Hyperspectral dry pea thermal damage identification and quality inversion method fused with multilayer wavelet decomposition

Abstract

The invention discloses a hyperspectral dry pea thermal damage identification method fused with multilayer wavelet decomposition characteristics. According to the method, spectral information of dry peas within the range of 400-1000 nm is obtained through a hyperspectral imaging technology, seven-layer decomposition is carried out on an original spectrum by adopting a db6 wavelet function, and multi-scale detail features are extracted. And an optimal wavelet layer is screened in combination with correlation analysis, and a multi-scale fusion feature set is constructed, so that the discrimination capability of the model on the heat loss grade is improved. And carrying out modeling analysis on the features by adopting a plurality of classification models such as PLS, SVM, LR and RF, and comparing the recognition performance of the features under different feature combinations. According to the method, redundant information interference is effectively reduced, the expression ability of the spectrum to the thermal damage characteristics is enhanced, and the accuracy and stability of the model are remarkably improved. The method has the advantages of non-contact, high throughput and intelligence, and is suitable for heat damage identification and quality grading of agricultural products such as dry peas.

Description

Technical Field

[0001] The present invention relates to the field of optical response detection, and in particular to a method for spectral recognition and quality inversion of heat-damaged dry peas based on multi-layer wavelet decomposition and correlation analysis. Background Art

[0002] Dry peas are an important food and feed crop rich in protein, starch, and various micronutrients, and are widely used in food processing, plant protein extraction, and livestock and poultry feed. However, during long-distance transportation and storage, dry peas are susceptible to environmental influences such as high temperature and high humidity, resulting in varying degrees of thermal damage. This, in turn, increases mold and deteriorates quality, seriously affecting their commercial value and processing efficiency. Therefore, establishing an efficient, non-destructive, and intelligent dry pea quality identification and grading technology is of great significance for ensuring the quality and safety of agricultural products and improving the efficiency of import and export supervision.

[0003] Currently, dry pea quality testing still relies primarily on visual inspection and physical and chemical analysis. While these methods have some value, they suffer from low efficiency, subjectivity, time consumption, and the inability to perform online testing. These limitations make them unable to meet the modern agricultural industry's demand for "fast, accurate, and intelligent" testing. With the advancement of optical sensing and information processing technologies, methods based on visible-near-infrared hyperspectral imaging are becoming an important tool for nondestructive testing of agricultural products. This technology simultaneously captures both spatial image and spectral reflectance information of a sample, enabling a comprehensive assessment of its internal quality status. It offers the advantages of nondestructiveness, high throughput, and simultaneous multi-parameter analysis.

[0004] Although hyperspectral imaging technology has shown great potential in the field of agricultural product testing, its application in seed crops such as dry peas still faces many challenges. First, hyperspectral data has problems such as high dimensionality, high redundant information, and significant noise interference. Direct modeling often leads to complex models, heavy computational burdens, and poor robustness. Secondly, the thermal damage characterization signal of dry peas is weak and may only appear in certain specific bands or subtle spectral features. Traditional dimensionality reduction and modeling methods are difficult to effectively extract key features. How to extract feature information that is highly correlated with quality from high-dimensional spectra is the key to improving modeling accuracy and efficiency.

[0005] In recent years, wavelet analysis technology has been gradually introduced into the field of hyperspectral signal processing due to its excellent time-frequency localization characteristics and multi-scale decomposition capabilities. Wavelet transforms can decompose the original spectrum into approximate coefficients and detail coefficients at different scales, helping to separate low-frequency information from high-frequency noise and enhancing the model's ability to identify local features. The db6 wavelet function in the Daubechies series, in particular, performs well in signal compression and denoising, making it suitable for processing complex agricultural product spectral signals. However, wavelet decomposition alone still struggles to accurately identify the features most strongly correlated with the target indicator. Therefore, introducing correlation analysis and performing feature selection on the decomposed multi-scale coefficients can further screen spectral dimensions significantly associated with dry pea quality, thereby improving modeling efficiency.

[0006] Therefore, the development of a hyperspectral intelligent recognition method for dry peas based on the combination of wavelet decomposition and correlation analysis has important theoretical significance and broad application prospects. It can effectively improve the intelligent quality inspection and grading capabilities of agricultural products and promote the development of intelligent agricultural detection technology. Summary of the Invention

[0007] To address these issues, the present invention aims to address the technical challenges of low detection efficiency and poor model robustness in existing dry pea quality identification methods. A method for spectral identification and quality inversion of heat-damaged dry peas based on multi-layer wavelet decomposition and correlation analysis is proposed. This method enables rapid, non-destructive identification of the degree of heat damage in dry peas and accurate inversion of their quality indicators, providing effective technical support for intelligent quality inspection and grading of agricultural products. To achieve these objectives, the present invention provides the following solutions:

[0008] The present invention proposes a dry pea spectral recognition and quality inversion method based on multi-layer wavelet decomposition and correlation analysis, which is used to achieve efficient identification and quality characteristic information extraction of dry peas with different degrees of heat damage. It has the advantages of non-contact, rapid, high-throughput, and intelligent.

[0009] The method comprises the following steps:

[0010] Step 1: Select imported Russian dry pea samples provided by Tianjin Customs and remove defective kernels such as those with cracks, insect damage, and mildew, ultimately obtaining a complete sample of 4,500 kernels.

[0011] Step 2: The samples were divided into three groups, each containing 1,500 grains. One group was set as the control group, and the other two groups were subjected to heat damage treatment for 20 and 30 days, respectively, to simulate samples with different heat damage levels.

[0012] Step 3: Visible-near-infrared hyperspectral images of the three groups of samples were collected and the acquired spectral data were subjected to multi-layer wavelet decomposition. The db6 wavelet function was used to perform a seven-layer decomposition to extract detailed information at different scales for correlation analysis. Based on the strength of the correlation between each layer and the target variable, the three wavelet decomposition layers with the highest correlation were selected, and their characteristic coefficients were extracted separately or in combination as modeling inputs.

[0013] Step 4: Based on the characteristic coefficients extracted from each wavelet decomposition layer, a dry pea recognition model is established;

[0014] In step five, visible-near-infrared hyperspectral image data of the dry pea sample to be tested is collected and preprocessed according to the parameters described in step three. The corresponding wavelet decomposition features are then extracted. These extracted features are then input into the established dry pea thermal damage identification model, which automatically outputs a thermal damage classification result, enabling rapid and intelligent identification of the sample's thermal damage status. Preferably, after outputting the thermal damage classification result in step five, each pixel or region in the hyperspectral image is assigned a different color based on the classification label, generating a visualization of the thermal damage classification, enabling identification of dry pea thermal damage and image visualization inversion analysis.

[0015] Preferably, imported Russian dry pea samples provided by Tianjin Customs are selected, which have a full appearance, uniform color, and no surface defects such as cracks, insect bites, and mildew, to ensure the consistency of sample quality.

[0016] Preferably, a visible-near-infrared hyperspectral imaging system is used to collect spectral images of dry peas in the 400–1000 nm band, and the acquisition parameters are set as follows: lens-sample object distance of 300 mm, spectral resolution of 1.43 nm, scanning speed of 6.5 mm / s, image size of 804×1097 pixels, and window smoothing points of 3. After threshold segmentation of each dry pea, the spectral data of all pixels in the region of interest are extracted. The db6 wavelet function is used to perform a seven-layer wavelet decomposition on the original spectral data. After extracting the detail coefficients of each layer, the correlation analysis method is used to screen the feature layers and coefficient information that are highly correlated with the thermal damage level. Based on the screened feature data, partial least squares regression (PLS), support vector machine (SVM), logistic regression (LR), and random forest (RF) models are established to identify and inversely predict the thermal damage level and quality indicators of dry peas.

[0017] Preferably, modeling and calculation are The data were completed on MathWorks Inc. (R2019b, USA) and PyTorch (1.13.1, Meta Platforms Inc., USA) platforms.

[0018] According to the specific embodiment of the present invention, the following technical effects are disclosed:

[0019] This study, based on a visible-near-infrared hyperspectral imaging system, effectively extracts multi-scale feature information related to dry pea heat damage and quality status through multi-layer wavelet decomposition combined with correlation analysis. The results significantly improve the model's prediction accuracy. The study further identifies key response bands at varying degrees of heat damage, eliminates redundant or noisy information, and enhances the hyperspectral data's ability to capture dry pea quality changes. This study provides theoretical support and an applied foundation for quality detection and intelligent grading of seed crops such as dry peas. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a method for dry pea spectral recognition and quality inversion based on multi-layer wavelet decomposition and correlation analysis.

[0021] Figure 2 These are real RGB images of dry peas at different heat damage times.

[0022] Figure 3 The detail coefficients of each layer after seven-layer wavelet decomposition of dry pea hyperspectral data using db6 wavelet function (taking normal dry pea spectrum decomposition as an example).

[0023] Figure 4 This is a correlation result analysis chart.

[0024] Figure 5 This is the inversion visualization result diagram. DETAILED DESCRIPTION

[0025] The technical solution of the present invention is further described in detail with reference to the following specific examples.

[0026] The purpose of this invention is to propose a dry pea spectral recognition and quality inversion method based on multi-layer wavelet decomposition and correlation analysis, which is used to achieve efficient identification and quality information extraction of dry peas with different degrees of heat damage, and has the advantages of non-contact, fast and intelligent.

[0027] In order to intuitively demonstrate the above-mentioned objects, features and advantages, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] like Figure 1 As shown, the present invention proposes a dry pea spectral recognition and quality inversion method based on multi-layer wavelet decomposition and correlation analysis, which is used to achieve efficient recognition and quality information extraction of dry peas with different degrees of heat damage, and has the advantages of non-contact, fast, and intelligent.

[0029] The method comprises the following steps:

[0030] Step 1: Select imported Russian dry pea samples provided by Tianjin Customs and remove defective kernels such as those with cracks, insect damage, and mildew, ultimately obtaining a complete sample of 4,500 kernels.

[0031] Step 2: Divide the samples into 3 groups, 1500 grains in each group. One group is set as the control group, and the other two groups are subjected to heat damage treatment for 20 days and 30 days respectively to simulate samples with different heat damage levels. Figure 2 These are RGB images of dry peas at different heat damage times;

[0032] Step 3: Visible-near-infrared hyperspectral images of the three groups of samples were collected and the acquired spectral data were subjected to multi-layer wavelet decomposition. The db6 wavelet function was used to perform a seven-layer decomposition to extract detailed information at different scales for correlation analysis. Based on the strength of the correlation between each layer and the target variable, the three wavelet decomposition layers with the highest correlation were selected, and their characteristic coefficients were extracted separately or in combination as modeling inputs.

[0033] Step 4: Based on the characteristic coefficients extracted from each wavelet decomposition layer, multiple classification models are established to identify samples with different heat loss levels and invert quality parameters;

[0034] Step 5: Compare the differences in classification accuracy between different decomposition layers and model combinations, and visualize the recognition and prediction results.

[0035] Among them, step one and step two specifically include:

[0036] Imported Russian dry pea samples provided by Tianjin Customs were selected. First, poor quality kernels with obvious appearance defects such as cracks, insect bites, and mildew were removed, and healthy samples with intact, plump appearance were retained. After preliminary screening, a total of 4,500 intact dry pea kernels were obtained. The samples were evenly divided into three groups, each containing 1,500 kernels. Groups 1 and 2 were used to simulate varying degrees of heat damage, while Group 3 served as a control group stored at room temperature. Groups 1 and 2 were sealed in 100 mL wide-mouth bottles and stored in a 50°C thermostat for 20 and 30 days, respectively, to obtain samples with mild and severe heat damage. Group 3 samples were sealed and stored at room temperature for comparative analysis. After the heat treatment, the three groups of samples were uniformly numbered and grouped for subsequent spectral data acquisition and heat damage classification modeling.

[0037] Step three specifically includes:

[0038] Three groups of dry pea samples (normal temperature control, 20-day heat damage, and 30-day heat damage) were used as the research subjects, and visible-near-infrared (VIS-NIR) hyperspectral image data were collected. Before hyperspectral image acquisition, the stability of the light source should be ensured and the influence of the spectrometer itself should be eliminated. Therefore, the hyperspectral instrument was preheated for 30 minutes before image acquisition. The optimal object distance between the hyperspectral lens and the sample was first determined, and then the various hyperspectral data acquisition parameters were adjusted. The specific parameters were: ImSpector V10E hyperspectrometer (Specim, Finland, 380-1010 nm), lens-to-sample object distance of 300 mm, spectral resolution of 1.43 nm, scan speed of 6.5 mm / s, exposure time of 4 ms, and image size of 804 × 1097.

[0039] After collecting hyperspectral information for all dry pea samples, we scanned a Teflon white plate (99.99% reflectivity) under the same acquisition conditions to obtain a completely white calibration plate image. The camera lens cap was then used to obtain a completely black background image. A black-white correction method was used to reduce the effects of the instrument's dark current and the sample's own reflection of the light source. The black-white correction formula is:

[0040]

[0041] Where: R represents the corrected signal strength, R0 represents the original signal strength, B represents the calibration signal strength of full black, and W represents the calibration signal strength of full white.

[0042] To fully explore the detailed features of spectral data at different scales, this study used a multi-layer discrete wavelet transform (DWT) to process the raw hyperspectral data. db6 from the Daubechies wavelet family was selected as the mother wavelet, and a seven-layer decomposition was performed on the dry pea spectrum. Figure 3 The paper shows the variation of detail coefficients extracted from each layer after seven-layer decomposition using the db6 wavelet function, taking the spectrum of normal dry peas as an example, reflecting the distribution of spectral features at different scales. The method was implemented on the MATLAB R2019b (MathWorks Inc., USA) platform and used its built-in function wavedec to perform a seven-layer decomposition operation on each spectral data to obtain the detail coefficients (Detail coefficients) of different decomposition layers.

[0043] For each layer of decomposition results, its detail coefficient is extracted as feature information and correlation analysis is performed with the corresponding heat loss level of the sample. The Pearson correlation coefficient between the wavelet detail coefficients of the 1st to 7th layers and the original spectrum is calculated and displayed in the form of a heat map, as shown in Figure 4 As shown in the correlation matrix, the seventh layer's detail coefficients have the highest correlation with the original spectrum, reaching a coefficient of 0.4828, indicating that this layer's features are highly representative in preserving the original spectral information. The sixth and fifth layers' detail coefficients have correlation coefficients of 0.2383 and 0.1334 with the original spectrum, respectively, also exhibiting a certain degree of positive correlation, suggesting potential for use as feature variables.

[0044] In contrast, the detail coefficients of layers 1 through 4 exhibit low correlations with the original spectrum, with some even exhibiting a weak negative correlation. This suggests that the features extracted from these layers may contain more high-frequency noise information, contributing less to the effectiveness of subsequent modeling. Therefore, in the subsequent feature construction process, the wavelet detail coefficients of layers 5 through 7, which have a strong correlation with the original spectrum, are preferentially selected as modeling input variables. The detail coefficients in these three layers are further processed to extract individual features or combine them to form high-dimensional feature vectors, which serve as input variables for subsequent modeling analysis. This fully preserves the discriminative spectral feature information while reducing redundancy and noise interference.

[0045] Step 4 specifically includes:

[0046] Using the detail feature coefficients extracted from different wavelet decomposition layers, various classification models were constructed to identify dry pea samples of varying heat damage grades and to inversely predict their internal quality parameters. Least squares discriminant analysis (PLS), support vector machine (SVM), logistic regression (LR), and random forest (RF) models were used to comprehensively evaluate the adaptability and classification performance of the different algorithms in processing multi-scale spectral features. During the modeling process, three groups of samples (control, mild heat damage, and severe heat damage) were used as classification labels, and the extracted wavelet detail coefficients were used as input features. To examine the effectiveness of each layer's features, separate models were constructed based on features from layers 5 to 7. Furthermore, these features were combined as fusion input to explore the impact of multi-scale information on model performance. Model construction and training were implemented in the PyTorch deep learning framework (version 1.13.1, Meta Platforms Inc., USA), leveraging its flexible data loading and model encapsulation mechanisms to achieve a unified modeling process. A five-fold cross-validation strategy was employed during training to ensure the robustness and generalization of the classification models. In addition, we use model accuracy metrics (accuracy, precision, and recall) to evaluate the pros and cons of different algorithms and feature combinations and identify the optimal modeling strategy. (Note: Accuracy measures the proportion of correct predictions overall; precision measures the proportion of predicted positives that are actually positive; and recall measures the proportion of actual positives that are correctly predicted.)

[0047] Step five specifically includes:

[0048] The analysis results, as shown in Table 1, show that the PLS model achieved relatively high recognition accuracy using raw spectral data (79.17% for the calibration set and 78.44% for the prediction set), essentially distinguishing the three heat loss grades. Model accuracy showed some variation as wavelet decomposition was used to extract features. Among the single-layer features, the model constructed using the seventh-layer detail coefficients performed best, achieving a prediction set accuracy of 78.44%, comparable to the raw spectra, with slightly improved recall and precision. The sixth layer performed second, while the fifth-layer features performed slightly worse, likely due to their excessive granularity and the significant influence of noise. Regarding feature fusion, combining the fifth to seventh layers of features slightly improved model performance. Prediction accuracy remained at 78.22%, but recall increased to 73.73%, indicating that feature fusion enhances model stability to a certain extent. However, increasing the number of fusion layers (e.g., combining the fifth and seventh layers) resulted in a decrease in performance, suggesting that blindly increasing dimensionality can introduce redundant information.

[0049] Table 1: Classification effect evaluation table of feature optimization based on wavelet layer combination in PLS model

[0050]

[0051] The SVM model has the most significant feature extraction capability among all combinations. As shown in Table 2, the prediction accuracy is 91.22% with the original spectral input. After fusing the features of the fifth and seventh wavelet layers, the prediction accuracy increases to 96.22%. The recall and precision rates are both over 96%, far exceeding the PLS model. When the fifth layer is used alone, the prediction set accuracy is as high as 92.78%, indicating that this layer contains strong discriminant information. Although the effects of the sixth and seventh layers used alone are slightly lower than the fifth layer, they are both better than the original data, indicating that wavelet features can effectively extract key information. After multi-layer fusion (fifth + sixth + seventh layers), the accuracy decreases slightly (92.56%), which may be due to the risk of overfitting the model to high-dimensional features.

[0052] Table 2: Evaluation of classification effect of feature optimization based on wavelet layer combination in SVM model

[0053]

[0054]

[0055] Table 3 shows that the LR model achieved a prediction accuracy of 93.11% using the original data input, second only to the SVM. When modeling using wavelet features, the combined features of the fifth and seventh layers performed optimally, achieving a prediction accuracy of 91.88%. Both precision and recall were above 91%, demonstrating that low-dimensional linear models can also learn effective discriminative information from wavelet features. Furthermore, the modeling performance of the sixth layer was slightly weaker, suggesting that this layer may suffer from insufficient information or redundant interference.

[0056] Table 3: Classification effect evaluation table of feature optimization based on wavelet layer combination in LR model

[0057]

[0058]

[0059] As an ensemble method, the RF model demonstrates excellent generalization capabilities. Table 4 shows that the model combining the fifth and seventh layers achieved the highest accuracy (95.44%) on the prediction set, with a precision of 95.48%. Furthermore, the model combining the fifth, sixth, and seventh layers also achieved high performance, with a prediction accuracy of 94.11%, demonstrating that RF is more robust to high-dimensional inputs than other models. In the single-layer analysis, the fifth layer still performed best (93.56%), confirming its stable advantage across multiple models.

[0060] Table 4: Classification effect evaluation table of feature optimization based on wavelet layer combination in RF model

[0061]

[0062] Comparing the modeling results of different decomposition layers reveals that the seventh layer's characteristic coefficients have the highest correlation with the original spectrum, making it suitable for preserving full spectral information. The fifth layer performs well across all models, indicating that its detail scale is suitable for expressing structural changes caused by thermal damage. Fusion features (fifth + seventh layers, or fifth + sixth + seventh layers) significantly improve performance in both SVM and RF, making them particularly suitable for recognition tasks with complex nonlinear boundaries. Excessive feature combinations (e.g., fifth + sixth layers) lead to performance degradation in some models (e.g., PLS), suggesting the need to control dimensionality and information redundancy. Therefore, for hyperspectral wavelet decomposition feature modeling, a balance should be struck between information expression and feature redundancy, and appropriate feature layers should be selected based on model characteristics. Visualization results further validate the feasibility and accuracy of recognition systems based on wavelet decomposition and multi-model fusion in practical applications, demonstrating that hyperspectral technology, when combined with appropriate modeling methods, possesses excellent nondestructive testing capabilities. Figure 5 The visualization images of the inversion results are displayed, which intuitively reflect the performance of the model in sample identification and classification.

[0063] The present invention discloses a method for spectral identification and quality inversion of dry peas based on multi-layer wavelet decomposition and correlation analysis. This method addresses the problem that the internal quality changes of dry peas after heat damage treatment are difficult to accurately identify through traditional means. It uses visible-near-infrared hyperspectral imaging technology to collect multi-dimensional spectral information, and combines wavelet transform to perform multi-scale decomposition processing on the spectral data to extract detailed features at different scales. By calculating the correlation between the characteristics of each decomposition layer and the heat damage grade, the wavelet decomposition layer with the closest relationship to the target variable is screened out, and a variety of classification models such as PLS, SVM, LR, and RF are further constructed to achieve inversion prediction of the identification of samples with different heat damage grades. This method can effectively eliminate redundant bands and noise interference, strengthen the expression of quality-related information in spectral data, significantly improve recognition accuracy and model stability, and provide a new efficient, non-destructive, and intelligent detection solution for storage and transportation monitoring and quality grading of seed crops.

[0064] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for a person skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to replace some of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions claimed to be protected by the present invention.

Claims

1. A method for identifying heat damage in dry peas by integrating hyperspectral features with multi-layer wavelet decomposition, characterized in that: The following steps are involved: Step 1, selecting a dry pea sample; Step 2: The samples were divided into three groups, one of which was the control group, and the other two groups were subjected to heat damage treatment by heating, with different heat damage times; Step 3: Use a hyperspectral imaging system to collect high-dimensional spectral data of visible-near infrared light in the range of 400-1000 nm for three groups of samples; The db6 wavelet function was used to perform a seven-layer discrete wavelet decomposition of the original spectrum to extract detail information at different scales. A correlation analysis was then conducted between the detail coefficients of all decomposition layers and the heat loss level. Based on the strength of the correlation, three optimal wavelet layers were selected and their characteristic coefficients were extracted separately or in combination as model input variables. Step 4: Establish a dry pea recognition model based on different heat damage levels; In step 5, visible-near-infrared hyperspectral image data of the dry pea sample to be tested is collected, spectral acquisition and preprocessing are performed according to the parameters described in step 3, and the corresponding wavelet decomposition features are extracted; the extracted features are input into the constructed dry pea thermal damage identification model, and the thermal damage grade classification results are automatically output to achieve rapid and intelligent identification of the thermal damage status of the sample.

2. The method according to claim 1, wherein: In step 1, intact dry pea seeds without cracks, insect damage, or mildew are selected as samples.

3. The method according to claim 1, wherein: The two groups of samples subjected to heat damage treatment in step 2 were subjected to heat damage treatment for 20 days and 30 days respectively.

4. The method according to claim 1, wherein: The acquisition parameters of the visible-shortwave near-infrared spectrometer in step 3 are: the object distance between the lens and the sample is 300 mm, the spectral resolution is 1.43 nm, the scanning speed is 6.5 mm / s, the image size is 804×1097, and the window smoothing point number is 3.

5. The method according to claim 1, wherein: In step 3, the three best wavelet layers are selected as layers 5-7 based on the strength of the correlation.

6. The method according to claim 1, wherein: The dry pea recognition model described in step 4 is the least squares prediction model PLS, the support vector machine model SVM, the logistic regression model LR and the random forest model RF.

7. The method according to claim 6, characterized in that: The dry pea recognition model described in step 4 is a random forest model RF, which uses the 5th and 7th layers of wavelet, or the 5th layer + 6th layer + 7th layer fusion features.

8. The method according to claim 6, wherein: The dry pea recognition model described in step 4 is a support vector machine model SVM, which uses the fusion features of the 5th and 7th layers of wavelet.

9. The method according to claim 1, wherein: Step 5 outputs the classification results of the thermal damage level of the sample; then assign different colors to each pixel or area in the hyperspectral image according to the classification label, generate a thermal damage grade visualization map, and realize the identification of dry pea thermal damage and image visualization inversion analysis.

Images