Soil salinity monitoring method based on multi-source satellite remote sensing and deep learning

Abstract

The invention discloses a soil salinity monitoring method based on multi-source satellite remote sensing and deep learning, and is applied to the technical field of soil salinity monitoring. Comprising the following steps: obtaining a vector graph and a satellite data set of a farmland, carrying out cloud removal screening and median synthesis on the vector graph of the farmland, carrying out resampling on image data of the satellite data set, selecting features in the data set through a feature selection algorithm, and comparing a traditional machine learning model with a deep learning model through an SHAP value. Evaluating the performance of different models, determining the key influence characteristics of the soil salinity, and in the model training process, updating the parameters of each model through a minimization logarithmic loss function, selecting a prediction regression model, and generating a farmland analysis visualization graph. According to the invention, the problems of untimely monitoring, time-consuming and labor-consuming process, complex data processing, unreliable monitoring precision and the like of the farmland in the prior art are solved, and key decision support is provided for soil salinization treatment of the farmland.

Description

Technical Field

[0001] The present invention relates to the technical field of soil salinity monitoring, and more specifically to a soil salinity monitoring method based on multi-source satellite remote sensing and deep learning. Background Art

[0002] Soil salinization has become a global environmental problem, posing a severe challenge to agricultural production and ecosystem health. The region is vast and has complex terrain. Many places face challenges of lack of data and information lag, which makes it difficult to obtain accurate and timely monitoring data on soil salinization. Therefore, there is an urgent need for an efficient and accurate large-area monitoring technology to assist in making timely soil management decisions.

[0003] In arid areas, the climate conditions of little rainfall and high evaporation lead to the accumulation of salt on the soil surface, forming saline-alkali land, which not only restricts the normal growth of crops such as cotton, but also reduces the availability of key nutrients in the soil, directly affecting the yield and quality of cotton. In addition, soil salinization increases the complexity and cost of land management, and traditional soil sampling methods are inefficient and difficult to provide timely and accurate support for agricultural production. With the development of precision agriculture, remote sensing technology has been widely used in agricultural soil salinization monitoring. UAV remote sensing has the advantages of high flexibility and high precision, and can quickly capture surface information, so it has been widely used in small-area fine monitoring tasks. However, due to the limitations of the endurance and flight speed of drones, their monitoring range is usually limited to a small area, and it is difficult to cover a large area of ​​farmland or conduct long-term continuous monitoring. In contrast, satellite remote sensing has the characteristics of wide coverage and high update frequency, and can perform periodic monitoring of large-scale farmland. Therefore, satellite remote sensing has received increasing attention in monitoring farmland changes, assessing crop growth status, and supporting agricultural decision-making. However, the information provided by a single data source of satellite remote sensing is insufficient to meet the needs of accurate and efficient monitoring in some cases. Specifically, although satellite remote sensing can provide a wide range of spectral information, its spatial resolution and temporal resolution are relatively low, which makes it difficult to provide accurate and reliable farmland management needs, especially in soil salinity monitoring tasks. Therefore, the limitations of a single satellite data source highlight the importance of multi-source remote sensing data fusion. Therefore, how to provide a soil salinity monitoring method based on multi-source satellite remote sensing and deep learning is an urgent problem that technicians in this field need to solve. Summary of the invention

[0004] In view of this, the present invention provides a soil salinity monitoring method based on multi-source satellite remote sensing and deep learning to improve the accuracy and efficiency of soil salinity monitoring in large areas of arid farmland.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A soil salinity monitoring method based on multi-source satellite remote sensing and deep learning includes the following steps:

[0007] S1, obtaining vector graphics of farmland, image data of Landsat 8 and Sentinel-2 satellite data sets, and performing cloud removal screening and median synthesis on the vector graphics of farmland;

[0008] S2, resample the image data of the Landsat 8 satellite dataset to ensure the same resolution of each band;

[0009] S3, select features in the data set through feature selection algorithm as input data for subsequent modeling;

[0010] S4. Use SHAP value to evaluate the contribution of features to soil salinity prediction and determine the key influencing features of soil salinity;

[0011] S5. During the model training process, the parameters of each model are updated by minimizing the logarithmic loss function, which evaluates the performance of the model by measuring the matching degree between the probability distribution of the model output and the true label;

[0012] S6. Select a predictive regression model to conduct a detailed analysis of farmland soil and generate a visualization of the farmland analysis.

[0013] Optionally, S1 is:

[0014] S11. Use the Google Earth Engine API to import the vector graphics of the farmland;

[0015] S12. Using the satellite datasets of Landsat 8 and Sentinel-2, a declouding algorithm was applied in Google Earth Engine to screen the cloud coverage percentage of each pixel, apply the cloud detection function cloudScore to detect the cloud amount, set a threshold T to determine the cloud amount, and if the result of the cloudScore detection is less than T, the pixel is considered to be a cloud, and the image frames with high cloud cover are removed;

[0016] S13. For the image set after cloud removal and filtering, a median synthesis method is applied to generate a synthetic image by calculating the median of each pixel in the time series.

[0017] Optionally, bicubic interpolation is used in S2 to implement resampling, and the interpolated pixel value I'(x',y') is:

[0018]

[0019] Where I'(x',y') is the interpolated pixel value, (x',y') is the coordinate of the target pixel, the original pixel coordinate is (x,y), and the position of (x',y') in the original image is between (i,j) and (i+1,j+1).

[0020] Optionally, S3 specifically includes: performing feature selection through continuous projection algorithm, competitive adaptive reweighted sampling, and uninformative variable removal algorithm, presetting a preliminary threshold for the number of features, calculating the importance of each feature in the model through a random forest algorithm, and selecting features whose feature importance scores are greater than the average importance scores of all features in the data set as input data for subsequent modeling.

[0021] Optionally, calculate the SHAP value of each feature in S4:

[0022]

[0023] In the formula, S represents the feature subset, N represents the full feature set, f(S) represents the output value of the model trained with feature subset S, f(S∪{i}) represents the output value of the model after feature i is added to feature subset S, and |S|! and |N|! represent the factorial number of feature subset S and full feature set N, respectively.

[0024] Optionally, S5 is specifically: The logarithmic loss function evaluates the performance of the model by measuring the matching degree between the probability distribution of the model output and the true label, and its formula is:

[0025]

[0026] In the formula, N represents the total number of samples, M represents the number of categories, and y ij represents the true label of the i-th sample in the j-th category; for the i-th sample, if its true category is the j-th category, then y ij =1, otherwise y ij =0; p ij represents the predicted probability that the i-th sample belongs to the j-th category; log(p ij ) is the logarithm of the predicted probability, which is used to measure the match between the model's prediction results and the true label.

[0027] Optionally, S6 is specifically:

[0028] S61, point set generation: based on the geographic information data of the farmland area, a series of geographic point sets are generated for the farmland soil at a resolution of 10 meters. The point set generation process is based on the grid division of the coordinate system, and the important characteristic attributes of the soil salinity corresponding to each point set coordinate are extracted;

[0029] S62. Selection of the best regression model: selecting the model with the best performance as the soil salinity prediction model, predicting the point set based on the soil salinity prediction model, and obtaining the predicted salinity value of each point;

[0030] S63, exporting the predicted salinity value of the soil salinity prediction model to an Excel file in a table format, wherein the data of each point includes the geographic coordinates and the corresponding predicted salinity value, and importing the Excel file into ArcGIS software, and performing spatial visualization of the data through the geographic information system. ArcGIS generates a soil salinity distribution map of the farmland area by mapping the predicted salinity value to the coordinate position of the point set;

[0031] S64. Soil classification and visualization mapping: Based on the predicted salinity values, the soil in the farmland area is classified according to the degree of soil salinization.

[0032] Optionally, the deep learning model includes a KAN network, which consists of a hidden layer and an output layer. Each layer processes the input data through learnable parameters and optimizes the learnable parameters through forward propagation and back propagation to minimize the loss function. The calculation formula of the hidden layer is:

[0033] hi(x)=k·A(w1·x+φ1)+B

[0034] Where x is the input feature, hi(x) is the output of the hidden layer, k, A, and B are learnable parameters; w1 and φ1 are the weight and bias parameters of the hidden layer, and A(·) is a learnable activation function;

[0035] The output layer is responsible for further converting the output of the hidden layer into the final prediction result. Its calculation formula is:

[0036] y(x)=k·A(w2·hi(x)+φ2)+B

[0037] Where y(x) is the final output of the model, w2 and φ2 are the weight and bias parameters of the output layer.

[0038] Optionally, the loss function of the KAN network during training is:

[0039]

[0040] In the formula, n is the number of samples, θ represents the parameters in the network, and y i is the true value, is the predicted value.

[0041] Through the above technical solutions, it can be seen that compared with the prior art, the present invention provides a soil salinity monitoring method based on multi-source satellite remote sensing and deep learning, which has the following beneficial effects: the present invention obtains and pre-processes large-area vector image data through Google Earth Engine, and simultaneously obtains the salinity information of farmland soil sampling points, and uses feature selection algorithms to create different feature data sets on spectra and environmental variables to ensure the representativeness and diversity of the data. Then, based on the comparison of machine learning and deep learning models, their performance in soil salinity prediction tasks is evaluated, and the key characteristic variables affecting soil salinity are identified. Finally, the model with the best performance in the prediction task is selected for soil The present invention not only significantly improves the application potential of remote sensing technology in farmland soil salinity monitoring, but also ensures the accuracy and effectiveness of the model in salt monitoring. By integrating multi-source satellite data and advanced machine learning algorithms, it can efficiently process soil data in complex environments and maintain high adaptability and robustness in different climatic conditions, geographical locations and farmland management practices, especially in farmland areas with more serious salinization. It can provide timely warnings to help farmers take intervention measures in advance, thereby improving the sustainability and economic benefits of agricultural production. At the same time, the present invention is scalable and can be applied to other types of agricultural monitoring tasks, such as crop health assessment, pest and disease monitoring, etc., to further promote the development of smart agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0043] Figure 1 It is a flow chart of the soil salinity monitoring method of the present invention;

[0044] Figure 2 Schematic diagram of spectral reflectance of Landsat 8 multi-source satellite image in an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of the spectral reflectance of the Sentinel-2 multi-source satellite image in an embodiment of the present invention;

[0046] Figure 4 is a structural diagram of an interpretable neural network KAN in an embodiment of the present invention;

[0047] Figure 5is an inversion diagram of the degree of salinization of arid farmland in an embodiment of the present invention;

[0048] Figure 6 Schematic diagram of model reliability based on SHAP and logarithmic loss in an embodiment of the present invention;

[0049] Figure 7 A schematic diagram of model feature contribution based on SHAP and logarithmic loss in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] The embodiment of the present invention discloses a soil salinity monitoring method based on multi-source satellite remote sensing and deep learning, such as Figure 1 As shown, the following steps are included:

[0052] S1. Obtain the vector graphics of farmland and the image data of Landsat 8 and Sentinel-2 satellite datasets, perform cloud removal screening and median synthesis on the vector graphics of farmland; the spectral reflectance of Landsat 8 and Sentinel-2 satellite datasets is as follows: Figure 2 and Figure 3 As shown;

[0053] S2, resample the image data of the Landsat 8 satellite dataset to ensure the same resolution of each band;

[0054] S3, select features in the data set through feature selection algorithm as input data for subsequent modeling;

[0055] S4. Use SHAP value to evaluate the contribution of features to soil salinity prediction and determine the key influencing features of soil salinity;

[0056] S5. During the model training process, the parameters of each model are updated by minimizing the logarithmic loss function, which evaluates the performance of the model by measuring the matching degree between the probability distribution of the model output and the true label;

[0057] S6. Select a predictive regression model to conduct a detailed analysis of farmland soil and generate a visualization of the farmland analysis.

[0058] Furthermore, S1 is specifically:

[0059] S11. Use Google EarthEngine API to import vector graphics of farmland;

[0060] In the embodiment of the present invention, S11 can be implemented by importing an existing Geographic Information System (GIS) file (Shapefile format), assuming that the vector layer is V, which includes the geographic coordinates and boundaries of each farmland block;

[0061] S12. Using the satellite datasets of Landsat 8 and Sentinel-2, a cloud removal algorithm was applied in Google EarthEngine to screen the cloud coverage percentage of each pixel, and the cloud detection function cloudScore was applied to detect the cloud amount. A threshold T was set to determine the cloud amount. If the result of the cloudScore detection was less than T, the pixel was considered to be a cloud, and image frames with high cloud cover were removed.

[0062] S13. For the image set after cloud removal screening, a median synthesis method is applied to generate a synthetic image by calculating the median of each pixel in the time series.

[0063] In the embodiment of the present invention, the formula for median synthesis is specifically:

[0064] I comp (x,y)=median({I t (x,y)∣t∈[t0,t1],cloudScore(I t (x,y) <T})

[0065] In the formula, I comp For synthetic images, I t is a single image at time t, x and y are pixel coordinates, cloudScore is the cloud detection function, and T is the cloud amount judgment threshold.

[0066] Furthermore, bicubic interpolation is used in S2 to achieve resampling. Bicubic interpolation (cubic convolution interpolation) has a good smooth interpolation effect. Bicubic interpolation usually involves a 4x4 neighborhood pixel matrix and uses a cubic polynomial to calculate the value of the target pixel; the interpolated pixel value I'(x',y') is:

[0067]

[0068] Where I'(x',y') is the interpolated pixel value, (x',y') is the coordinate of the target pixel, the original pixel coordinate is (x,y), and the position of (x',y') in the original image is between (i,j) and (i+1,j+1).

[0069] Furthermore, S3 is specifically as follows: feature selection is performed through the successive projection algorithm (SPA), competitive adaptive reweighted sampling (CARS), and uninformative variable removal algorithm (UVE), a preliminary threshold for the number of features is preset, the importance of each feature in the model is calculated through the random forest algorithm (RF), and features whose feature importance scores are greater than the average importance scores of all features in the data set are selected as input data for subsequent modeling.

[0070] In the embodiment of the present invention, the calculation formula of the feature importance average FeatureImportanceMean is:

[0071]

[0072] In the formula, n is the total number of all features, FeatureImportance i is the importance score of the i-th feature.

[0073] Furthermore, the SHAP value of each feature is calculated in S4:

[0074]

[0075] In the formula, S represents the feature subset, N represents the full feature set, f(S) represents the output value of the model trained with feature subset S, f(S∪{i}) represents the output value of the model after feature i is added to feature subset S, and |S|! and |N|! represent the factorial number of feature subset S and full feature set N, respectively.

[0076] In the embodiment of the present invention, in order to determine the key influencing features of soil salinity and evaluate the performance of machine learning in soil salinity prediction regression, traditional machine learning methods (support vector machine, random forest, extremely random tree) and deep learning methods (neural network, residual network, multi-layer perceptron) were selected for comparison. In order to achieve visual comparison and select features that significantly affect soil salinity, the SHAP value (SHapleyAdditive exPlanations) method was used. This method can explain the contribution of each feature to the prediction result. The SHAP method is an explanatory model tool based on game theory, which can assign a "contribution value" to each feature to quantify the marginal contribution of the feature to the model output. The feature contribution is as follows: Figure 7As shown in the figure. The SHAP value explains the impact of each feature on the prediction result by calculating the marginal impact of the feature on the model prediction result when the feature is added or removed from the model. Compared with other feature importance analysis methods, SHAP has the advantages of consistency and local accuracy, that is, it can ensure that the overall feature contribution of the model and the contribution of each feature are relatively consistent and explainable.

[0077] Furthermore, S5 is specifically: The logarithmic loss function evaluates the performance of the model by measuring the matching degree between the probability distribution of the model output and the true label, and its formula is:

[0078]

[0079] In the formula, N represents the total number of samples, M represents the number of categories, and y ij represents the true label of the i-th sample in the j-th category; for the i-th sample, if its true category is the j-th category, then y ij =1, otherwise y ij =0; p ij represents the predicted probability that the i-th sample belongs to the j-th category; log(p ij ) is the logarithm of the predicted probability, which is used to measure the match between the model's prediction results and the true label.

[0080] In the embodiment of the present invention, in order to evaluate the degree of soil salinization, the machine learning and deep learning models with good performance are selected in the prediction task to complete the soil salinization classification task. In order to evaluate the reliability of the classification model, the logarithmic loss function is added to the model training. The reliability evaluation is as follows: Figure 6 As shown. The logarithmic loss function (Log Loss), also known as the cross-entropy loss function (Cross-EntropyLoss), is one of the commonly used loss functions in classification problems. It evaluates the performance of the model by measuring the matching degree between the probability distribution of the model output and the true label, and is particularly suitable for multi-classification problems. During the model training process, the parameters of each model are updated by minimizing the logarithmic loss function to ensure that the model has high accuracy and stability in predicting the degree of soil salinization.

[0081] Furthermore, S6 is specifically:

[0082] S61, point set generation: based on the geographic information data of the farmland area, a series of geographic point sets are generated for the farmland soil at a resolution of 10 meters. The point set generation process is based on the grid division of the coordinate system, and the important characteristic attributes of the soil salinity corresponding to each point set coordinate are extracted;

[0083] In the embodiment of the present invention, the formula for generating the point set is:

[0084] P grid ={(xi ,y i )|i=1,2,…,N}

[0085] Where P grid Represents a series of geographic point sets generated at a resolution of 10 meters, (x i ,y i ) is the geographic coordinate of each point, and N represents the number of point sets generated;

[0086] S62. Selection of the best regression model: selecting the model with the best performance as the soil salinity prediction model, predicting the point set based on the soil salinity prediction model, and obtaining the predicted salinity value of each point;

[0087] In the embodiment of the present invention, the prediction model is expressed as:

[0088]

[0089] In the formula, is the predicted salinity value of the i-th point; f(·) represents the best trained soil salinity prediction model; θ is the parameter of the model; x i ,y i The input geographic coordinate data.

[0090] S63, exporting the predicted salinity value of the soil salinity prediction model to an Excel file in a table format, wherein the data of each point includes the geographic coordinates and the corresponding predicted salinity value, and importing the Excel file into ArcGIS software, and performing spatial visualization of the data through the geographic information system. ArcGIS generates a soil salinity distribution map of the farmland area by mapping the predicted salinity value to the coordinate position of the point set;

[0091] S64. Soil classification and visualization mapping: Based on the predicted salinity values, the soil in the farmland area is classified according to the degree of soil salinization.

[0092] In the embodiment of the present invention, the degree of soil salinization includes mild, moderate and severe, forming a clear soil type interval. The soil classification result is as follows: Figure 5 shown.

[0093] Furthermore, the deep learning model includes a KAN network, and the KAN network structure is as follows: Figure 4 As shown in the figure, the purpose is to improve the flexibility and expressiveness of the model while maintaining the interpretability of the model; the KAN network consists of hidden layers and output layers. Each layer processes the input data through learnable parameters and optimizes the learnable parameters through forward propagation and back propagation to minimize the loss function; the calculation formula of the hidden layer is:

[0094] hi(x)=k·A(w1·x+φ1)+B

[0095] In the formula, x is the input feature, hi(x) is the output of the hidden layer, k, A, and B are learnable parameters; w1 and φ1 are the weight and bias parameters of the hidden layer, and A(·) is the learnable activation function. Through the formula, the hidden layer can adaptively adjust the shape and weight parameters of the activation function to better express the nonlinear relationship of the input features.

[0096] The output layer is responsible for further converting the output of the hidden layer into the final prediction result. Its calculation formula is:

[0097] y(x)=k·A(w2·hi(x)+φ2)+B

[0098] In the formula, y(x) is the final output of the model, w2 and φ2 are the weight and bias parameters of the output layer. Similar to the hidden layer, the output layer also uses a learnable activation function to enhance the model's ability to express complex functions, so that the KAN network can not only learn complex patterns in the input data more flexibly, but also ensure that the model has high interpretability and flexibility by using a learnable activation function.

[0099] Furthermore, the loss function of the KAN network during training is:

[0100]

[0101] In the formula, n is the number of samples, θ represents the parameters in the network, and y i is the true value, is the predicted value.

[0102] The design of the KAN network is based on the Kolmogorov-Arnold representation theorem, which states that a multivariate continuous function can be represented by a finite number of single-variable continuous functions and binary addition operations. Therefore, the KAN network introduces a learnable activation function to replace the fixed activation function in the traditional neural network, processing input data in a more flexible way and enhancing the model's ability to learn complex relationships.

[0103] In the present embodiment, the model performance is evaluated to evaluate the prediction review and classification tasks. In the soil salinity prediction model, the coefficient of determination (R 2 ) and root mean square error (RMSE) are used to quantify the degree of data fitting and evaluate the model error. In salinity classification, accuracy, precision, recall and F1 score are used to evaluate the performance of distinguishing different salinity areas.

[0104]

[0105] Where TP, FP and FN represent true positive, false positive and false negative respectively. Precision is based on the prediction results of the model, and measures how many samples of a certain class (positive class) label predicted by the model as the degree of salinization are actually consistent with the class label. It reflects the accuracy of the positive class prediction, that is, the proportion of positive predictions. Recall is a statistical result from the real salt data. It is the number of real salt classifications found by the model in the total salt classification samples, that is, the proportion of full findings. The F1 score is a method to balance the detection model, balancing the accuracy and recall of the model. Accuracy is used to measure the overall performance of the classification model. It represents the ratio of the number of samples correctly classified by the model to the total number of samples. In the soil salt classification task, it can help understand the performance of the model in all categories.

[0106] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the embodiments can be referred to each other. The above description of the disclosed embodiments enables professionals and technicians in this field to implement or use the present invention. Various modifications to these embodiments will be obvious to professionals and technicians in this field, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown in this article, but will comply with the widest range consistent with the principles and novel features disclosed herein.

Claims

1. A soil salinity monitoring method based on multi-source satellite remote sensing and deep learning, characterized in that: The following steps are involved: S1, obtaining vector graphics of farmland, image data of Landsat 8 and Sentinel-2 satellite data sets, and performing cloud removal screening and median synthesis on the vector graphics of farmland; S2, resample the image data of the Landsat 8 satellite dataset to ensure the same resolution of each band; S3, select features in the data set through feature selection algorithm as input data for subsequent modeling; S4. Use SHAP value to evaluate the contribution of features to soil salinity prediction and determine the key influencing features of soil salinity; S5. During the model training process, the parameters of each model are updated by minimizing the logarithmic loss function, which evaluates the performance of the model by measuring the matching degree between the probability distribution of the model output and the true label; S6. Select a predictive regression model to conduct a detailed analysis of farmland soil and generate a visualization of the farmland analysis.

2. The soil salinity monitoring method based on multi-source satellite remote sensing and deep learning according to claim 1 is characterized in that: S1 is specifically: S11. Use the Google Earth Engine API to import the vector graphics of the farmland; S12. Using the satellite datasets of Landsat 8 and Sentinel-2, a declouding algorithm was applied in Google Earth Engine to screen the cloud coverage percentage of each pixel, apply the cloud detection function cloudScore to detect the cloud amount, set a threshold T to determine the cloud amount, and if the result of the cloudScore detection is less than T, the pixel is considered to be a cloud, and the image frames with high cloud cover are removed; S13. For the image set after cloud removal screening, a median synthesis method is applied to generate a synthetic image by calculating the median of each pixel in the time series.

3. The soil salinity monitoring method based on multi-source satellite remote sensing and deep learning according to claim 1 is characterized in that: Bicubic interpolation is used in S2 to achieve resampling, and the interpolated pixel value I'(x',y') is: Where I'(x',y') is the interpolated pixel value, (x',y') is the coordinate of the target pixel, the original pixel coordinate is (x,y), and the position of (x',y') in the original image is between (i,j) and (i+1,j+1).

4. The soil salinity monitoring method based on multi-source satellite remote sensing and deep learning according to claim 1 is characterized in that: S3 is specifically as follows: feature selection is performed through continuous projection algorithm, competitive adaptive reweighted sampling, and uninformative variable removal algorithm, a preliminary threshold for the number of features is preset, the importance of each feature in the model is calculated through the random forest algorithm, and features whose feature importance scores are greater than the average importance scores of all features in the data set are selected as input data for subsequent modeling.

5. The soil salinity monitoring method based on multi-source satellite remote sensing and deep learning according to claim 1 is characterized in that: Calculate the SHAP value of each feature in S4: In the formula, S represents the feature subset, N represents the full feature set, f(S) represents the output value of the model trained with feature subset S, f(S∪{i}) represents the output value of the model after feature i is added to feature subset S, and |S|! and |N|! represent the factorial number of feature subset S and full feature set N, respectively.

6. The soil salinity monitoring method based on multi-source satellite remote sensing and deep learning according to claim 1 is characterized in that: S5 is as follows: The logarithmic loss function evaluates the performance of the model by measuring the matching degree between the probability distribution of the model output and the true label. Its formula is: In the formula, N represents the total number of samples, M represents the number of categories, and y ij Represents the true label of the i-th sample in the j-th category; For the i-th sample, if its true category is the j-th category, then y ij =1, otherwise y ij =0; p ij represents the predicted probability that the i-th sample belongs to the j-th category; log(p ij ) is the logarithm of the predicted probability, which is used to measure the match between the model's prediction results and the true label.

7. The soil salinity monitoring method based on multi-source satellite remote sensing and deep learning according to claim 1 is characterized in that: S6 is specifically: S61, point set generation: based on the geographic information data of the farmland area, a series of geographic point sets are generated for the farmland soil at a resolution of 10 meters. The point set generation process is based on the grid division of the coordinate system, and the important characteristic attributes of the soil salinity corresponding to each point set coordinate are extracted; S62. Selection of the best regression model: selecting the model with the best performance as the soil salinity prediction model, predicting the point set based on the soil salinity prediction model, and obtaining the predicted salinity value of each point; S63, exporting the predicted salinity value of the soil salinity prediction model to an Excel file in a table format, wherein the data of each point includes the geographic coordinates and the corresponding predicted salinity value, and importing the Excel file into ArcGIS software, and performing spatial visualization of the data through the geographic information system. ArcGIS generates a soil salinity distribution map of the farmland area by mapping the predicted salinity value to the coordinate position of the point set; S64. Soil classification and visualization mapping: Based on the predicted salinity values, the soil in the farmland area is classified according to the degree of soil salinization.

8. The soil salinity monitoring method based on multi-source satellite remote sensing and deep learning according to claim 1 is characterized in that: The deep learning model includes the KAN network, which consists of a hidden layer and an output layer. Each layer processes the input data through learnable parameters and optimizes the learnable parameters through forward propagation and back propagation to minimize the loss function. The calculation formula of the hidden layer is: hi(x)=k·A(w1·x+φ1)+B Where x is the input feature, hi(x) is the output of the hidden layer, k, A, and B are learnable parameters; w1 and φ1 are the weight and bias parameters of the hidden layer, and A(·) is a learnable activation function; The output layer is responsible for further converting the output of the hidden layer into the final prediction result. Its calculation formula is: y(x)=k·A(w2·hi(x)+φ2)+B Where y(x) is the final output of the model, w2 and φ2 are the weight and bias parameters of the output layer.

9. The soil salinity monitoring method based on multi-source satellite remote sensing and deep learning according to claim 8, characterized in that: The loss function of the KAN network during training is: In the formula, n is the number of samples, θ represents the parameters in the network, and y i is the true value, is the predicted value.

Images