Satellite hyperspectral coastal saline-alkali soil salinity inversion method based on machine learning model

Abstract

The invention discloses a satellite hyperspectral coastal saline-alkali soil salinity inversion method based on a machine learning model, and the method comprises the steps: S1, collecting an original multispectral image and ground actual measurement data, and obtaining standardized initial feature space data through the radiometric calibration correction preprocessing operation; s2, performing fractional order differential spectrum enhancement on the initial feature space data to obtain an enhanced high-dimensional feature set, and performing importance adaptive optimization on the enhanced high-dimensional feature set by adopting a Boruta feature selection algorithm to obtain an optimal high-dimensional feature subset; s3, based on the optimal high-dimensional feature subset, an initial inversion model is obtained by integrating Optuna hyper-parameter automatic optimization, and based on the initial inversion model, a salinity inversion model is obtained through feature contribution degree analysis optimization and embedding of an SHAP interpreter; and S4, performing salinity prediction on the global remote sensing image based on the salinity inversion model to realize sea saline-alkali land salinity inversion.

Description

Technical Field

[0001] This invention relates to the field of next-generation information technology, and in particular to a satellite hyperspectral method for salinity inversion of coastal saline-alkali land based on machine learning models. Background Technology

[0002] Remote sensing technology provides an efficient means for monitoring soil salinity at the regional scale, especially hyperspectral remote sensing, which can capture the fine spectral response of soil salinity in the visible-near-infrared band. Traditional methods mainly rely on linear statistical models or simple vegetation indices for salinity estimation, but due to the complexity of coastal saline-alkali land environments (such as tidal disturbances, vegetation cover, and soil moisture variability), the inversion accuracy is generally insufficient. In recent years, machine learning models (such as random forests and support vector machines) have been introduced into hyperspectral salinity inversion, improving prediction accuracy through nonlinear fitting, and gradually becoming the mainstream technical approach in this field.

[0003] However, existing machine learning methods for salinity inversion in coastal saline-alkali lands have two main problems: machine models are mostly black-box structures, making it difficult to explain the specific contribution of each spectral feature to salinity prediction, and they lack physical mechanism support, resulting in low reliability of the results. Furthermore, hyperspectral data has high dimensionality and strong redundancy between bands. Traditional feature selection methods (such as correlation coefficient method and principal component analysis) rely on manual threshold setting or linear assumptions, which cannot adaptively identify sensitive bands strongly correlated with salinity, and are prone to missing key features or introducing noise, thus limiting the generalization ability of the model. Therefore, a new satellite hyperspectral method for salinity inversion in coastal saline-alkali lands based on a machine learning model is proposed. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: A satellite hyperspectral method for salinity inversion of coastal saline-alkali land based on machine learning models includes: S1: Acquire raw multispectral images and ground-measured data, and obtain standardized initial feature spatial data through radiometric calibration and correction preprocessing. S2: Fractional-order differential spectral enhancement is performed on the initial feature space data to obtain the enhanced high-dimensional feature set, and the Boruta feature selection algorithm is used to adaptively optimize the importance of the enhanced high-dimensional feature set to obtain the optimal high-dimensional feature subset; S3: Based on the optimal high-dimensional feature subset, the initial inversion model is obtained by automatic optimization through integration of Optuna hyperparameters. Based on the initial inversion model, the salt inversion model is obtained by optimizing through feature contribution analysis and embedding the SHAP interpreter. S4: Based on the salinity inversion model, the salinity of remote sensing images across the entire region is predicted to achieve salinity inversion of saline-alkali land.

[0005] The process of obtaining the initial feature space data is as follows: The original image dataset a was constructed by collecting images of the saline-alkali land in the study area. Sampling points were set up in the study area using the grid method. The conductivity of the soil extract was measured by a DDSJ-307F conductivity meter and converted into the soil salt mass fraction S. Using Geographic Information System (GIS) technology, based on the original image dataset 'a' and soil salinity mass fraction 'S', the features corresponding to each ground sampling point are extracted after correction and resampling to the same projected coordinate system and spatial resolution, forming an original feature sample matrix X. All original features in the original feature sample matrix X are then processed. Standardize the data to obtain the initial feature space data B.

[0006] The process of fractional differential spectral enhancement is as follows: Extract the original image data sequence of all samples from the initial feature space data B. With soil salinity quality label Fractional differential transformation was performed on the original image data sequence using the Grünwald-Letnikov difference method. The spectral matrix is ​​obtained by performing a 0.5th-order differential calculation on each spectral sequence. ; The differentiated spectral matrix Spatial coordinates and soil salinity mass labels corresponding to the initial feature space data B Recombined, we obtain the enhanced high-dimensional feature set. .

[0007] The process of adaptively selecting the importance of the enhanced high-dimensional feature set is as follows: For the enhanced spectral matrix The shadow feature matrix is ​​obtained by performing a random permutation operation on each spectral feature band. The enhanced spectral matrix With shadow feature matrix By merging columns, we obtain the combined feature matrix. ; Combined feature matrix As input, to enhance the high-dimensional feature set Soil salinity quality label To predict the target, random forest regression is introduced to calculate the combined feature matrix. Importance score of each feature And extract the shadow feature matrix. Calculate the maximum value of the importance scores of all corresponding features. ; If The corresponding features are labeled as candidate important features. Only all candidate important features are retained to generate a shadow feature matrix, which yields the optimal feature matrix after filtering. The optimal feature matrix after screening , and the enhanced high-dimensional feature set The corresponding spatial coordinates and soil salinity quality labels are recombined to obtain the optimal high-dimensional feature subset. .

[0008] Introducing random forest regression to calculate the importance score of each feature in the combined feature matrix The process is as follows: For each decision tree in the random forest, calculate its mean squared error on the out samples, then randomly permutate the value of the feature and recalculate the mean squared error. The difference between the two errors is the importance contribution of the feature in the tree. The importance score of a feature is obtained by averaging the contributions of all decision trees. .

[0009] The process of integrating Optuna's automatic hyperparameter optimization is as follows: The optimal high-dimensional feature subset The optimal characteristic matrix in With soil salinity quality label The training set is obtained by randomly dividing the data in a 7:3 ratio. and verification set ; XGBoost was chosen as the basic model framework. The initial hyperparameter ranges were defined, and optimization objectives were set for the validation set. The prediction mean square error (MSE) is the optimization objective. The goal is to minimize the prediction error on the validation set. The mathematical expression of the optimization objective is as follows:

[0010] Where Θ is the combination of hyperparameters to be optimized. That is the number of samples in the validation set. It is the true salt content of the i-th sample. It is the predicted value of the i-th sample. Indicates the optimization error; Perform Bayesian optimization iterations and select the hyperparameter combination that minimizes the MSE of the validation set, denoted as the hyperparameter combination. .

[0011] The process of obtaining the salt inversion model is as follows:

[0012] The initial salt inversion model is obtained by minimizing the prediction error of the training set using the gradient descent algorithm. Embed the pre-prepared SHAP interpreter into the initial model The initial model was optimized through feature contribution analysis and tested on the validation set. Verify the model's prediction accuracy. After verification, the trained salt inversion model is obtained. .

[0013] The process of optimizing the initial model through feature contribution analysis is as follows: Calculate the feature SHAP value using the training set. Using the SHAP interpreter as input, the contribution of each feature to the model's prediction result is calculated, resulting in a sequence of SHAP values ​​for each feature. ; Set a threshold τ for the SHAP value, take the mean of the absolute values ​​of all feature SHAP values, and then remove them. The weak contribution characteristics from hyperparameter combination The columns corresponding to the strong contributing features are retained to obtain the updated modeling feature matrix. Retrain and optimize the model to model the updated feature matrix. With soil salinity quality label The optimal hyperparameter combination was obtained by re-integrating Optuna and performing automatic hyperparameter optimization. Training the initial salt inversion model .

[0014] The present invention has the following beneficial effects: 1. By combining fractional differential spectral enhancement and the Boruta feature selection algorithm, sensitive features closely related to soil salinity can be accurately extracted from hyperspectral data, effectively eliminating redundant information and noise. Furthermore, based on this, the Optuna framework is used to automatically optimize the hyperparameters of the XGBoost model, enabling the model to reach the best fit state. This significantly improves the accuracy, robustness, and generalization ability of salinity prediction in coastal saline-alkali land, ensuring the reliability of the inversion results at the regional scale.

[0015] 2. While constructing a high-precision salinity inversion model, this invention introduces the SHAP interpretability analysis method, which can quantify the contribution of each input feature, such as a specific spectral band, vegetation index, and topographic factor, to the model's prediction results. This makes the decision-making process of the salinity inversion model transparent, effectively overcoming the deficiency of traditional machine learning models in salinity inversion due to the lack of physical interpretation, and enhancing the credibility and scientific value of the inversion results.

[0016] 3. By employing the Boruta feature selection algorithm, which creates shadow features for importance comparison, features that significantly contribute to salinity prediction can be automatically and objectively identified. This avoids the subjective bias introduced by relying on manually set thresholds or single indicators in traditional methods, improving the efficiency and reliability of feature selection. It can not only generate high-precision spatial distribution maps of soil salinity, but also reveal the key driving factors affecting spatial differentiation of salinity and their contribution patterns based on SHAP analysis results. This provides a more comprehensive and in-depth scientific basis for the zoning, classification, precise management, and ecological restoration of coastal saline-alkali land. Attached Figure Description

[0017] Figure 1 This diagram illustrates the steps of the satellite hyperspectral method for salinity inversion of coastal saline-alkali land based on a machine learning model, as proposed in this invention. Detailed Implementation

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

[0019] Example 1

[0020] like Figure 1 As shown, the satellite hyperspectral method for salinity inversion of coastal saline-alkali land based on a machine learning model proposed in this invention includes: S1: Acquire raw multispectral images and ground-measured data, and obtain standardized initial feature spatial data through radiometric calibration and correction preprocessing. We selected typical saline-alkali lands in the study area, such as the coastal saline-alkali area of ​​the Yellow River Delta, using hyperspectral satellite images. Taking the data from the Gaofen-5 hyperspectral sensor as an example, the spectral resolution was 5nm and the spatial resolution was 30m. The images met the requirements of cloud cover <10% and the imaging time was during the dry season when the soil surface water content was stable to avoid water interference with the salinity spectral response. We constructed the original image dataset a. Sampling points were set up in the study area using a grid method with a depth of 500m×500m. Three samples of topsoil from the 0-20cm layer were collected at each sampling point. After natural air drying and grinding through a 2mm sieve, the conductivity of the soil extract was measured using a DDSJ-307F conductivity meter and converted into the soil salt mass fraction S. Then, spatial registration is performed. Using Geographic Information System (GIS) technology, the original image dataset a and soil salinity mass fraction S are corrected and resampled to the same projection coordinate system and spatial resolution. The data features of all original image datasets a and soil salinity mass fraction S corresponding to each ground sampling point are extracted to form a long feature vector. The feature vectors of all sample points are stacked to form the original feature sample matrix X. In the original feature sample matrix X, the numerical ranges and dimensions of different features (such as reflectivity and elevation) vary greatly, which can lead to difficulties in model training. Therefore, standardization is performed to eliminate the influence of dimensions. Standardization is performed on each column of the matrix (i.e., each feature). For the j-th feature of the i-th sample, the standardization formula is:

[0021] in, Let i represent the j-th feature of the i-th sample. , Let j be the average value of the j-th feature. Let be the standard deviation of the j-th feature; For all original features in the original feature sample matrix X Standardize to obtain initial feature space data B, including the preprocessed original image dataset a and soil salinity mass fraction S; Specifically, the initial feature space data B is a standardized matrix that incorporates information from multiple sources and has undergone rigorous preprocessing to form a high-quality, dimensionless initial feature space, providing a reliable and uniform input for subsequent steps.

[0022] S2: Fractional-order differential spectral enhancement is performed on the initial feature space data to obtain the enhanced high-dimensional feature set, and the Boruta feature selection algorithm is used to adaptively optimize the importance of the enhanced high-dimensional feature set to obtain the optimal high-dimensional feature subset; Extract the original image data sequence of all samples from the initial feature space data B. With soil salinity quality label ,in: Image raw data sequence: Each element corresponds to the original image dataset a of the original sample; Soil salinity quality label: Each element corresponds to the soil salinity mass fraction S of the Central Plains sample; For saline-alkali soils, the response of salinity to the spectrum exhibits a weakly nonlinear characteristic. Preliminary experiments verified that the correlation coefficient between the spectral characteristics and salinity reaches its maximum when the derivative order is 0.5. Therefore, a fractional derivative order was selected. ; Fractional differential transformation was performed on the original image data sequence using the Grünwald-Letnikov difference method. Each spectral sequence in the image is subjected to a 0.5th-order differential calculation for the original image data sequence. Reflectivity of the t-th band The calculation rule for its fractional derivative result is to perform a weighted difference on the reflectance of the current band and all previous bands, expressed as:

[0023] in, Represents the original image data sequence Reflectivity of the t-th band The spectral matrix obtained after the 0.5th order fractional differentiation operation, where Γ is the gamma function. The factorial of index k, These are the historical reflectance values ​​used in the calculation. For bands, Indicates the first k bands; The differentiated spectral matrix Spatial coordinates and soil salinity mass labels corresponding to the initial feature space data B Recombined, we obtain the enhanced high-dimensional feature set. Its structure is as follows: ,in, Representing spatial coordinates, enhanced high-dimensional feature set The spectral characteristics of the sample better reflect the subtle spectral differences caused by salinity, while suppressing the interference of background noise, and are the only input source for the subsequent Boruta algorithm. The process of adaptively optimizing the importance of the enhanced high-dimensional feature set using the Boruta feature selection algorithm is as follows: From the enhanced high-dimensional feature set Separate the enhanced spectral matrix Soil salinity quality label ; Among them, the enhanced spectral matrix This contains the spectral sequences of all samples after 0.5-order fractional derivative processing. Each element corresponds to the differential spectral response value of the sample in a specific band, serving as the basic input for subsequent feature selection. (Soil salinity quality label) The measured soil salinity mass fraction of all samples will serve as the reference benchmark for assessing the importance of the feature. For the enhanced spectral matrix Each spectral feature band is randomly permuted to shuffle the order of eigenvalues ​​within a column, resulting in a spectral matrix. Shadow feature matrix with completely identical dimensions The role of shadow features is to provide a random benchmark to distinguish the true importance of the original features from random noise; The enhanced spectral matrix With shadow feature matrix By merging columns, we obtain the combined feature matrix. This combined matrix contains both the effective features to be evaluated and the random features used for comparison, and is the input to the Boruta feature selection algorithm. Combined feature matrix As input, to enhance the high-dimensional feature set Soil salinity quality label To predict the target, random forest regression is introduced to calculate the combined feature matrix. The reduction in mean squared error for each feature is as follows: For each decision tree in the random forest, calculate its mean squared error on the out samples, then randomly permutate the value of the feature and recalculate the mean squared error. The difference between the two errors is the importance contribution of the feature in the tree. The contribution values ​​of all decision trees are averaged to obtain the importance score Imp of the feature. This score reflects the degree of influence of the feature on the salt label prediction result. The higher the score, the stronger the association between the feature and salt. From the combined feature matrix In the process, extract the shadow feature matrix. Calculate the maximum value of the importance scores of all corresponding features. , This represents the highest level of importance that random features can achieve, serving as a benchmark for judging whether the original features are effective; For the spectral matrix For each feature in the dataset, its importance score Imp is compared with... Comparison: If This indicates that the importance of this feature exceeds that of random noise, and it is marked as a candidate important feature. If this feature appears multiple times consecutively (set to 10 times), it is considered an important feature. This indicates that its importance does not exceed the random level, and it is marked as a non-important feature and removed from [the list]. Remove from the middle; After removing the labeled non-important features, update the original feature matrix and regenerate the shadow feature matrix based on the updated matrix. The iteration continues until all features in the original feature matrix have been marked as important or unimportant features, at which point the iteration terminates. Collect all spectral features marked as important features from the spectral matrix. Extract the columns corresponding to these features to obtain the optimal feature matrix after filtering. The optimal feature matrix after screening , and the enhanced high-dimensional feature set The corresponding spatial coordinates G and soil salinity quality label Recombining yields the optimal high-dimensional feature subset. Optimal high-dimensional feature subset The spectral features strongly correlated with soil salinity were preserved, while redundant and noisy features were eliminated, and the features were directly used to construct subsequent salinity inversion models.

[0024] S3: Based on the optimal high-dimensional feature subset, the initial inversion model is obtained by automatic optimization through integration of Optuna hyperparameters. Based on the initial inversion model, the salt inversion model is obtained by optimizing through feature contribution analysis and embedding the SHAP interpreter. The optimal high-dimensional feature subset The optimal characteristic matrix in With soil salinity quality label The training set is obtained by randomly dividing the data in a 7:3 ratio. It contains 70% of the sample data, which will be used for the model's parameter learning and hyperparameter optimization process; Validation set It contains 30% sample data, which will be used to evaluate the model's generalization ability on data not used in the training, providing an objective performance benchmark for hyperparameter optimization; With training set As input, the XGBoost model was selected as the basic model framework, and automatic hyperparameter optimization was performed using Optuna. The specific process is as follows: Define the hyperparameter search space. Based on the characteristics of the XGBoost model, set the value range of the initial hyperparameters, including tree depth (3-10), learning rate (0.01-0.3), and subsample ratio (0.6-1.0). Set optimization goals to validate the set. The prediction mean square error (MSE) is the optimization objective, which is to minimize the prediction error on the validation set. The mathematical expression of this optimization objective is as follows:

[0025] Where Θ is the combination of hyperparameters to be optimized. That is the number of samples in the validation set. It is the true salt content of the i-th sample. Given the hyperparameter Θ, L(Θ) represents the model's prediction for the i-th sample. The goal is to find the Θ that minimizes L(Θ). Indicates the optimization error; Then, Bayesian optimization iteration is performed. Optuna uses the Bayesian optimization algorithm, which adaptively searches for a better hyperparameter combination based on the hyperparameter combinations of historical iterations and the corresponding validation set error. The number of iterations is set to 50 to ensure the sufficiency of the search. After the iteration is complete, the hyperparameter combination that minimizes the MSE of the validation set is selected, and denoted as the hyperparameter combination. This combination will serve as the parameter benchmark for subsequent model training; By minimizing the prediction error on the training set using the gradient descent algorithm, the training set is learned. The nonlinear correlation between spectral characteristics and salinity labels ultimately yields the initial salinity inversion model. The training process of this model can be described as solving the following optimization problem:

[0026] Where F is the XGBoost model space, and f is the specific model function. It is a loss function. and These are the feature vector and the true salt value of the j-th sample in the training set, respectively. The process of optimizing and embedding the SHAP interpreter through feature contribution analysis is as follows: Embed the pre-prepared SHAP interpreter into the initial model. The model is optimized through feature contribution analysis. The process is as follows: First, calculate the feature SHAP value using the training set. Using the SHAP interpreter as input, the contribution of each feature to the model's prediction result is calculated, resulting in a sequence of SHAP values ​​for each feature. , The larger the absolute value of the value, the stronger the influence of the feature on salinity prediction; Then, strong contributing features are selected by setting a SHAP value threshold τ, and the mean of the absolute values ​​of all feature SHAP values ​​is taken to eliminate those features. The weak contribution characteristics from hyperparameter combination The columns corresponding to the strong contributing features are retained to obtain the updated modeling feature matrix. Retrain and optimize the model to model the updated feature matrix. With soil salinity quality label The optimal hyperparameter combination was obtained by re-integrating Optuna and performing automatic hyperparameter optimization. Training the initial salt inversion model ; With validation set The model is used as input to test its prediction accuracy. After validation, the trained salt inversion model is obtained. ; Specifically, the salinity inversion model We learned to select key spectral features using fractional-order differential enhancement and the Boruta algorithm, and to accurately infer the complex nonlinear mapping relationship of soil salinity content, as well as the salinity inversion model. Based on SHAP analysis, features that contribute little to the prediction or may even introduce noise are eliminated, allowing the model ontology to focus more on strongly correlated features. This significantly improves the accuracy, stability, and generalization ability of the prediction, ensuring its reliability when applied in unknown regions.

[0027] S4: Based on the salinity inversion model, the salinity of the entire remote sensing image is predicted to realize the salinity inversion of the sea saline-alkali land; Acquire hyperspectral remote sensing images of the entire study area, perform radiometric calibration and correction preprocessing on them with S1, and obtain initial feature spatial data; Extracting the optimal high-dimensional feature subset pixel-by-pixel from the initial feature space data. The optimal eigenmatrix Consistent spectral bands, i.e., the characteristic bands after fractional-order differential enhancement and Boruta algorithm filtering, are used to generate a global spectral feature matrix, which is then loaded into the salinity inversion model. The global spectral feature matrix is ​​input into the salinity inversion model. In the process, the soil salinity mass fraction at the corresponding location is predicted pixel by pixel, resulting in a global salinity prediction sequence. ; Sequence of global salinity prediction results By matching the spatial coordinate information of pixels with the geographical boundaries of the study area, the prediction results are rendered into a thematic map of the spatial distribution of salinity in the sea saline-alkali land using GIS tools. The salinity is classified into levels according to the salinity content, such as light, moderate and heavy saline-alkali areas, to complete the salinity inversion of the entire sea saline-alkali land.

[0028] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0029] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for retrieving salt content of coastal saline-alkali soil based on a machine learning model, characterized in that, Comprise: S1: Collect original multispectral images and ground measured data, and obtain standardized initial feature space data through radiation calibration correction preprocessing operation; S2: Perform fractional order differential spectral enhancement on the initial feature space data to obtain an enhanced high-dimensional feature set, and use the Boruta feature selection algorithm to adaptively optimize the importance of the enhanced high-dimensional feature set to obtain an optimal high-dimensional feature subset; S3: Based on the optimal high-dimensional feature subset, an initial inversion model is obtained by integrating Optuna hyperparameter automatic optimization, and based on the initial inversion model, a salt inversion model is obtained by feature contribution degree analysis optimization and embedding SHAP interpreter; S4: Based on the salt inversion model, the salt content of the global remote sensing image is predicted to realize the salt content inversion of the coastal saline land.

2. The satellite hyperspectral coastal salinized soil salinity inversion method based on a machine learning model according to claim 1, characterized in that, The process of obtaining the initial feature space data is: Select the coastal saline land in the study area to collect and construct image original data set a, and use the grid point distribution method to distribute the sampling points in the study area, measure the conductivity of the soil extract by using the DDSJ-307F conductivity meter, and convert it into the soil salt content mass fraction S; Using the geographic information system (GIS) technology, based on the image original data set a and the soil salt mass fraction S, the corresponding features of each ground sampling point are extracted after correction and resampling to the same projection coordinate system and spatial resolution, to form an original feature sample matrix X. All original features in the original feature sample matrix X are normalized to obtain initial feature space data B. Standardization is performed to obtain initial feature space data B.

3. The satellite hyperspectral coastal salinized soil salinity inversion method based on a machine learning model according to claim 2, characterized in that, The process of fractional order differential spectral enhancement is: From the initial feature space data B, extract the image original data sequence of all samples With the soil salt quality label , perform fractional differential transformation, and use Grünwald-Letnikov difference method to calculate 0.5 order differential for each spectrum sequence in the image original data sequence Obtain the differential spectrum matrix ; The differentiated spectral matrix is recombined to obtain an enhanced high-dimensional feature set corresponding to the spatial coordinates and soil salinity quality labels in the initial feature space data B recombined to obtain an enhanced high-dimensional feature set .

4. The machine learning model based satellite hyperspectral coastal saline soil salinity-inversion method according to claim 1, characterized in that, The process of adaptively optimizing the importance of the enhanced high-dimensional feature set is: For the enhanced spectral matrix The shadow feature matrix is ​​obtained by performing a random permutation operation on each spectral feature band. The enhanced spectral matrix With shadow feature matrix By merging columns, we obtain the combined feature matrix. ; Combined feature matrix As input, to enhance the high-dimensional feature set Soil salinity quality label To predict the target, random forest regression is introduced to calculate the combined feature matrix. Importance score of each feature And extract the shadow feature matrix. Calculate the maximum value of the importance scores of all corresponding features. ; If The corresponding feature marks are candidate important features, and only the shadow feature matrix generated by retaining all candidate important features is obtained to generate the screened optimal feature matrix , the screened optimal feature matrix , and the enhanced high-dimensional feature set The corresponding spatial coordinates and soil salt quality labels are recombined to obtain the optimal high-dimensional feature subset .

5. The machine learning model based satellite hyperspectral coastal saline soil salinity inversion method according to claim 4, characterized in that, Importing the random forest regression computes the importance score for each feature in the combined feature matrix The process is as follows: For each decision tree in the random forest, calculate the mean square error on the external sample, then randomly replace the value of the feature and recalculate the mean square error, and the difference between the two errors is the importance contribution of the feature in the tree; The contribution values for all decision trees are averaged to obtain the importance score for the feature .

6. The satellite hyperspectral coastal salinized soil salinity inversion method based on a machine learning model according to claim 5, characterized in that, The process of integrating Optuna hyperparameter automatic optimization is: optimal feature matrix in the optimal high-dimensional feature subset optimal feature matrix in the optimal high-dimensional feature subset soil salt quality label randomly divided in a ratio of 7:3 to obtain a training set and a verification set ; XGBoost is chosen as the base model framework to define the range of values for the initial hyperparameters and set the optimization objective for the validation set The mean squared error (MSE) of the predictions on the validation set is chosen as the optimization objective, which is to minimize the prediction error on the validation set. The mathematical expression of the optimization objective is as follows: where Θ is the combination of hyperparameters to be optimized, is the number of samples of the validation set, is the true salinity value of the i-th sample, is the predicted value of the i-th sample, denotes the optimization error; performing a Bayesian optimization iteration, selecting a combination of hyperparameters that minimizes the validation set MSE, denoted as hyperparameters .

7. The satellite hyperspectral coastal salinized soil salinity inversion method based on a machine learning model according to claim 6, characterized in that, The process of obtaining the salt inversion model is: An initial salinity inversion model is obtained by minimizing the prediction error of the training set through a gradient descent algorithm , the SHAP interpreter prepared in advance is embedded into the initial model , the initial model is optimized through feature contribution degree analysis, and the prediction accuracy of the model is verified through the verification set , after the verification is completed, the trained salinity inversion model is obtained .

8. The machine learning model based satellite hyperspectral coastal saline soil salinity inversion method according to claim 7, characterized in that, The process of optimizing the initial model by feature contribution degree analysis is: The SHAP value of the calculation feature is calculated, taking the training set As input, the SHAP interpreter is used to calculate the contribution of each feature to the model prediction result, and the SHAP value sequence of each feature is obtained ; Set SHAP value threshold τ, take the mean of the absolute value of all feature SHAP values, remove weakly contributing features from the hyperparameter combination , and keep the columns corresponding to strongly contributing features to obtain an updated modeling feature matrix . Retrain the optimization model with the updated modeling feature matrix and the soil salinity quality label to obtain the optimal hyperparameter combination obtained by re-integrating Optuna for automatic hyperparameter optimization Train the initial salinity inversion model .

Images