z.; Duan, Y.; Wang, L.; Dai, X.; Chen, Z.; Zhang, X.; Chen, J. The effects of different temperature and humidity conditions on the ripening and cracking of Annona atemoya fruit during storage by regulating the conversion of starch into soluble sugars. LWT; 2024; 208, 116703. [DOI: https://dx.doi.org/10.1016/j.lwt.2024.116703]
8. Richter Reis, F.; de Oliveira, A.C.; Gadelha, G.G.P.; de Abreu, M.B.; Soares, H.I. Vacuum Drying for Extending Litchi Shelf-Life: Vitamin C, Total Phenolics, Texture and Shelf-Life Assessment. Plant Foods Hum. Nutr.; 2017; 72, pp. 120-125. [DOI: https://dx.doi.org/10.1007/s11130-017-0602-9]
9. Xiao, Q.; Ye, S.; Wang, H.; Xing, S.; Zhu, W.; Zhang, H.; Zhu, J.; Pu, C.; Zhao, D.; Zhou, Q. et al. Soluble sugar, organic acid and phenolic composition and flavor evaluation of plum fruits. Food Chem. X; 2024; 24, 101790. [DOI: https://dx.doi.org/10.1016/j.fochx.2024.101790]
10. Mujanović, I.; Balijagić, J.; Bajagić, M.; Poštić, D.; Đurović, S. Variations in polyphenol content and anthocyanin composition in bilberry populations (Vaccinium myrtillus L.) due to environmental factors. J. Food Compos. Anal.; 2024; 136, 106732. [DOI: https://dx.doi.org/10.1016/j.jfca.2024.106732]
11. Su, D.; Ti, H.; Zhang, R.; Zhang, M.; Wei, Z.; Deng, Y.; Guo, J. Structural elucidation and cellular antioxidant activity evaluation of major antioxidant phenolics in lychee pulp. Food Chem.; 2014; 158, pp. 385-391. [DOI: https://dx.doi.org/10.1016/j.foodchem.2014.02.134] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24731359]
12. Jie, Z.; Hou, X.; Wang, J.; Zhang, W.; Zhang, A. Non-destructive detection of adulteration of weight-loss drugs in the field of spectral feature fusion. Infrared Phys. Technol.; 2023; 131, 104591. [DOI: https://dx.doi.org/10.1016/j.infrared.2023.104591]
13. Zhang, Z.; Chang, Z.; Huang, J.; Leng, G.; Xu, W.; Wang, Y.; Xie, Z.; Yang, J. Enhancing soil texture classification with multivariate scattering correction and residual neural networks using visible near-infrared spectra. J. Environ. Manag.; 2024; 352, 120094. [DOI: https://dx.doi.org/10.1016/j.jenvman.2024.120094] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38237335]
14. Xia, K.; Cheng, Q.; Xia, S.; Zhang, S.; Shen, Q.; Li, C.; Zhou, J. Optimization of a Soil Particle Content Prediction Model Based on a Combined Spectral Index and Successive Projections Algorithm Using Vis-NIR Spectroscopy. Spectroscopy; 2020; 35, pp. 24-34.
15. Wahyudi, I.; Munawar, A.A.; Yu, P.; Samadi, S. Optical characterization of NIR spectra for chemomectric model of cocoa pod husk fermented for animal feed. IOP Conf. Ser. Earth Environ. Sci.; 2023; 1183, 012003. [DOI: https://dx.doi.org/10.1088/1755-1315/1183/1/012003]
16. Ding, D.; Yu, H.; Yin, Y.; Yuan, Y.; Li, Z.; Li, F. Determination of Chlorophyll and Hardness in Cucumbers by Raman Spectroscopy with Successive Projections Algorithm (SPA)—Extreme Learning Machine (ELM). Anal. Lett.; 2023; 56, pp. 1216-1228. [DOI: https://dx.doi.org/10.1080/00032719.2022.2123922]
17. Yang, D.; Hu, J. A detection method of oil content for maize kernels based on CARS feature selection and deep sparse autoencoder feature extraction. Ind. Crops Prod.; 2024; 222, 119464. [DOI: https://dx.doi.org/10.1016/j.indcrop.2024.119464]
18. Lu, Y.; Du, C.; Yu, C. Determination of the contents of magnesium and potassium in rapeseeds using FTIR-PAS combined with least squares support vector machines and uninformative variable elimination. Anal. Methods; 2014; 6, pp. 2586-2591. [DOI: https://dx.doi.org/10.1039/C3AY41460D]
19. Rong, L.; Wang, Y.; Wang, Y.; Jiang, D.; Bai, J.; Wu, Z.; Li, L.; Wang, T.; Tan, H. A fresh-cut papaya freshness prediction model based on partial least squares regression and support vector machine regression. Heliyon; 2024; 10, e30255. [DOI: https://dx.doi.org/10.1016/j.heliyon.2024.e30255]
20. Tan, F.; Mo, X.; Ruan, S.; Yan, T.; Xing, P.; Gao, P.; Xu, W.; Ye, W.; Li, Y.; Gao, X. et al. Combining Vis-NIR and NIR Spectral Imaging Techniques with Data Fusion for Rapid and Nondestructive Multi-Quality Detection of Cherry Tomatoes. Foods; 2023; 12, 3621. [DOI: https://dx.doi.org/10.3390/foods12193621]
21. Chawla, P.K.; Nair, M.S.; Malkhede, D.G.; Patil, H.Y.; Jindal, S.K.; Chandra, A.; Gawas, M.A. Parkinson’s disease classification using nature inspired feature selection and recursive feature elimination. Multimed. Tools Appl.; 2023; 83, pp. 35197-35220. [DOI: https://dx.doi.org/10.1007/s11042-023-16804-w]
22. Sun, J. Application and optimization of RFECV feature selection and random forest prediction model. Digit. Commun. World; 2024; 9, pp. 114-116.
23. Gong, B.; An, A.; Shi, Y.; Jia, W. Fault diagnosis of photovoltaic array with multi-module fusion under hyperparameter optimization. Energy Convers. Manag.; 2024; 319, 118974. [DOI: https://dx.doi.org/10.1016/j.enconman.2024.118974]
24. Zhu, S.J. Research on Data Ownership Methods for Multi-Source Information Fusion Systems Based on Machine Learning. Ph.D Thesis; Xi’an University of Technology: Xi’an, China, 2024.
25. Tchakoucht, T.A.; Elkari, B.; Chaibi, Y.; Kousksou, T. Random forest with feature selection and K-fold cross validation for predicting the electrical and thermal efficiencies of air based photovoltaic-thermal systems. Energy Rep.; 2024; 12, pp. 988-999. [DOI: https://dx.doi.org/10.1016/j.egyr.2024.07.002]
26. Harif, A.; Kassimi, A.M. Predictive Modeling of Student Performance Using RFECV-RF for Feature Selection and Machine Learning Techniques. Int. J. Adv. Comput. Sci. Appl.; 2024; 15, pp. 231-240. [DOI: https://dx.doi.org/10.14569/IJACSA.2024.0150723]
27. Zhang, L.; Xia, Q.; Chen, J.; Cao, Y.; Guan, R.; Huang, H. Determination of Anthocyanin Content in Blueberry Pomace Using Three Near-Infrared Spectroscopy Methods. Spectrosc. Spectr. Anal.; 2020; 40, pp. 2246-2252.
28. Ding, G.; Jin, K.; Chen, X.; Li, A.; Guo, Z. Non-destructive prediction of ready-to-eat kiwifruit firmness based on Fourier transform near-infrared spectroscopy. Postharvest Biol. Technol.; 2024; 212, 112908. [DOI: https://dx.doi.org/10.1016/j.postharvbio.2024.112908]
29. Hafsah, S.; Ahmad, F.; Arianti, N.D.; Saputra, E.; Hartuti, S. Rapid and non-destructive determination of vitamin C and antioxidant activity of intact red chilies using visible near-infrared spectroscopy and machine learning tools. Case Stud. Chem. Environ. Eng.; 2023; 8, 100435.
30. Yang, Q.; Yang, X.; Zhang, Q.; Wang, Y.; Song, H.; Huang, F. Quantifying soluble sugar in super sweet corn using near-infrared spectroscopy combined with chemometrics. Optik; 2020; 220, 165128. [DOI: https://dx.doi.org/10.1016/j.ijleo.2020.165128]
31. Amoghin, M.L.; Gilandeh, Y.A.; Tahmasebi, M.; Arribas, J.I. Automatic non-destructive estimation of polyphenol oxidase and peroxidase enzyme activity levels in three bell pepper varieties by Vis/NIR spectroscopy imaging data based on machine learning methods. Chemom. Intell. Lab. Syst.; 2024; 250, 105137. [DOI: https://dx.doi.org/10.1016/j.chemolab.2024.105137]
32. Nan, X.; Zhou, Q.; Li, Y.; Li, J.; Wang, H.; Xu, W.; Liu, Y.; Ni, Y.; Guo, T.; Ji, W. et al. Maturity Discrimination and Polyphenol Content Detection Model of Black-fruited Mountain Ash Based on Hyperspectral Technology. Food Ind. Technol.; 2023; 44, pp. 292-301.
33. Wu, D. Detection of Anthocyanin Content in the Skin of Wine Grapes Based on Hyperspectral Imaging Technology; Northwest A&F University: Xianyang, China, 2014.
34. Qi, J.; Sun, S.; Zhang, L.; Zhu, Y.; Zhou, H.; Gan, X.; Li, B.; Chen, Y.; Li, W.; Li, T. et al. Seasonal variation of antioxidant bioactive compounds in southern highbush blueberry leaves and non-destructive quality prediction in situ by a portable near-infrared spectrometer. Food Chem.; 2024; 457, 139925. [DOI: https://dx.doi.org/10.1016/j.foodchem.2024.139925] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38917567]
35. Wang, Z.; Wu, S.; Zuo, C.; Jiang, M.; Song, J.; Ding, F.; Tu, K.; Lan, W.; Pan, L. Exploring the variability and heterogeneity of apple firmness using visible and near-infrared hyperspectral imaging. LWT; 2024; 192, 115704. [DOI: https://dx.doi.org/10.1016/j.lwt.2023.115704]
36. He, H.J.; Zhang, C.; Bian, X.; An, J.; Wang, Y.; Ou, X.; Kamruzzaman, M. Improved prediction of vitamin C and reducing sugar content in sweetpotatoes using hyperspectral imaging and LARS-enhanced LASSO variable selection. J. Food Compos. Anal.; 2024; 132, 106350. [DOI: https://dx.doi.org/10.1016/j.jfca.2024.106350]
37. Zhang, C.; Liu, F.; Kong, W.; Cui, P.; He, Y.; Zhou, W. Estimation and Visualization of Soluble Sugar Content in Oilseed Rape Leaves Using Hyperspectral Imaging. Trans. ASABE; 2016; 59, pp. 1499-1505.
38. Garcia, A.; Martinez, G.; Gomez, J. X-ray imaging for the detection of insect damage in fruits: A novel approach. Postharvest Biol. Technol.; 2021; 175, 111456.
39. Xu, J.T.; Zhang, Z.; Guo, Y.H.; Liu, Y.F.; Ning, X.F. Detection of Cucumber Powdery Mildew Based on Spectral and Image Information. J. Biosyst. Eng.; 2023; 48, pp. 115-122. [DOI: https://dx.doi.org/10.1007/s42853-023-00178-w]
40. Li, J.J.; Qu, Z.P.; Wang, Y.W.; Guo, J. Research on multi-parameter fusion non-invasive blood glucose detection method based on machine learning. Eur. Rev. Med. Pharmacol. Sci.; 2022; 26, pp. 6040-6049.
41. Al-Abdaly, N.M.; Al-Taai, S.R.; Imran, H.; Ibrahim, M. Development of prediction model of steel fiber-reinforced concrete compressive strength using random forest algorithm combined with hyperparameter tuning and k-fold cross-validation. East. Eur. J. Enterp. Technol.; 2021; 5, pp. 59-65. [DOI: https://dx.doi.org/10.15587/1729-4061.2021.242986]
42. Kim, J.H.; Lee, D.H.; Mendoza, J.A.; Lee, M.Y. Applying machine learning random forest (RF) method in predicting the cement products with a co-processing of input materials: Optimizing the hyperparameters. Environ. Res.; 2024; 248, 118300. [DOI: https://dx.doi.org/10.1016/j.envres.2024.118300]
© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.