Near infrared spectrum-based leaf moisture content and chlorophyll detection method

Abstract

The embodiment of the invention provides a leaf moisture content and chlorophyll detection method based on a near infrared spectrum. The method comprises the following steps: acquiring original spectrum data of a sample leaf; preprocessing the original spectral data to obtain processed spectral data; carrying out key wave band screening on the processed spectral data by adopting a competitive self-adaptive reweighted sampling algorithm or a continuous projection algorithm to obtain target spectral data; and inputting the target spectral data into a pre-constructed neural network model, and detecting the water content and chlorophyll of the leaves to obtain a detection result. In the scheme, the CNN-KAN integrates the advantages of the convolutional neural network and the KAN network, so that the nonlinear fitting capability of the model is improved, the interpretability of the model is enhanced, the efficient extraction capability of fine features in the spectral data is shown, the detection cost is reduced, the detection efficiency is high, and field detection can be realized.

Description

Technical Field

[0001] This application relates to the field of detection methods, and in particular to a method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy. Background Technology

[0002] Seedlings are young plants cultivated from seeds or other propagation materials during tree reproduction, possessing a certain growth potential and vitality. The quality of seedlings directly affects the later growth rate, disease and pest resistance, physiological health, stress resistance, and final yield of the plant.

[0003] Currently, seedling testing typically employs chemical analysis methods. For example, the moisture content of leaves in *Betula sclerophylla* seedlings is determined using the drying and weighing method, which involves drying the sample to a constant weight and calculating the water loss to determine the moisture content. Chlorophyll is measured using spectrophotometry, which measures the absorbance of chlorophyll extract at a fixed wavelength and then calculates the chlorophyll content using a formula.

[0004] Traditional chemical methods are costly and time-consuming, requiring specific testing laboratories and failing to meet the requirements for rapid and non-destructive testing. Summary of the Invention

[0005] The purpose of this application is to provide a method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy, to solve the problems of existing detection methods, which have high detection costs, relatively long detection times, and require completion in a specific testing laboratory, thus failing to meet the application requirements of rapid and non-destructive detection. The specific technical solution is as follows: In a first aspect of this application, a method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy is provided, comprising: Obtain the raw spectral data of the sample blade; The original spectral data is preprocessed to obtain processed spectral data, wherein the preprocessing includes at least one of the following: multivariate scattering correction, standard normal variable transformation, SG smoothing, SG smoothing + multivariate scattering correction, and SG smoothing + standard normal variable transformation. The target spectral data is obtained by using either the Competitive Adaptive Reweighted Sampling (CARS) algorithm or the Continuous Projection (SPA) algorithm to screen key bands in the processed spectral data. The target spectral data is input into a pre-built neural network model to detect leaf water content and chlorophyll content, and the detection results are obtained.

[0006] Optionally, the neural network includes: a convolutional neural network, such as the Kolmogorov-Arnold network or a convolutional neural network.

[0007] Optionally, the convolutional neural network-Kolmogorov-Arnold network includes: Three convolutional pooling layers, a KAN network layer, and an output layer; The structure of the KAN network includes: The spectral reflectance data to be processed is used as the system input data. This input data contains data from n different spectral bands, namely X1, X2, ..., Xn. The system is set with (2n+1) independent processing channels q. In each channel, the following operations are performed: Using the first preset nonlinear processing function θq within the channel, preliminary transformations are performed on the n band data respectively. Each result after the preliminary transformation is multiplied by the preset weight coefficient λqp corresponding to each band within the channel. All weighted results are summed to generate an intermediate result for the channel. The intermediate results obtained from each channel are then input into the second preset nonlinear processing function Ψq within the channel for depth transformation to generate the final contribution value for the channel. Finally, the system sums the final contribution values ​​of all (2n+1) channels to obtain the final output result of the KAN network.

[0008] Optionally, the convolutional neural network includes: Three convolutional pooling layers, a fully connected layer, and an output layer; The convolutional kernel of the convolutional layer has a size of 1×5 and is used to extract local features; The pooling layer is a max pooling layer with a pooling window size of 1×3 to reduce the dimension of the feature map and reduce the computational cost. After being processed by three convolutional pooling layers, the size of the feature map is gradually reduced and finally passed to the fully connected layer to further integrate the features and perform non-linear mapping.

[0009] Optionally, the ReLU activation function is used in the fully connected layer to enhance the nonlinear expressive power of the model. The gradient of the loss function is calculated through the backpropagation algorithm, and the model parameters are updated based on the gradient information to optimize the learning process of the network. The output of the fully connected layer is used to predict the water content and chlorophyll content of the leaves.

[0010] In a second aspect of this application, a leaf water content and chlorophyll detection device based on near-infrared spectroscopy is also provided, comprising: The data acquisition module is used to acquire the raw spectral data of the sample blades; The preprocessing module is used to preprocess the original spectral data to obtain processed spectral data. The preprocessing includes at least one of the following: multivariate scattering correction, standard normal variable transformation, SG smoothing, SG smoothing + multivariate scattering correction (MSC), and SG smoothing + standard normal variable transformation. The band selection module is used to select key bands from the processed spectral data using either the Competitive Adaptive Reweighted Sampling (CARS) algorithm or the Continuous Projection (SPA) algorithm to obtain the target spectral data. The detection module is used to input the target spectral data into a pre-built neural network model to detect the leaf water content and chlorophyll content, and obtain the detection results.

[0011] Optionally, the neural network includes: a convolutional neural network, such as the Kolmogorov-Arnold network or a convolutional neural network.

[0012] Optionally, the convolutional neural network-Kolmogorov-Arnold network includes: Three convolutional pooling layers, a KAN network layer, and an output layer; The structure of the KAN network includes: The spectral reflectance data to be processed is used as the system input data. This input data contains data from n different spectral bands, namely X1, X2, ..., Xn. The system is set with (2n+1) independent processing channels q. In each channel, the following operations are performed: Using the first preset nonlinear processing function θq within the channel, preliminary transformations are performed on the n band data respectively. Each result after the preliminary transformation is multiplied by the preset weight coefficient λqp corresponding to each band within the channel. All weighted results are summed to generate an intermediate result for the channel. The intermediate results obtained from each channel are then input into the second preset nonlinear processing function Ψq within the channel for depth transformation to generate the final contribution value for the channel. Finally, the system sums the final contribution values ​​of all (2n+1) channels to obtain the final output result of the KAN network.

[0013] Optionally, the convolutional neural network includes: Three convolutional pooling layers, a fully connected layer, and an output layer; The kernel size of the convolutional layer is 1×5, which is used to extract local features; The pooling layer is a max pooling layer with a pooling window size of 1×3 to reduce the dimension of the feature map and reduce the computational cost. After three convolutional pooling layers, the size of the feature map is gradually reduced and finally passed to the fully connected layer to further integrate the features and perform non-linear mapping.

[0014] In a third aspect of this application, an electronic device is also provided, characterized in that it includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements any of the steps described above.

[0015] The beneficial effects of this application are: This application provides a method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy. The method involves acquiring raw spectral data of sample leaves; preprocessing the raw spectral data to obtain processed spectral data; using a competitive adaptive reweighted sampling algorithm (CARS) or a continuous projection algorithm (SPA) to select key bands in the processed spectral data to obtain target spectral data; and inputting the target spectral data into a pre-built neural network model to detect leaf water content and chlorophyll, thus obtaining the detection results. In this scheme, the CNN-KAN model is optimized by introducing CARS and SPA band selection methods to improve its modeling performance in predicting leaf water content and chlorophyll. CNN-KAN, by integrating the advantages of convolutional neural networks and KAN networks, not only improves the model's nonlinear fitting ability but also enhances its interpretability. In water content prediction, the CNN-KAN model, after SG+SNV preprocessing, achieves a prediction accuracy of 0.9017, far exceeding traditional methods such as RF and PLS, demonstrating better generalization performance and stability. In the chlorophyll content prediction task, the CNN-KAN model, after SG preprocessing, significantly outperformed other methods, demonstrating a highly efficient ability to extract subtle features from spectral data. This not only reduced detection costs but also increased detection efficiency, enabling on-site detection. A comparison of the results of the CARS and SPA band selection methods shows that the CARS method is superior in improving the prediction accuracy of the CNN-KAN model and reducing errors, making it suitable as a preferred band selection scheme under complex spectral feature conditions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0017] Figure 1 This is a flowchart of a method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy according to this application; Figure 2 The figures show the original spectral curves of this application and the spectral changes after processing with different preprocessing methods. In the figures, (a) is the original spectrum, (b) is the MSC spectrum, (c) is the SNV spectrum, (d) is the SG spectrum, (e) is the SG+MSC spectrum, and (f) is the SG+SNV spectrum. Figure 3 This is a diagram of the CNN model structure in this application; Figure 4 This is a schematic diagram of the KAN network in this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To address the problems in the existing technology, this application provides a method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy. This method solves the problems that existing detection methods have high detection costs, relatively long detection times, and need to be completed in a specific testing room, thus failing to meet the requirements for rapid and non-destructive detection.

[0020] It should be noted that the leaf water content and chlorophyll detection method based on near-infrared spectroscopy provided in this application embodiment can be applied to electronic devices. In practical applications, the electronic device can be a smartphone, tablet computer, laptop computer, etc., which is all reasonable.

[0021] The following section first introduces a method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy, provided in the embodiments of this application.

[0022] like Figure 1 As shown in the embodiments of this application, a method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy may include the following steps: S101, Obtain the raw spectral data of the sample blade.

[0023] For example, the sample blade in this embodiment can be obtained through on-site sampling. The original spectral data of the sample blade in this embodiment can be obtained using any spectroscopic instrument in the prior art, such as a portable near-infrared spectrometer.

[0024] For example, the equipment used to collect near-infrared spectral data can be a portable near-infrared spectrometer, model NIR-R210, manufactured by Shenzhen Spectrum Interconnect Technology Co., Ltd. To improve the representativeness of the measurement data, three measurement points were randomly selected for each leaf to ensure the diversity and comprehensiveness of the sampling. After the measurement was completed, the average of the three collected spectral data was taken as the final spectral data for that leaf. All operations throughout the experiment were performed under uniform experimental conditions, including ambient light intensity, ambient temperature, and humidity, to strictly control the impact of external interference factors on data acquisition.

[0025] S102, preprocess the original spectral data to obtain processed spectral data, wherein the preprocessing includes at least one of: multivariate scattering correction (MSC), standard normal variable transformation (SNV), SG smoothing, SG smoothing + multivariate scattering correction, and SG smoothing + standard normal variable transformation.

[0026] Raw spectral data may be affected by noise, scattering effects, and baseline drift, thus reducing the accuracy of spectral analysis. To improve modeling performance, this embodiment performs various preprocessing steps on the spectral data, including multivariate scattering correction (MSC), standard normal variable transformation (SNV), SG smoothing, and combined processing of SG+MSC and SG+SNV.

[0027] Figure 2 The paper presents the original spectral curves and the spectral changes after processing with different preprocessing methods. MSC processing significantly reduces scattering effects in the spectrum, improving spectral consistency between different samples; SNV further eliminates scattering and baseline drift, making the spectra more comparable across samples. SG smoothing effectively reduces noise while preserving key spectral features. Building upon this, SG+MSC combines the advantages of denoising, smoothing, and scattering correction, further enhancing data quality; SG+SNV, while strengthening the expressive power of spectral features, effectively suppresses interference, improving the stability and accuracy of modeling.

[0028] S103 uses either the Competitive Adaptive Reweighted Sampling (CARS) algorithm or the Continuous Projection (SPA) algorithm to screen key bands in the processed spectral data to obtain the target spectral data.

[0029] Since full-spectrum data contains a large number of bands, some of which may contribute little to the prediction of target parameters (chlorophyll and water content) or even introduce noise, reducing the model's prediction accuracy and stability. Therefore, this embodiment uses CARS and SPA to screen key bands to improve the model's prediction performance, thereby constructing a prediction model for chlorophyll and water content.

[0030] Table 1 shows that spectral preprocessing significantly impacts the performance of the Random Forest (RF) model in predicting leaf moisture content. Using the raw spectral data, the model's coefficient of determination (R²) is 0.7815, indicating that the RF model possesses a certain modeling capability without preprocessing and can effectively reveal the relationship between spectral features and leaf moisture content. The predictive performance of the model varies after applying different preprocessing methods. Among them, the model performs best after using the SG+SNV combined preprocessing method, with R² increasing to 0.8551 and RMSE and MAE decreasing to 0.0559 and 0.0397, respectively, significantly outperforming other processing methods. This demonstrates that SG+SNV preprocessing can effectively remove noise and enhance feature signals, thereby improving the model's accuracy in predicting moisture content.

[0031] In contrast, MSC and SNV preprocessing methods performed relatively poorly on this dataset, with R² values ​​of 0.6271 and 0.6893, respectively, and correspondingly high RMSE and MAE values. This indicates that these two preprocessing methods failed to effectively enhance model performance on this dataset, possibly due to insufficient preservation of key information or the introduction of interference, which affected the prediction results. SG and SG+MSC preprocessing improved model performance to some extent, achieving R² values ​​of 0.7947 and 0.7213, respectively, with relatively low prediction errors, but still not as good as SG+SNV. Comprehensive analysis shows that appropriate spectral preprocessing methods, especially the combination of SG+SNV, can significantly improve the performance of the RF model in leaf water content prediction.

[0032] Table 1. Performance of random forest water content prediction (full band)

[0033] Table 2 shows the performance of chlorophyll content prediction based on the RF model under different spectral preprocessing methods.

[0034] When using raw spectral data, the model's R... 2 The R² value was 0.8203, RMSE was 2.9112, and MAE was 2.1358, indicating that the RF model has good regression ability. Among all preprocessing methods, SG smoothing had the best effect, with R² = 0.8203. 2The R² value increased to 0.8244, while the RMSE and MAE decreased to 2.8778 and 2.1105, respectively. This indicates that the SG method effectively suppressed noise while preserving the characteristic information of the original spectrum, thus improving model performance. The SG+SNV combined preprocessing also performed well, with an R² of 0.8183, slightly lower than the original data and SG processing, but with RMSE and MAE of 2.9275 and 2.1338, respectively, showing good comprehensive predictive ability. In contrast, the SG+MSC processing performed relatively poorly, with an R² of 0.7897. Although there was some improvement, it did not reach the level of SG or the original spectrum. Standardization methods such as MSC and SNV performed poorly in this study, with R² values ​​of 0.7184 and 0.7181, respectively. Both RMSE and MAE were significantly higher than the original data, indicating that these methods failed to effectively improve model performance under the current data conditions and may even have caused information loss or noise enhancement, thus weakening the predictive ability of the RF model.

[0035] Table 2. Chlorophyll prediction performance of random forests (full band)

[0036] Table 3 shows the impact of the CARS and SPA band selection methods on the predictive performance of the random forest model under optimal preprocessing conditions. In leaf water content prediction, the CARS method performed best, with a coefficient of determination R0. 2 The R² value was 0.8511, RMSE was 0.0567, and MAE was 0.0402, demonstrating its significant advantages in removing redundant variables and extracting effective features, thereby significantly improving the model's accuracy and stability. In comparison, the SPA method's R² value was... 2 The R² value was 0.8442, slightly lower than CARS, but its RMSE and MAE were only 0.0580 and 0.0431 respectively, still demonstrating good predictive ability. In chlorophyll content prediction, the performance difference between the two methods was relatively small. The R² value of the CARS method... 2 The R value was 0.8147, RMSE was 2.9568, MAE was 2.1307, while the R value of the SPA method was... 2 The accuracy was slightly higher at 0.8220, with RMSE and MAE of 2.8975 and 2.1223 respectively, indicating that SPA had a slight advantage in fit for this task. However, considering both overall accuracy and model robustness, CARS still showed more stable performance. Overall, both CARS and SPA can effectively simplify the model input dimension and improve prediction performance, but CARS has a more significant advantage in water content prediction, while their performance is similar in the chlorophyll prediction task, showing some complementarity.

[0037] Table 3 Comparison of Random Forest CARS and SPA Band Selection

[0038] Table 4 shows the predictive performance of the PLS model for leaf water content under different spectral preprocessing methods. Using the original spectral data, the model's coefficient of determination (R²) was 0.8798, RMSE was 0.0509, and MAE was 0.0382, indicating that the PLS model possesses good modeling ability without preprocessing and can effectively reveal the relationship between spectral features and leaf water content. Among all preprocessing methods, the SG+SNV combination performed best, increasing R² to 0.8984 and reducing RMSE and MAE to 0.0468 and 0.0348, respectively. This indicates that the method achieved a good balance between enhancing effective features and suppressing noise, significantly improving the model's predictive performance. SG+MSC and SNV preprocessing also showed good modeling ability, with R² both at 0.8810, and their error indices were slightly lower than those of the original spectral data, indicating that they improved the model's stability and accuracy to some extent. The effects of SG and smoothing processing compared to MSC preprocessing were comparable to those of the original data, with R² values ​​of 0.8781 and 0.8728, respectively, indicating that their improvement on model performance was relatively limited. Overall, most preprocessing methods improved the performance of the PLSR model in water content prediction to varying degrees, with the SG+SNV combination showing the most significant improvement, demonstrating its advantages in spectral feature extraction and modeling.

[0039] Table 4. Performance of PLS ​​in predicting moisture content (full band)

[0040] Table 5 shows the impact of different spectral preprocessing methods on the PLS model's prediction of chlorophyll content. When using raw spectral data, the model's R-value is... 2 The R² was 0.8242, RMSE was 2.8794, and MAE was 2.0700, indicating that the PLS model has good regression performance without preprocessing and can effectively characterize the relationship between spectral features and chlorophyll content. Among all preprocessing methods, the SG+SNV combination showed the best results, increasing R² to 0.8327 and reducing RMSE and MAE to 2.8088 and 2.0161, respectively. This indicates that the method can effectively enhance spectral features related to chlorophyll while suppressing noise, significantly improving the model's prediction accuracy. The SG+MSC treatment also performed slightly better than the original data, with an R² of 0.8219 and a decrease in error indicators, showing that it improved the modeling effect to some extent. The R² of SG smoothing alone was 0.8125, which is slightly lower than the original spectrum, but the RMSE and MAE were 2.9736 and 2.1702, respectively, indicating that it has a certain effect on noise reduction, but its improvement on modeling is limited. The prediction effects of MSC and SNV were relatively average, with R² of 0.8242, RMSE of 2.8794, and MAE of 2.0700. 2Both scores were around 0.80, with RMSEs of 3.0357 and 3.0342 respectively, and MAEs exceeding 2.2, indicating that these two standardization methods failed to significantly improve model performance in the current task and may even have resulted in the loss of some useful information. Overall, the SG+SNV combination is the optimal preprocessing scheme, followed by SG+MSC; while MSC and SNV alone did not perform well. The results show that appropriate combined preprocessing methods have a positive effect on improving the performance of the PLS model in the chlorophyll content prediction task.

[0041] Table 5. PLS chlorophyll prediction performance (full band)

[0042] Table 6 shows the performance comparison of the PLS model in leaf water content and chlorophyll prediction tasks under optimal preprocessing conditions, using both CARS and SPA band selection methods. In water content prediction, the CARS method has an R² of 0.8463, an RMSE of 0.0576, and a MAE of 0.0405. The SPA method has a higher R². 2 The accuracy of the two methods was 0.8468, RMSE was 0.0575, and MAE was 0.0445. The difference in prediction accuracy between the two methods was minimal, indicating that both can effectively extract feature bands highly correlated with water content. However, the CARS method showed slightly better stability in error control. In the chlorophyll prediction task, the CARS method demonstrated a more significant advantage, with its R² value being 0.8468. 2 The R value was 0.8264, RMSE was 2.8619, MAE was 2.0949, while the R value of the SPA method was... 2 The accuracy scores were 0.8148, RMSE was 2.9559, and MAE was 2.1778, indicating that CARS more effectively preserved spectral information related to chlorophyll during band selection, improving the overall model fit and prediction accuracy. In this task context, CARS not only outperformed SPA in all accuracy metrics but also demonstrated stronger robustness.

[0043] Table 6 Comparison of CARS and SPA band selection in PLS models

[0044] S104, The target spectral data is input into a pre-built neural network model to detect leaf water content and chlorophyll, and the detection results are obtained.

[0045] For example, the neural network includes: a convolutional neural network - a Kolmogorov-Arnold network or one of the convolutional neural networks.

[0046] For example, in one implementation, the convolutional neural network includes: Three convolutional pooling layers, a fully connected layer, and an output layer; The kernel size of the convolutional layer is 1×5, which is used to extract local features; The pooling layer is a max pooling layer with a pooling window size of 1×3 to reduce the dimension of the feature map and reduce the computational cost. After three convolutional pooling layers, the size of the feature map is gradually reduced and finally passed to the fully connected layer to further integrate the features and perform non-linear mapping.

[0047] In the fully connected layer, the ReLU activation function is used to enhance the nonlinear expressive power of the model. The gradient of the loss function is calculated through the backpropagation algorithm, and the model parameters are updated based on the gradient information to optimize the learning process of the network. The output of the fully connected layer is used to predict the water content and chlorophyll content of the leaves.

[0048] Convolutional neural networks (CNNs) have demonstrated significant advantages in near-infrared spectral analysis. Their powerful feature extraction and nonlinear modeling capabilities make them an important tool for processing spectral data and performing efficient prediction tasks. This application constructs a CNN-based model, such as... Figure 3 As shown, this method enables non-destructive testing of chlorophyll and moisture content in the leaves of *Betula sclerophylla* seedlings.

[0049] The spectral data, after preprocessing, is used as input. Since the collected spectral data contains 228 bands, the input layer dimension is set to 228×1. To handle missing or incomplete feature data, a zero-padding strategy is employed to ensure consistency in the input dimension. The designed network structure includes three convolutional pooling layers, a fully connected layer, and an output layer. The kernel size of each convolutional layer is set to 1×5 for extracting local features; the pooling window size of the max-pooling layer is set to 1×3 to reduce the dimension of the feature map and decrease computational cost. After processing by three convolutional pooling layers, the size of the feature map is gradually reduced and finally passed to the fully connected layer for further feature integration and nonlinear mapping. In the fully connected layer, the ReLU activation function is used to enhance the model's nonlinear expressive power, thereby improving learning efficiency and prediction accuracy. Finally, the gradient of the loss function is calculated using the backpropagation algorithm, and the model parameters are updated based on the gradient information to optimize the network's learning process. Ultimately, the output of the fully connected layer is used to predict the leaf water content and chlorophyll content.

[0050] Table 7 shows the performance of the CNN model in predicting leaf water content under different spectral preprocessing conditions. CNN can effectively mine deep features in spectral data and achieves high prediction accuracy even when using the original spectrum. R 2The R² value was 0.8479, RMSE was 0.0573, and MAE was 0.0420, indicating good modeling capabilities. Significant differences in model performance were observed after introducing different preprocessing methods, with SNV showing the most significant effect. 2 The R² value increased to 0.8776, while RMSE and MAE decreased to 0.0514 and 0.0376, respectively, demonstrating the advantage of this method in improving the regression performance of CNN models. SG+MSC and SG alone also achieved performance comparable to the original spectra, with R² values ​​of 0.8484 and 0.8449, respectively, and relatively small changes in error metrics, indicating that this type of smoothing preprocessing method can improve model stability without destroying the original structure. The R² value of SG+SNV was 0.8376, slightly lower than the original data, suggesting a complex interaction between different preprocessing combinations and model performance. MSC alone performed relatively poorly, with R² decreasing to 0.8024 and the error increasing significantly, possibly because this method failed to effectively preserve spectral features related to water content.

[0051] In summary, CNN models are quite sensitive to preprocessing methods, with SNV preprocessing showing the best results. SG-type smoothing methods also improve model stability and prediction accuracy to some extent. Appropriate selection of spectral preprocessing strategies is crucial for fully leveraging the performance of deep learning models.

[0052] Table 7. CNN Water Content Prediction Performance (Full Band)

[0053] Table 8 shows the regression performance of the CNN model for chlorophyll content prediction under different spectral preprocessing methods. The results show that spectral processing has varying degrees of impact on model performance. Overall, the SNV preprocessing method performed best in this study, with an R² of 0.8621, significantly higher than the original spectrum (0.8415). Furthermore, the RMSE and MAE decreased to 2.5501 and 1.8678, respectively, indicating that this method has a strong ability to enhance effective features related to chlorophyll content in the spectrum, thus significantly improving the modeling accuracy of the CNN model. SG and its combinations (SG+SNV, SG+MSC) also improved model performance to some extent. Specifically, the R² after SG preprocessing was 0.8577, slightly higher than the original data, and the RMSE and MAE were 2.5908 and 1.8849, respectively, indicating that this method retained key feature information while removing spectral noise. The R² values ​​for SG+SNV and SG+MSC are 0.8526 and 0.8494, respectively, which are similar to the performance of SG processing and do not bring further improvement, indicating that the superposition effect of different preprocessing methods is limited.

[0054] Compared to the methods mentioned above, MSC processing performed poorly, with R² decreasing to 0.7852 and RMSE and MAE increasing to 3.1827 and 2.2890 respectively, indicating insufficient enhancement of spectral features and potentially weakening the model's sensitivity to the target variable. In conclusion, CNN models are highly sensitive to preprocessing methods, and a reasonable preprocessing strategy is crucial for improving model performance. Under the conditions of this study, SNV proved to be the optimal preprocessing method, followed by SG, while MSC was less effective and should be used with caution.

[0055] Table 8. CNN chlorophyll prediction performance (full band)

[0056] Table 9 shows the impact of CARS and SPA band selection methods on the performance of CNN models in predicting leaf water content and chlorophyll content under optimal preprocessing conditions. In the water content prediction task, the CARS method achieved R... 2 The R² value was 0.8708, slightly better than the SPA method, demonstrating its superior performance in modeling accuracy and error control. The chlorophyll prediction results showed a similar trend, with the CARS method exhibiting a lower R² value. 2 The CARS spectral density was 0.8571, with RMSE and MAE of 2.5959 and 1.8886 respectively. Compared to the SPA method, it still maintains a certain advantage in regression ability and prediction error. Although the RMSE of SPA in chlorophyll prediction was slightly lower than that of CARS, its MAE was higher, indicating that the overall deviation of the predicted values ​​was greater. Overall, the CARS method showed better comprehensive performance in both prediction tasks, indicating that it can more effectively extract key variables in the spectral feature screening process, enhancing the accuracy and stability of the model.

[0057] Table 9 Comparison of CARS and SPA band selection in CNN models

[0058] For example, in one implementation, the convolutional neural network-Kolmogorov-Arnold network includes: Three convolutional pooling layers, a KAN network layer, and an output layer; The structure of the KAN network includes: The spectral reflectance data to be processed is used as the system input data. This input data contains data from n different spectral bands, namely X1, X2, ..., Xn. The system is set with (2n+1) independent processing channels q. In each channel, the following operations are performed: Using the first preset nonlinear processing function θq within the channel (θq is not a fixed function, but an arbitrarily complex function that can be learned and adjusted through training data, usually a spline function, which is equivalent to automatically learning an optimal nonlinear encoding method for each input data), preliminary transformations are performed on the n band data respectively. Each result after the preliminary transformation is multiplied by the preset weight coefficient λqp corresponding to each band within the channel (λqp is a learnable parameter with a real value, the specific value of which is adaptively optimized by the network during training based on spectral data and target variables). All weighted results are summed to generate an intermediate result for the channel. The intermediate results obtained from each channel are then input into the second preset nonlinear processing function Ψq within the channel (Ψq is not a fixed function, but is implemented using a learnable parameterized function, such as a B-spline function) for depth transformation to generate the final contribution value for the channel. Finally, the system sums the final contribution values ​​of all (2n+1) channels to obtain the final output result of the KAN network.

[0059] CNNs are used to efficiently extract spatial features from input data, while KANs improve model performance and interpretability by replacing traditional fully connected layers. Specifically, fully connected layers are replaced with KAN network layers. The KAN network is a novel neural network structure based on the Kolmogorov-Arnold approximation theorem.

[0060] Figure 4 A schematic diagram of a 3-layer KAN network is shown. Unlike traditional neural networks that use fixed activation functions, KAN networks are built based on the Kolmogorov-Arnold representation theorem. Its core innovation lies in representing weight parameters as B-spline functions instead of traditional real values. These B-spline functions directly connect two neurons, replacing the linear weights in MLPs, allowing the network edges to have learnable activation functions. In KAN networks, neurons simply summarize the function outputs of each connection edge (simple summation), unlike traditional neurons that perform complex nonlinear transformations, thus efficiently approximating complex functions.

[0061] Table 10 shows the predictive performance of the CNN-KAN model for leaf water content under different spectral preprocessing conditions. Under the unprocessed raw spectral conditions, the model already demonstrated strong modeling ability, with R² of 0.8581, RMSE of 0.0553, and MAE of 0.0388. After processing with different preprocessing methods, the model performance was further improved. The SG+SNV combination achieved the best results, increasing R² to 0.9017 and reducing RMSE and MAE to 0.0460 and 0.0344, respectively, indicating that this combination excels in enhancing effective spectral features and improving model accuracy. The effect of using SG smoothing alone was also very close, with R² of 0.8999, RMSE and MAE of 0.0465 and 0.0351, respectively, showing its good effect in preserving key information and suppressing noise.

[0062] SNV and MSC processing also improved the model performance to some extent, with R² values ​​of 0.8665 and 0.8402, respectively, and the error index was slightly improved compared to the original data. The combination of SG and MSC (R²=0.8579) was similar to the original spectral results, indicating that this combination did not significantly enhance the model performance.

[0063] In summary, preprocessing methods have a significant impact on the prediction performance of the CNN-KAN model. The SG+SNV and SG methods perform best in the water content regression task and can effectively improve the model's feature extraction and generalization capabilities.

[0064] Table 10. CNN-KAN water content prediction performance (full band)

[0065] Table 11 shows the performance of the CNN-KAN model in predicting chlorophyll content under different preprocessing methods. Under the original spectral conditions without preprocessing, the model already exhibits good fitting ability, with R0... 2 The R² value was 0.8503, RMSE was 2.6034, and MAE was 1.8736. By introducing different preprocessing strategies, the model performance was improved to some extent. Among them, SG smoothing achieved the best results, with R² reaching 0.8648, and RMSE and MAE decreasing to 2.5253 and 1.8492 respectively, indicating that this method effectively suppressed noise while preserving key spectral features related to chlorophyll.

[0066] The SG+SNV combination performed relatively well, with an R² of 0.8588, RMSE of 2.5810, and MAE of 1.8551, indicating that this combination has certain advantages in enhancing feature representation. The SG+MSC combination yielded results close to the original spectrum, with an R² of 0.8417 and a slight decrease in error metrics, suggesting that this combination had limited improvement on model performance. In contrast, SNV and MSC treatments performed poorly individually. MSC treatment resulted in the worst performance, with R² decreasing to 0.7596, RMSE increasing to 3.3671, and MAE to 2.4711. SNV treatment, with an R² of 0.8063 and an RMSE of 3.0228, also failed to effectively improve model performance.

[0067] Overall, different preprocessing methods have significantly different impacts on model performance. The SG and SG+SNV methods performed best in the chlorophyll prediction task, demonstrating their effectiveness in enhancing features and improving prediction accuracy.

[0068] Table 11. Chlorophyll prediction performance of the CNN-KAN model with different preprocessing methods (full band).

[0069] Under optimal preprocessing conditions, two band selection methods, CARS and SPA, were further introduced to optimize the features of the CNN-KAN model, thereby improving its modeling performance in leaf water content and chlorophyll prediction tasks. Table 12 presents the comparison results of the two methods. In water content prediction, the CARS method achieved higher accuracy, R... 2 The R² value is 0.9183, RMSE is 0.0420, and MAE is 0.0311, demonstrating its strong capabilities in feature extraction and redundant information removal. In comparison, the SPA method performs slightly worse, with R² value... 2 The RSI was 0.9067, with RMSE and MAE of 0.0449 and 0.0335 respectively. While these figures were also high, they were slightly lower than CARS in terms of error control. In the chlorophyll prediction task, CARS also demonstrated superior performance, with RSI... 2 The R² value was 0.8713, RMSE was 2.4636, and MAE was 1.7736, which is better than the R² value of the SPA method overall. 2 =0.8479, RMSE=2.6783, and MAE=1.9314. CARS shows a significant advantage in both prediction metrics, reflecting its stronger stability and adaptability in constructing efficient feature subsets. The comparison shows that the CARS method performs better in improving the prediction accuracy of the CNN-KAN model and reducing errors, making it suitable as a preferred band selection scheme under complex spectral feature conditions.

[0070] Table 12 Comparison of band selection methods under optimal preprocessing for CNN-KAN models

[0071] In this application, the original spectral data of the sample leaves are obtained; the original spectral data is preprocessed to obtain processed spectral data; the competitive adaptive reweighted sampling algorithm (CARS) and the continuous projection algorithm (SPA) are used to screen key bands in the processed spectral data to obtain target spectral data; the target spectral data is input into a pre-constructed neural network model to detect leaf water content and chlorophyll content, and the detection results are obtained. In this scheme, the optimal processing method of the CNN-KAN model in water content prediction is SG+SNV+CARS, R 2 The R value reached 0.9183, and the optimal treatment method for chlorophyll prediction was SG+CARS, with R... 2 The R² value is 0.8713. Under the optimal preprocessing conditions of SG+SNV and SG, two band selection methods, CARS and SPA, are further introduced to optimize the features of the CNN-KAN model, thereby improving its modeling performance in leaf water content and chlorophyll prediction tasks. The comparison results of the two methods are presented. In water content prediction, the CARS method achieves higher accuracy, with R² = 0.8713. 2 The R² value is 0.9183, RMSE is 0.0420, and MAE is 0.0311, demonstrating its strong capabilities in feature extraction and redundant information removal. In comparison, the SPA method performs slightly worse, with R² value... 2 The RSI was 0.9067, with RMSE and MAE of 0.0449 and 0.0335 respectively. While these figures were also high, they were slightly lower than CARS in terms of error control. In the chlorophyll prediction task, CARS also demonstrated superior performance, with RSI... 2 The R² value was 0.8713, RMSE was 2.4636, and MAE was 1.7736, which is better than the R² value of the SPA method overall. 2 =0.8479, RMSE=2.6783, and MAE=1.9314. CARS shows a significant advantage in both prediction metrics, reflecting its stronger stability and adaptability in constructing efficient feature subsets. The comparison shows that the CARS method performs better in improving the prediction accuracy of the CNN-KAN model and reducing errors, making it suitable as a preferred band selection scheme under complex spectral feature conditions.

[0072] Compared to the embodiments of the above methods, this application provides a leaf water content and chlorophyll detection device based on near-infrared spectroscopy, which may include: The data acquisition module is used to acquire the raw spectral data of the sample blades; The preprocessing module is used to preprocess the original spectral data to obtain processed spectral data. The preprocessing includes at least one of the following: multivariate scattering correction, standard normal variable transformation, SG smoothing, SG smoothing + multivariate scattering correction (MSC), and SG smoothing + standard normal variable transformation. The band selection module is used to select key bands from the processed spectral data using the Competitive Adaptive Reweighted Sampling (CARS) algorithm and the Continuous Projection (SPA) algorithm to obtain the target spectral data. The detection module is used to input the target spectral data into a pre-built neural network model to detect the leaf water content and chlorophyll content, and obtain the detection results.

[0073] This application obtains the raw spectral data of sample leaves; preprocesses the raw spectral data to obtain processed spectral data; uses a competitive adaptive reweighted sampling algorithm (CARS) and a continuous projection algorithm (SPA) to screen key bands in the processed spectral data to obtain target spectral data; inputs the target spectral data into a pre-constructed neural network model to detect leaf water content and chlorophyll content, obtaining the detection results. In this scheme, the optimal processing method for the CNN-KAN model in water content prediction is SG+SNV+CARS, R 2 The R value reached 0.9183, and the optimal treatment method for chlorophyll prediction was SG+CARS, with R... 2 The R² value is 0.8713. Under the optimal preprocessing conditions of SG+SNV and SG, two band selection methods, CARS and SPA, are further introduced to optimize the features of the CNN-KAN model, thereby improving its modeling performance in leaf water content and chlorophyll prediction tasks. The comparison results of the two methods are presented. In water content prediction, the CARS method achieves higher accuracy, with R² = 0.8713. 2 The R² value is 0.9183, RMSE is 0.0420, and MAE is 0.0311, demonstrating its strong capabilities in feature extraction and redundant information removal. In comparison, the SPA method performs slightly worse, with R² value... 2 The RSI was 0.9067, with RMSE and MAE of 0.0449 and 0.0335 respectively. While these figures were also high, they were slightly lower than CARS in terms of error control. In the chlorophyll prediction task, CARS also demonstrated superior performance, with RSI... 2 The R² value was 0.8713, RMSE was 2.4636, and MAE was 1.7736, which is better than the R² value of the SPA method overall. 2=0.8479, RMSE=2.6783, and MAE=1.9314. CARS shows a significant advantage in both prediction metrics, reflecting its stronger stability and adaptability in constructing efficient feature subsets. The comparison shows that the CARS method performs better in improving the prediction accuracy of the CNN-KAN model and reducing errors, making it suitable as a preferred band selection scheme under complex spectral feature conditions.

[0074] Optionally, the neural network includes: a convolutional neural network, such as the Kolmogorov-Arnold network or a convolutional neural network.

[0075] Optionally, the convolutional neural network-Kolmogorov-Arnold network includes: Three convolutional pooling layers, a KAN network layer, and an output layer; The structure of the KAN network includes: The spectral reflectance data to be processed is used as the system input data. This input data contains data from n different spectral bands, namely X1, X2, ..., Xn. The system is set with (2n+1) independent processing channels q. In each channel, the following operations are performed: Using the first preset nonlinear processing function θq within the channel, preliminary transformations are performed on the n band data respectively. Each result after the preliminary transformation is multiplied by the preset weight coefficient λqp corresponding to each band within the channel. All weighted results are summed to generate an intermediate result for the channel. The intermediate results obtained from each channel are then input into the second preset nonlinear processing function Ψq within the channel for depth transformation to generate the final contribution value for the channel. Finally, the system sums the final contribution values ​​of all (2n+1) channels to obtain the final output result of the KAN network.

[0076] Optionally, the convolutional neural network includes: Three convolutional pooling layers, a fully connected layer, and an output layer; The kernel size of the convolutional layer is 1×5, which is used to extract local features; The pooling layer is a max pooling layer with a pooling window size of 1×3 to reduce the dimension of the feature map and reduce the computational cost. After three convolutional pooling layers, the size of the feature map is gradually reduced and finally passed to the fully connected layer to further integrate the features and perform non-linear mapping.

[0077] This application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus. Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the leaf water content and chlorophyll detection based on near-infrared spectroscopy provided in the embodiments of this application.

[0078] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0079] The communication interface is used for communication between the aforementioned terminal and other devices.

[0080] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0081] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0082] In another embodiment provided in this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, it implements the leaf water content and chlorophyll detection method based on near-infrared spectroscopy as described in any of the above embodiments.

[0083] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the near-infrared spectroscopy-based leaf water content and chlorophyll detection methods described in the above embodiments.

[0084] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0087] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A method for detecting leaf water content and chlorophyll based on near-infrared spectroscopy, characterized in that, include: Obtain the raw spectral data of the sample blade; The original spectral data is preprocessed to obtain processed spectral data, wherein the preprocessing includes at least one of the following: multivariate scattering correction, standard normal variable transformation, SG smoothing, SG smoothing + multivariate scattering correction, and SG smoothing + standard normal variable transformation. The target spectral data is obtained by using a competitive adaptive reweighted sampling algorithm or a continuous projection algorithm to screen key bands in the processed spectral data. The target spectral data is input into a pre-built neural network model to detect leaf water content and chlorophyll content, and the detection results are obtained.

2. The method according to claim 1, characterized in that, The neural network includes: a convolutional neural network, such as the Kolmogorov-Arnold network or one of the convolutional neural networks.

3. The method according to claim 2, characterized in that, The convolutional neural network-Kolmogorov-Arnold network includes: Three convolutional pooling layers, a KAN network layer, and an output layer; The structure of the KAN network includes: The spectral reflectance data to be processed is used as the system input data. This input data contains data from n different spectral bands, namely X1, X2, ..., Xn. The system is set with (2n+1) independent processing channels q. In each channel, the following operations are performed: Using the first preset nonlinear processing function θq within the channel, preliminary transformations are performed on the n band data respectively. Each result after the preliminary transformation is multiplied by the preset weight coefficient λqp corresponding to each band within the channel. All weighted results are summed to generate an intermediate result for the channel. The intermediate results obtained from each channel are then input into the second preset nonlinear processing function Ψq within the channel for depth transformation to generate the final contribution value for the channel. Finally, the system sums the final contribution values ​​of all (2n+1) channels to obtain the final output result of the KAN network.

4. The method according to claim 2, characterized in that, The convolutional neural network includes: Three convolutional pooling layers, a fully connected layer, and an output layer; The kernel size of the convolutional layer is 1×5, which is used to extract local features; The pooling layer is a max pooling layer with a pooling window size of 1×3 to reduce the dimension of the feature map and reduce the computational cost. After three convolutional pooling layers, the size of the feature map is gradually reduced and finally passed to the fully connected layer to further integrate the features and perform non-linear mapping.

5. The method according to claim 4, characterized in that, In the fully connected layer, the ReLU activation function is used to enhance the nonlinear expressive power of the model. The gradient of the loss function is calculated through the backpropagation algorithm, and the model parameters are updated based on the gradient information to optimize the learning process of the network. The output of the fully connected layer is used to predict the water content and chlorophyll content of the leaves.

6. A device for detecting leaf water content and chlorophyll based on near-infrared spectroscopy, characterized in that, include: The data acquisition module is used to acquire the raw spectral data of the sample blade; The preprocessing module is used to preprocess the original spectral data to obtain processed spectral data. The preprocessing includes at least one of the following: multivariate scattering correction, standard normal variable transformation, SG smoothing, SG smoothing + multivariate scattering correction (MSC), and SG smoothing + standard normal variable transformation. The band selection module is used to select key bands from the processed spectral data using a competitive adaptive reweighted sampling algorithm or a continuous projection algorithm to obtain target spectral data. The detection module is used to input the target spectral data into a pre-built neural network model to detect the leaf water content and chlorophyll content, and obtain the detection results.

7. The apparatus according to claim 6, characterized in that, The neural network includes: a convolutional neural network, such as the Kolmogorov-Arnold network or one of the convolutional neural networks.

8. The apparatus according to claim 7, characterized in that, The convolutional neural network-Kolmogorov-Arnold network includes: Three convolutional pooling layers, a KAN network layer, and an output layer; The structure of the KAN network includes: The spectral reflectance data to be processed is used as the system input data. This input data contains data from n different spectral bands, namely X1, X2, ..., Xn. The system is set with (2n+1) independent processing channels q. In each channel, the following operations are performed: Using the first preset nonlinear processing function θq within the channel, preliminary transformations are performed on the n band data respectively. Each result after the preliminary transformation is multiplied by the preset weight coefficient λqp corresponding to each band within the channel. All weighted results are summed to generate an intermediate result for the channel. The intermediate results obtained from each channel are then input into the second preset nonlinear processing function Ψq within the channel for depth transformation to generate the final contribution value for the channel. Finally, the system sums the final contribution values ​​of all (2n+1) channels to obtain the final output result of the KAN network.

9. The apparatus according to claim 7, characterized in that, The convolutional neural network includes: Three convolutional pooling layers, a fully connected layer, and an output layer; The kernel size of the convolutional layer is 1×5, which is used to extract local features; The pooling layer is a max pooling layer with a pooling window size of 1×3 to reduce the dimension of the feature map and reduce the computational cost. After three convolutional pooling layers, the size of the feature map is gradually reduced and finally passed to the fully connected layer to further integrate the features and perform non-linear mapping.

10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-5.

Images