Apple leaf ring spot early diagnosis and chlorophyll content prediction method based on hyperspectral imaging and CNN-LSTM mixed model

Abstract

The invention discloses an apple leaf ring spot early diagnosis and chlorophyll content prediction method based on hyperspectral imaging and a CNN-LSTM mixed model. According to the method, hyperspectral images of apple leaves within the range of 400-1000 nm are collected through a hyperspectral imaging system, characteristic wave bands related to the disease progress and chlorophyll change are screened through a competitive adaptive reweighted sampling (CARS) and a continuous projection algorithm (SPA), and a convolutional neural network and long-short term memory network (CNN-LSTM) mixed model is input; early recognition of diseases and synchronous prediction of chlorophyll content are realized. And the chlorophyll prediction model is used for pixel-by-pixel prediction of the leaf spectral image, so that a spatial distribution visual image of the chlorophyll content of the apple leaf is realized, and lossless, rapid and quantitative evaluation of the health state of the apple leaf is realized.

Description

Technical Field

[0001] This invention relates to the field of non-destructive detection of plant diseases, and in particular to a method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM. Background Technology

[0002] Apples, as a widely cultivated and important economic fruit tree globally, directly impact the economic benefits for fruit growers and influence the overall development level of the fruit industry. Apple leaves, as the primary photosynthetic organs, directly reflect the tree's physiological state and photosynthetic capacity, serving as crucial indicators for evaluating tree health and yield potential. However, in recent years, apple leaf ring rot has frequently occurred in several major producing areas, becoming one of the significant diseases restricting the healthy growth of apple trees and the high yield and quality of fruit. This pathogenic fungus inhibits chlorophyll synthesis and reduces photosynthetic efficiency by infecting leaves, leading to early leaf drop and poor fruit development. Currently, the detection of apple leaf ring rot typically relies on manual visual investigation or laboratory physicochemical analysis: the former suffers from low efficiency, strong subjectivity, and susceptibility to environmental interference; the latter, while providing relatively accurate physiological and biochemical indicators, has a long detection cycle, complex operation, and is destructive, making it unsuitable for real-time field monitoring. Therefore, there is an urgent need to develop a non-destructive method that combines early disease diagnosis with physiological status assessment to achieve early and accurate identification of apple leaf ring rot and dynamic quantification of physiological status, providing reliable technical support for orchard disease prevention and control and precision management.

[0003] Hyperspectral imaging, as an emerging non-destructive testing method, can acquire spectral and spatial information of leaves in the visible to near-infrared range, reflecting changes in the physiological and biochemical characteristics of plant tissues. Existing research has shown that this technology can identify plant diseases and estimate physiological parameters, but it is difficult to achieve synergistic analysis of disease identification and chlorophyll content prediction. Therefore, there is an urgent need to propose a method for early diagnosis of apple leaf diseases and prediction of chlorophyll content based on hyperspectral imaging and deep learning, in order to achieve real-time, rapid, and non-destructive detection of leaf health status, providing technical support for intelligent monitoring and precise management of fruit tree diseases. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the above-mentioned background technology and propose a method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on hyperspectral imaging and CNN-LSTM hybrid model, thereby providing efficient and reliable technical support for intelligent monitoring and precise management of fruit tree diseases.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Step 1: Using a hyperspectral imaging system to obtain images of apple leaves at different post-inoculation days (400-1000 mm). The process involves: 1) Obtaining a hyperspectral image of the apple leaf at nm; 2) Performing radiometric correction on the hyperspectral image of the apple leaf and selecting a Region of Interest (ROI) to extract the average spectrum of all pixels within the ROI as the original spectral data of the sample; 3) Measuring the chlorophyll content of the apple leaf samples; 4) Selecting feature bands related to ring rot and chlorophyll content and dividing the sample set; 5) Using the spectral reflectance of the feature bands as input, constructing a ring rot diagnosis model using a hybrid model of convolutional neural network (CNN) and long short-term memory (LSTM), and determining the optimal model; 6) Using the spectral reflectance and chlorophyll content of the feature bands as input, constructing a chlorophyll content prediction model using a hybrid model of convolutional neural network (CNN) and long short-term memory (LSTM), and determining the optimal model; 7) Applying the optimal chlorophyll content prediction model to each pixel of the spectral image of the apple leaf at different infection times, generating a visual image of the spatial distribution of chlorophyll content based on each pixel, intuitively showing the changes in chlorophyll content of leaves under ring rot infection. The technical problem this invention aims to solve is to develop a comprehensive detection system that integrates disease identification and physiological parameter prediction. It proposes a method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model combining hyperspectral imaging and convolutional neural networks (CNN-LSTM).

[0006] Furthermore, in step 1, the apple tree variety is "Red Fuji". Day 0 is considered an uninoculated healthy leaf. Hyperspectral images of the apple leaves were acquired at 0, 2, 4, 6, and 8 days after inoculation. The radiometric correction formula is as follows: (1) In the formula, The corrected spectral image, The original spectral image of the sample. To calibrate the spectral image of the whiteboard, A spectral image with a dark background.

[0007] Furthermore, in step 2, the method for measuring the chlorophyll content of apple leaf samples is as follows: The relative chlorophyll content (SPAD) of leaves was measured using a SPAD-502 Plus chlorophyll meter. Measurements were taken once at three locations (upper, middle, and lower) along both sides of the midrib on each leaf, for a total of six measurements. The average of the six measurements was used as the reference SPAD value for that leaf. To ensure data consistency, all measurements were performed by the same operator.

[0008] Furthermore, in step 3, the continuous projection algorithm SPA and the competitive adaptive reweighted sampling algorithm CARS are used to select the feature bands.

[0009] Furthermore, in step 3, stratified sampling is used simultaneously to divide the sample set. In the construction of the diagnostic model, the samples are stratified according to different inoculation days, and then divided into a modeling set, a validation set, and a prediction set in a 6:2:2 ratio. In the construction of the chlorophyll content prediction model, continuous chlorophyll values ​​are stratified according to quartiles, and then divided into a modeling set, a validation set, and a prediction set in a 6:2:2 ratio.

[0010] Furthermore, in step 4, using the segmented dataset's feature band data and sample labels as input, a ring-shaped disease diagnosis model is constructed using a hybrid model of convolutional neural network (CNN) and long short-term memory (LSTM). This model first extracts features through two convolutional layers (Conv1 and Conv2), followed by batch normalization (BN) and ReLU activation after each layer to accelerate convergence and enhance nonlinear expression. Subsequently, dimensionality reduction is achieved through max pooling (Pool1 and Pool2), and the extracted feature sequences are input into the LSTM layer to capture long-range dependencies. Further, a multilayer perceptron (MLP) is used to nonlinearly map and combine the temporal features extracted by the LSTM, improving feature discrimination capability. Finally, a softmax classifier is used to predict different post-vaccination days.

[0011] Furthermore, in step 4, a diagnostic model for apple leaf ring rot is simultaneously constructed using Support Vector Machine (SVM) and Random Forest (RF) to compare the performance of the model constructed using CNN-LSTM. The modeling effect is evaluated using accuracy P, F1 score F1, and Kappa coefficient K.

[0012] Furthermore, in step 5, the chlorophyll content prediction model is constructed using the segmented dataset's feature band data and corresponding chlorophyll content values ​​as input, employing a hybrid model of convolutional neural network (CNN) and long short-term memory (LSTM). The front-end structure is consistent with the diagnostic model: first, spectral features are extracted through two convolutional layers (Conv1 and Conv2), followed by batch normalization (BNU) and ReLU activation after each layer to accelerate convergence and enhance nonlinear expression; then, downsampling and dimensionality reduction are performed using max pooling (Pool1 and Pool2). The resulting feature sequence is input into the LSTM layer to capture the long-range dependencies of the spectral sequence; further, a multilayer perceptron (MLP) is used to perform nonlinear mapping and combination of the temporal features. Unlike the diagnostic model, the regression model does not use a softmax classifier in the output layer; instead, it directly outputs continuous chlorophyll content prediction values ​​through the MLP to predict chlorophyll content.

[0013] Furthermore, in step 5, a prediction model for chlorophyll content in apple leaves infected with ring rot is constructed simultaneously using Partial Least Squares Regression (PLSR) and Backpropagation Neural Network (BPNN) to compare the performance of the model constructed using CNN-LSTM. The modeling effect is evaluated using the coefficient of determination (R²), root mean square error (RMSE), and residual prediction bias (RPD).

[0014] Furthermore, in step 6, representative apple leaf samples are selected, and the spectral reflectance values ​​of each pixel in the feature band image are extracted according to the optimal apple leaf chlorophyll content prediction model. The chlorophyll content of each pixel of the sample is calculated using the trained optimal prediction model, a grayscale image is generated, and the grayscale image is processed with pseudo-color to obtain a visualized color distribution map of chlorophyll content.

[0015] Furthermore, in step 6, as the infection progresses, the visualization smoothly transitions from orange to yellow-green; areas with high chlorophyll content are predominantly orange, covering the vast majority of the leaf pixels. Specifically: at day 0 post-inoculation, the leaves are predominantly red-orange, indicating a high chlorophyll content; from day 0 to 4 post-inoculation, localized green patches appear, reflecting a decrease in chlorophyll content; from day 0 to 4 post-inoculation, the overall color turns yellow-green, indicating a continued decrease in chlorophyll content. The decrease is particularly significant at the inoculation site and its adjacent areas.

[0016] The beneficial effects of this invention are as follows: By introducing a detection approach combining hyperspectral imaging and deep learning, this invention achieves early non-destructive diagnosis of apple leaf ring rot and simultaneous prediction of chlorophyll content. This invention can capture the spectral feature changes of leaves during the disease incubation period in the visible to near-infrared range (400-1000 nm). Through collaborative modeling using a hybrid CNN-LSTM model of convolutional neural networks and long short-term memory networks, it effectively extracts the spatial and temporal features of the spectrum, significantly improving the accuracy and stability of disease identification and physiological parameter prediction. This method can not only quantitatively assess leaf chlorophyll content but also visualize the spatial distribution of chlorophyll through pixel-by-pixel prediction, intuitively revealing the spatiotemporal variation patterns under disease infection. The entire detection process is non-destructive, rapid, highly accurate, and scalable, and can be applied to disease monitoring and health assessment of apple and other fruit tree leaves, providing reliable technical support for intelligent identification and precise management of orchard diseases. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0018] Figure 2 The image shows the spectral curves of apple leaves in this invention (where: (a) spectral curves of all samples, and (b) average spectral curves of samples at different infection stages).

[0019] Figure 3 This is a feature band selection diagram used to construct the diagnostic model in this invention (where: (a) the process of selecting feature bands based on the CARS algorithm, and (b) the distribution of feature bands optimized based on the SPA algorithm).

[0020] Figure 4 This is a structural diagram of the early diagnostic model for apple leaf ring spot disease in this invention.

[0021] Figure 5 The results of the early diagnosis model for apple leaf ring spot disease in this invention are as follows: (a) radar chart of the performance of each diagnosis model; (b) the change of accuracy and loss of the CARS–CNN-LSTM model with the number of training rounds; (c) confusion matrix of the modeling set of the CARS–CNN-LSTM model; (d) confusion matrix of the prediction set of the CARS–CNN-LSTM model.

[0022] Figure 6 The diagram shows the chlorophyll content analysis of apple leaves in this invention (wherein: (a) box plot of chlorophyll content at different post-inoculation days, and (b) distribution of chlorophyll content in each sample set).

[0023] Figure 7 This is a feature band selection diagram for chlorophyll content prediction in this invention (where: (a) the process of selecting feature bands based on the CARS algorithm, and (b) the feature band distribution optimized based on the SPA algorithm).

[0024] Figure 8 This is a structural diagram of the apple leaf chlorophyll content prediction model used in this invention for ring rot infection.

[0025] Figure 9 The results of the chlorophyll content prediction model in this invention are shown in the figure (where: (a) performance radar chart of chlorophyll prediction model; (b) changes of RMSEC and loss with each round during the training process of CARS–CNN-LSTM model; (c) scatter plot of measured and predicted values ​​of CARS–CNN-LSTM model model set; (d) scatter plot of measured and predicted values ​​of CARS–CNN-LSTM model prediction set).

[0026] Figure 10 This is a visualization of the chlorophyll content in apple leaves at different post-inoculation days in this invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0028] like Figure 1As shown, a method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM includes the following steps: Step 1: Use a hyperspectral imaging system to acquire hyperspectral images of apple leaves at 400-1000 nm at different post-inoculation days. After radiometric correction of the hyperspectral images of apple leaves, select the region of interest (ROI) and extract the average spectrum of all pixels within the ROI as the original spectral data of the sample. In step 1, the apple tree variety was "Red Fuji". Day 0 was considered as an uninoculated healthy leaf, and the days after inoculation were 0, 2, 4, 6 and 8 days respectively.

[0029] In step 1, the hyperspectral imaging system consists of a hyperspectral imaging sensor (GaiaField-V10E), an imaging lens, a halogen lamp light source, a standard white board, and a computer equipped with SoecView software. Its spectral range is 370-1015 nm, with 256 consecutive bands, a spectral resolution of 2.8 nm, an entrance slit width of 30 μm, a sampling interval of 2.3 nm, and a field of view of 22°. To capture a distortion-free, clear image, the camera lens is positioned 38 cm from the sample, and the camera exposure time is 11.8 ms.

[0030] In step 1, to reduce image inhomogeneity and the impact of dark current on the image, a standard white board with a reflectivity of 99% is used to perform radiometric correction on the image. The correction formula is as follows: (1) In the formula, The corrected spectral image, The original spectral image of the sample. To calibrate the spectral image of the whiteboard, A spectral image with a dark background.

[0031] In step 1, the main area of ​​the leaf is manually outlined as the Region of Interest (ROI) on each image, ensuring that the ROI covers representative parts of the leaf. The leaf is then accurately separated from the black background. Subsequently, the average reflectance of all pixels within the ROI is calculated as the spectral data for a single leaf. The obtained sample spectral curves are shown below. Figure 2 As shown.

[0032] Step 2: Measure the chlorophyll content of apple leaf samples; In step 2, after acquiring hyperspectral images of the apple leaves, the relative chlorophyll content (SPAD) of the leaves was measured using a SPAD-502 Plus chlorophyll meter. Measurements were taken once at three locations (upper, middle, and lower) along both sides of the midrib on each leaf, for a total of six measurements. The average of the six measurements was used as the reference SPAD value for that leaf. To ensure data consistency, all measurements were performed by the same operator. Chlorophyll content at each stage of infection is shown below. Figure 6 As shown in (a).

[0033] Step 3: Select characteristic bands related to ring spot disease and chlorophyll content, and divide the sample set; In step 3, CARS and SPA are used for feature band selection. CARS is based on the weights of the partial least squares (PLS) regression coefficients. By combining Monte Carlo sampling and a stepwise weight reduction strategy, it gradually eliminates bands that contribute less to the target variable, thus retaining feature bands highly correlated with disease classification or chlorophyll prediction. SPA, on the other hand, is based on the projected distance between variables and uses a stepwise selection strategy to effectively reduce multicollinearity among spectral variables, improve model interpretability and generalization ability, and is suitable for feature compression requirements in the modeling process. The feature bands selected for constructing the diagnostic model are shown in Table 1. Feature band selection is as follows: Figure 3 As shown in Table 2, the preferred feature bands used to construct the prediction model are as follows: The feature band selection is as follows... Figure 7 As shown.

[0034] Table 1 Table 2 In step 3, stratified random sampling is used to divide the sample set. In the diagnostic model construction, the samples are stratified according to different inoculation days, and then divided into a modeling set, validation set, and prediction set in a 6:2:2 ratio. In the chlorophyll content prediction model construction, continuous chlorophyll values ​​are stratified by quartiles, and then divided into a modeling set, validation set, and prediction set in a 6:2:2 ratio. The chlorophyll distribution of each dataset is shown below. Figure 6 As shown in (b).

[0035] Step 4: Using the spectral reflectance of the characteristic band as input, construct a ring-shaped disease diagnosis model using a hybrid model of convolutional neural network and long short-term memory network CNN-LSTM, and determine the optimal model with the best performance; In step 4, a diagnostic model for apple leaf ring rot was constructed using Support Vector Machine (SVM), Random Forest (RF), and CNN-LSTM. SVM achieves maximum margin separation of sample classes by constructing an optimal hyperplane in the feature space, exhibiting good generalization ability, especially suitable for classification tasks with small samples and high-dimensional data. It employs a Radial Basis Function (RBF) kernel function and optimizes the penalty parameter C and kernel parameter γ through cross-validation. RF effectively reduces the risk of overfitting by constructing multiple decision trees and using a voting strategy for result determination, demonstrating strong robustness and noise resistance. Based on the performance on the validation set, 500 decision trees were ultimately selected, with a maximum tree depth of 20, a minimum number of sample splits, and a minimum number of leaf node samples of 5. CNN-LSTM is a hybrid deep learning structure combining CNN and LSTM. In classification tasks, the CNN module first extracts features from the input, and the LSTM module then performs temporal modeling and information integration on the extracted feature sequences. Finally, the classification layer outputs the category, achieving accurate sample identification.

[0036] In step 4, during training of the proposed CNN-LSTM model, the data is first processed through two convolutional layers, Conv1 and Conv2, to extract features. Each layer is followed by Batch Normalization and ReLU activation to accelerate convergence and enhance nonlinear expression. Subsequently, dimensionality reduction is achieved through max pooling (Pool1 and Pool2), and the extracted feature sequences are input into the LSTM layer to capture long-range dependencies. Furthermore, a Multilayer Perceptron (MLP) is used to perform nonlinear mapping and combination of the temporal features extracted by the LSTM, improving feature discrimination capability. Finally, a Softmax classifier is used to predict different vaccination days. The CNN-LSTM structure is as follows: Figure 4 As shown.

[0037] In step 4, the accuracy P of the model set is calculated. C Modeling set F1-scoreF1 C Prediction set accuracy P P Prediction set F1-score F1 P and the prediction set Kappa coefficient K P The model performance was evaluated. The CARS–CNN-LSTM combination achieved the best results across all metrics: P on the modeling set. C =0.9422、F1 C =0.9404; P on the prediction set P =0.9321、F1 P =0.9304, K P =0.913. The training results of the early diagnostic model for apple leaf ring spot disease are as follows: Figure 5 As shown.

[0038] Step 5: Using the spectral reflectance and chlorophyll content of the characteristic bands as inputs, construct a chlorophyll content prediction model using a hybrid model of convolutional neural network and long short-term memory network CNN-LSTM, and determine the optimal model with the best performance. In step 5, a chlorophyll content prediction model is constructed by combining Partial Least Squares Regression (PLSR), Backpropagation Neural Network (BPNN), and a CNN-LSTM regression model. PLSR simultaneously reduces dimensionality and performs regression, projecting high-dimensional correlated independent variables into a low-dimensional latent variable space, maximizing the covariance between independent and dependent variables, and reducing the impact of multicollinearity, making it suitable for linear regression modeling of hyperspectral data. BPNN uses multi-layer nonlinear mapping to fit the complex relationship between input features and target variables, possessing strong nonlinear expression and adaptive learning capabilities, making it suitable for regression prediction of continuous variables such as spectral data and chlorophyll content. In regression tasks, CNN-LSTM combines the efficient extraction of local features by CNNs with the modeling of long-range dependencies by LSTMs, accurately characterizing the relationship between spectra and continuous chlorophyll content, thereby improving prediction accuracy and model generalization ability.

[0039] In step 5, the CNN-LSTM regression model takes the selected feature band data and corresponding chlorophyll content values ​​as input. The front-end structure is consistent with the classification model: first, spectral features are extracted through two convolutional layers, Conv1 and Conv2, followed by Batch Normalization and ReLU activation after each layer to accelerate convergence and enhance nonlinear expression; then, downsampling and dimensionality reduction are performed through max pooling (Pool1 and Pool2). The resulting feature sequence is input into the LSTM layer to capture the long-range dependencies of the spectral sequence; further, a multilayer perceptron (MLP) is used to perform nonlinear mapping and combination of the temporal features. Unlike the classification model, the regression model does not use a Softmax classifier in the output layer, but instead directly outputs continuous chlorophyll content prediction values ​​through the MLP to predict chlorophyll content. The structure of the CNN-LSTM regression model is as follows: Figure 8 As shown.

[0040] In step 5, the coefficients of determination are modeled. Coefficient of determination of the prediction set The model performance was evaluated using the root mean square error (RMSEC) of the modeling set, the root mean square error (RMSEP) of the prediction set, and the residual prediction bias (RPD). Among these, the CARS–CNN-LSTM combination achieved the best performance across all metrics. =0.9359, RMSEC=1.9269; Prediction set =0.8868, RMSEP=2.3931 and RPD=2.7186. The results of the predictive model for chlorophyll content in apple leaves infected with ring rot are as follows: Figure 9 As shown.

[0041] Step 6: Apply the optimal chlorophyll content prediction model to each pixel of the spectral images of apple leaves at different infection times, and generate a visualization image of the spatial distribution of chlorophyll content based on each pixel to intuitively show the changes in leaf chlorophyll content under ring rot infection.

[0042] In step 6, the constructed CARS–CNN-LSTM model was applied to each pixel of the sample spectral image, realizing pixel-by-pixel prediction and visualization of chlorophyll content, such as... Figure 10 As shown, the visualization smoothly transitions from orange to yellow-green as the infection progresses; areas with high chlorophyll content are predominantly orange, covering the vast majority of pixels on the leaf. Specifically: at day 0 post-inoculation, the leaves are predominantly red-orange, indicating high chlorophyll content; from day 0 to 4 post-inoculation, localized green patches appear, reflecting a decrease in chlorophyll content; from day 0 to 4 post-inoculation, the overall color turns yellow-green, indicating a continued decrease in chlorophyll content. The decrease is particularly significant at the inoculation site and its adjacent areas.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention.

[0044] All parts not covered in this invention are the same as or can be implemented using existing technologies.

Claims

1. A method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM, characterized in that, Includes the following steps: Step 1: Use a hyperspectral imaging system to acquire hyperspectral images of apple leaves at 400-1000 nm at different post-inoculation days. After radiometric correction of the hyperspectral images of apple leaves, select the region of interest (ROI) and extract the average spectrum of all pixels within the ROI as the original spectral data of the sample. Step 2: Measure the chlorophyll content of apple leaf samples; Step 3: Select the characteristic bands related to ring spot disease and chlorophyll content respectively, and divide the sample set; Step 4: Using the spectral reflectance of the characteristic band as input, construct a ring-shaped disease diagnosis model using a hybrid model of convolutional neural network and long short-term memory network CNN-LSTM, and determine the optimal model with the best performance; Step 5: Using the spectral reflectance and chlorophyll content of the characteristic bands as inputs, construct a chlorophyll content prediction model using a hybrid model of convolutional neural network and long short-term memory network CNN-LSTM, and determine the optimal model with the best performance. Step 6: Acquire hyperspectral images of the apple leaves to be tested, perform radiometric correction and ROI selection, extract the average spectral reflectance within the ROI as leaf spectral data, extract the spectral data corresponding to the characteristic bands related to the disease as input data, and use the optimal model of the ring rot disease diagnosis model to perform early diagnosis of ring rot disease on apple leaves; at the same time, extract the spectral data corresponding to the characteristic bands related to changes in chlorophyll content as input data, and use the optimal model of the chlorophyll content prediction model to predict chlorophyll content.

2. The method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM as described in claim 1, characterized in that, In step 1, The apple tree variety was "Red Fuji". Using day 0 as the baseline for uninoculated healthy leaves, hyperspectral images of the apple leaves were acquired at 0, 2, 4, 6, and 8 days post-inoculation. The radiometric correction formula is as follows: (1) In the formula, The corrected spectral image, The original spectral image of the sample. To calibrate the spectral image of the whiteboard, A spectral image with a dark background.

3. The method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM as described in claim 1, characterized in that, In step 1, the hyperspectral imaging system includes a hyperspectral imaging sensor, an imaging lens, a halogen lamp light source, a standard whiteboard, and a computer equipped with SpecView software.

4. The method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM as described in claim 1, characterized in that, In step 2, the instrument used to test the chlorophyll content of apple leaves is SPAD-502 Plus.

5. The method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM as described in claim 1, characterized in that, In step 3, the continuous projection algorithm SPA and the competitive adaptive reweighted sampling method CARS algorithm are used to select feature bands. In the construction of the diagnostic model, the samples are stratified according to different inoculation days, and then divided into modeling set, validation set and prediction set in a ratio of 6:2:

2. In the construction of the chlorophyll content prediction model, the continuous chlorophyll values ​​are stratified according to quartiles, and then divided into modeling set, validation set and prediction set in a ratio of 6:2:

2.

6. The method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM as described in claim 1, characterized in that, In step 4, the feature band data and sample labels of the divided dataset are used as input, and a wheel crack disease diagnosis model is constructed using a hybrid model of convolutional neural network and long short-term memory network CNN-LSTM. The model first extracts features through two layers of convolution Conv1 and Conv2, and each layer is followed by batch normalization and ReLU activation to accelerate convergence and enhance nonlinear expression. Subsequently, dimensionality reduction was achieved through max pooling (Pool1 and Pool2), and the extracted feature sequences were input into an LSTM layer to capture long-range dependencies. Furthermore, a multilayer perceptron (MLP) was used to perform nonlinear mapping and combination of the temporal features extracted by the LSTM to improve feature discrimination capability. Finally, a softmax classifier was used to predict different post-vaccination days.

7. The method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM as described in claim 1, characterized in that, In step 5, the chlorophyll content prediction model is constructed using the feature band data of the divided dataset and the corresponding chlorophyll content as input. The model is constructed using a hybrid model of convolutional neural network and long short-term memory network CNN-LSTM. First, spectral features are extracted through two layers of convolution Conv1 and Conv2. Batch normalization and ReLU activation are then applied after each layer to accelerate convergence and enhance nonlinear expression. Subsequently, downsampling and dimensionality reduction are performed using max pooling (Pool1 and Pool2). The resulting feature sequences are input into an LSTM layer to capture the long-range dependencies of the spectral sequences. The temporal features are then nonlinearly mapped and combined using a multilayer perceptron (MLP). The regression model directly outputs the predicted values ​​of continuous chlorophyll content in the output layer through the MLP, thus enabling the prediction of chlorophyll content.

8. The method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM as described in claim 1, characterized in that, In step 6, representative apple leaf samples are selected, and the spectral reflectance values ​​of each pixel in the feature band image are extracted according to the optimal apple leaf chlorophyll content prediction model. The chlorophyll content of each pixel in the sample is calculated using the trained optimal prediction model, a grayscale image is generated, and the grayscale image is processed with pseudo-color to obtain a visualized color distribution map of chlorophyll content.

9. The method for early diagnosis of apple leaf ring rot and prediction of chlorophyll content based on a hybrid model of hyperspectral imaging and CNN-LSTM as described in claim 1, characterized in that, It also includes applying the optimal chlorophyll content prediction model to each pixel of the spectral images of apple leaves at different infection times, normalizing the predicted chlorophyll content values ​​to the [0,1] interval according to the minimum and maximum values, and calling the pseudo-color mapping table to assign color values ​​according to the normalization results; low chlorophyll content areas correspond to blue, and high chlorophyll content areas correspond to red, thereby generating a visual pseudo-color distribution map of chlorophyll content.

Images