In this study, the slightly degraded strain of Pleurotus geesteranus (Xiu 57-1) was subcultured into eighth and fifteenth generation, respectively. And the non-degraded strain (Xiu 57) was used for comparison. The strains were obtained from the Mycological Research Center of Fujian Agriculture and Forestry University. The strains were inoculated into the center of culture dishes containing potato dextrose agar (PDA) medium (Ye et al., 2022) and placed in a 25°C room for 10 days to activate them ( Figure 1 ). The growth rate and enzyme activity of the four types of strains were tested, and the degradation degrees of these four strains were then ranked. The activity of laccase, carboxymethylcellulase (CMC) and xylanases were determined using test kits (Suzhou Comin Biotechnology Co., Ltd., Suzhou, Fujian, China), respectively. The samples were each duplicated three times. The samples were labeled Xiu57-1, Xiu57-2, and Xiu57-3 from low to high degradation degrees, and the non-degraded strain was labeled Xiu57-0. In the experiment, a piece of flora with a diameter of 5 mm at the end of the activated mycelium was inoculated into the center of a new 90 mm culture dish containing PDA. Then, these culture dishes were placed upside down in a 25°C room, and images were collected after 5 days of culture. In total, we collected 200 samples, comprising 50 samples for each degradation degree.

Dimension reduction is necessary to improve the accuracy of hyperspectral images in classification applications. Depending on the feature in spectra and spatial dimensions, dimension reduction methods can be band selection and feature extraction, respectively.

SVM is a supervised machine learning method based on generalized linear classifiers. Its main goal is to find the best hyperplane for the classification of new data points (Sharma et al., 2017; Tiwari and Ojha, 2019). This is achieved by maximizing the margins between different classes and reducing the distance between the hyperplane focuses. To identify a suitable SVM, it is necessary to find a_i and b by minimization, and then optimize and express them as follows (Tang et al., 2020):

where k ( x i , x j ) is the kernel function; C is the regularization parameter representing the error classification tolerance parameter.

In the input space, the training data will be projected into a higher dimensional feature space when the linear separation in the kernel function becomes easier. SVM utilizes different kernel functions in order to find a hyperplane that could divide the data into groups more efficiently. It has good classification performance when used for the minimum training set (Pitchai et al., 2023). Therefore, the SVM model needs an appropriate kernel function in order to correctly evaluate the hyperplane and reduce classification errors. In this study, the kernel function of SVM is defined by means of weighted summation. The specific algorithm formula is as follows:

Wherein, μ is the weighting factor, 0 ≤ μ ≤ 1; K S is the spatial kernel function and K w is the spectral kernel function.

In order to evaluate the classification effect of the model on samples, this study selected the accuracy of detection classification as an indicator to qualitatively evaluate the model. The accuracy rate is the ratio of the number of correctly classified samples to the total number of samples. A higher accuracy rate indicates a better classifier. The specific calculation formula is as follows (Jin et al., 2018):

In the above formula,

The determined extracellular enzyme activities of the samples were shown in Table 1 . It shows that there were significant difference of laccase CMC and xylanase activity among different groups(p<0.05). Compared to the non-degraded strain(Xiu 57), the degraded strains exhibited decreased enzyme activities. Furthermore, as subculture generation increased, all three enzymes activities of the strains significantly declined. In fact, these activities could decrease by more than half compared to those of the non-degraded strain.

In this study, a total of 600 micro-hyperspectral images of undegraded and three differently degraded strains of Pleurotus geesteranus were collected, including 150 images each of 57-0, 57-1, 57-2, and 57-3. Next, 600 average spectral data points were obtained from the ROI region of mycelium, and the curves are displayed in Figure 4 . As shown in the figure, the four Pleurotus geesteranus strains had similar mycelial textures and hyperspectral curve trends. With increased degradation, the mycelia became more narrow and the transmittance slightly increased. Correspondingly, the mycelia of the undegraded strain of Pleurotus geesteranus were relatively stronger, and the transmittance curve was slightly lower than that of the degraded strain.

Due to the detection limitations of the camera, there was a certain amount of instrument noise at the end of the original spectral curve. The spectral information of 336 bands in the 400-1000 nm wavelength range was selected as effective information. Then, the preprocessing of spectral data was completed with SG, MSC, and SNV, respectively. As shown in Figure 5 , the spectral curve after SG filtering was relatively smooth as spectral noise was removed and the signal-to-noise ratio was improved. The phenomenon of baseline shift and offset was essentially eliminated after MSC treatment. In addition, SNV effectively eliminated the error caused by scattering.

The Kennard Stone (KS) algorithm was used to divide the dataset. The training set and the testing set had a ratio of 4:1. The spectral data preprocessed by four different methods were used as input variables to establish the SVM classification model. The classification accuracy of the model established by different pretreatment data was compared and tabulated in Table 2 .

Compared with the model established from the original data, the accuracy of the model established from the spectral data processed by the three preprocessing methods was improved in both the training and test sets. In contrast, the classification accuracy of the spectral data model processed by SNV had the most significant increases at 5.9% and 1.6%, respectively. Therefore, SNV was used in subsequent model building to preprocess spectral data.

We used the SPA algorithm to extract characteristic bands from the original spectral data that were processed by the SNV algorithm. In this process, the root mean square error (RMSE) of the model was calculated based on the characteristic bands preselected at each iteration. The high accuracy of the model was based on the lowest RMSE. As shown in Figure 6A , the RMSE is roughly negatively correlated with the number of selected bands; when the number of characteristic bands was 16, the RMSE achieved its lowest value of 0.6628.

Figure 6B depicts the distribution of these 16 characteristic spectral wavelengths across the full spectral band. The selected characteristic wavelength bands were mainly concentrated around 450nm, 700nm, and 950nm, and a few characteristic wavelengths were selected at the remaining peaks and valleys. The main reason for this was that the feature selection process of the SPA algorithm was an unsupervised process, which was only analyzed from the distribution of independent variables. In addition, the wavelengths at 401.0, 404.4, 411.1, 412.8, 417.8, 422.9, 433.0, 527.0, 649.6, and 733.7 nm demonstrated that degradation degrees of Pleurotus geesteranus strains could be to a certain extent differentiated in the visible range. The wavelengths at 949.7 nm and 962.9 nm were mainly related to the O-H stretching (second overtone) in water and the CH and CH₂ stretching (third overtone) in fat. The wavelengths near 980 nm (985.7, 993.3 nm) were assigned to the second overtone NH stretching in protein (Yu et al., 2018). The wavelength at 980 nm was due to O–H stretching the second and first overtones (Wu and Sun, 2013).

The grayscale image at the characteristic wavelengths of each micro-hyperspectral image was selected as the input. The pre-trained ResNet50 model was used to extract the features of each image. Since the pre-trained ResNet50 model used the ImageFolder function to read the image, the function copied the single-channel image automatically and converted it to a three-channel image at the same time. After converting each grayscale image into the RGB format, it reduced the image from the center to a size of 224 * 224, i.e., the input image size specified by the ResNet50 structure through the resize function. At the end of the process, the fully connected layer compressed the features extracted after convolution and pooling. It processed and integrated the output features of the previous layer into a 2048-dimensional feature vector through different weights. Then, the features were sent to the SVM classifier as the input. The model structure is shown in Figure 7 .

An SVM classification model was established for the spectral data based on characteristic bands after SNV. The radial basis (REF) function was chosen as the kernel function, and the criss-cross method was used to select the optimal parameters during modeling. As the classification results show in Figure 8 , after the spectral data underwent pretreatment and feature extraction, the modeling accuracy was 91.0% for the training set and 83.3% for the testing set.

The accuracy of the model based on the feature wavelengths extracted by the SNV-SPA algorithm in the testing set was 4.1% higher than that of the original-spectra-input model. This difference indicated that the SNV-SPA algorithm could reduce the complexity of the model while improving its prediction ability. Therefore, it has better performance.

The features extracted from the depth residual model ResNet50 were used as the input of the SVM classifier, and the training accuracy was 87.7% and 80.8%. We also compared the results with the SVM model using all the grayscale images covering 336 bands as the input. With the “RBF” kernel function, the classification model accuracy was 80.2% in the training set and 71.1% in the testing set. The prediction accuracy of the ResNet50-SVM model was improved by 9.7% over the non-SVM model.

After feature extraction, the training kernel matrix and the testing kernel matrix were calculated based on the feature vectors under the spectral dimension and image dimension, respectively. All the results of SVM training and classification for each combination of fusion coefficients were compared using a grid search. After that, linear fusion was performed for the kernel matrix of the two different dimensions under the optimal weight. The final training results are shown in Figure 9 . The accuracy rates of the training and testing sets were 96.0% and 90.8%, respectively.

Generally, the classification errors were mainly concentrated between Class 2 and Class 3. This might be because the spectral trends of these two classes were very similar and the shapes of the peaks and valleys had little difference, whereas, the peaks and valleys of Class 0 and Class 1 were quite different from them. In terms of image dimension, some mycelia of Class 2 and Class 3 were too narrow to be distinguished. By contrast, the mycelia of Class 0 and Class 1 could be detected easier because they were stronger.

To evaluate the model’s classification capability, each model was trained by SVM. The experimental results of the training and testing datasets are shown in Table 3 . For spectral dimension data and image dimension data, it was found that the testing set accuracies of the models established after feature extraction using SG-SPA and ResNet50 algorithms were 4.1% and 9.7% higher than those of the original models, respectively.