Salinization type quantitative inversion method based on feature weighting and residual correction
By employing an integrated regression modeling framework based on feature weighting and residual correction in soil salinization monitoring, and through multi-model feature selection and fusion, an RF model and a residual correction model are constructed, and an integrated regression modeling framework with residual correction is used to quantitatively invert soil salinization types in conjunction with remote sensing image data. This solves the problems of difficult data acquisition and insufficient model expressive power in existing technologies, and achieves efficient and accurate monitoring of salinization types.
Patent Information
- Application Number
- CN202511327310.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-17
- Publication Date
- 2026-02-17
AI Technical Summary
Existing technologies for soil salinization monitoring suffer from problems such as difficulty in data acquisition, insufficient model expressive power, and poor robustness and prediction accuracy due to the limited range of feature selection strategies, making it difficult to achieve efficient and accurate regional-scale inversion.
An integrated regression modeling framework based on feature weighting and residual correction was adopted. Through feature selection and fusion of multiple models, RF model and residual correction model were constructed, and the soil salinization type was quantitatively inverted by combining remote sensing image data.
It significantly improves the accuracy and stability of the model, enhances its adaptability to heterogeneous data and complex environments, and improves the generalization performance and application reliability of salting type inversion.
Smart Images
Figure CN121545609A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of remote sensing technology and machine learning, and in particular to a quantitative inversion method for salinization type based on feature weighting and residual correction. Background Technology
[0002] Soil salinization is a significant ecological and agricultural problem commonly faced in arid and semi-arid regions, characterized by the abnormal accumulation of soluble salts in surface soil, severely impacting crop growth, soil structure stability, and regional ecosystem functions. Xinjiang, located in a typical arid region of my country, is one of the areas with the highest concentration of saline-alkali land in the country. Due to its complex hydrological, topographical, and climatic conditions, coupled with historical mismanagement of water resources, soil salinization is particularly severe, posing significant challenges to remediation and requiring robust monitoring. Although Xinjiang has made some progress in saline-alkali land improvement in recent years, the problem of soil salinization remains serious, necessitating more efficient and precise monitoring and management methods.
[0003] While relying on ground sampling and laboratory chemical analysis offers significant advantages in data accuracy, it also has considerable limitations. Firstly, ground sampling is constrained by personnel, time, and spatial conditions, especially in vast and topographically complex regions like Xinjiang, making high-frequency, large-scale data acquisition difficult. Secondly, traditional monitoring methods can only acquire discrete point information, failing to reflect the spatial continuity and dynamic trends of salinization. To improve monitoring efficiency, an increasing number of studies are attempting to utilize remote sensing and machine learning methods for spatial inversion of soil salinization. However, existing research generally relies on single machine learning algorithms. While these models perform well in some scenarios, they often suffer from limited information mining capabilities, insufficient model expressive power, and poor generalization when dealing with high-dimensional remote sensing features, multi-source heterogeneous data, and nonlinear relationships, making it difficult to effectively support high-precision inversion tasks at the regional scale.
[0004] Furthermore, the singularity of feature selection strategies is also a key factor limiting model performance. Current common methods, such as Pearson correlation analysis and model importance ranking, often focus on only one aspect of the data's correlation features, lacking cross-method collaboration mechanisms. This easily leads to the neglect of comprehensive identification and processing of feature redundancy and weakly sensitive features. Such a single-point dependent selection mechanism can easily result in the omission of important information or the retention of interfering features, thereby affecting the robustness and prediction accuracy of the final model.
[0005] To address the aforementioned issues, there is an urgent need to construct an integrated regression modeling framework that incorporates weighted features from multiple models and possesses residual correction capabilities. This framework should not only fully explore the multi-dimensional relationships between different types of features and soil salinity concentration, but also significantly improve the stability and interpretability of the inversion results through model structure optimization, providing a more efficient and accurate technical approach for the classification and mapping of salinization types in arid regions. Summary of the Invention
[0006] The purpose of this invention is to provide a quantitative inversion method for salinization types based on feature weighting and residual correction, in order to solve the technical problem that traditional soil salinization inversion methods cannot meet the needs of current practical applications.
[0007] The above-mentioned objective of this application is achieved through the following technical solution: S1: Collect measured concentration data of soil salt ions in the target area; acquire remote sensing images of the target area and extract its multi-source feature information; S2: Construct a quantitative inversion dataset for salinization types using measured concentration data and multi-source feature information; S3: Use feature selection methods to evaluate the correlation between multi-source feature information and soil salt ion concentration in the quantitative inversion dataset of salinization types, and form a unified feature importance weight vector through standardization and weighted fusion; construct a weighted feature set through the feature importance weight vector; S4: Construct an RF model and train it using a weighted feature set to obtain the prediction residual value of the prediction result; S5: Based on the predicted residual values, construct a residual correction model and train it using a weighted feature set; S6: Acquire the remote sensing image data to be tested; S7: Using the trained RF model and residual correction model and the remote sensing image data to be measured, predict and generate ion concentration data of the target area, and generate a two-dimensional spatiotemporal distribution map of soil salinization type for quantitative classification.
[0008] Optionally, the multi-source feature information includes: vegetation index, salinity index, and multi-band reflectance.
[0009] Optionally, step S2 includes: S21: Preprocess the remote sensing images of the study area; S22: Extract multi-source feature information of the corresponding locations in the remote sensing image based on the geographic coordinates of the soil sampling points in the target area; S23: Spatial matching of extracted multi-source feature information with measured concentration data to form image and ground observation data pairs; S24: Based on image and ground observation data pairs, calculate various vegetation indices and salinity indices to construct a comprehensive dataset that can be used for quantitative inversion of salinization types, namely, the salinization type quantitative inversion dataset.
[0010] Optionally, the preprocessing includes: radiometric calibration and atmospheric correction.
[0011] Optionally, step S3 includes: Feature selection methods include: XGBoost gain, Pearson correlation analysis, and mutual information; The gain index of the XGBoost model is used to calculate the contribution of each input feature in the quantitative inversion dataset of salinization type to the prediction of soil salt ion concentration in the regression model, and the gain score of each feature is obtained. Pearson correlation analysis was performed on the same batch of input features to calculate the linear correlation coefficient between each feature and the soil salt ion concentration, and the absolute value of the coefficient was taken as the correlation score. The mutual information method is used to evaluate the degree of nonlinear dependence between each input feature and soil salt ion concentration, and to obtain the mutual information score of each feature. The gain score, correlation score, and mutual information score were normalized separately. For the three normalized scores, a weighted average or weight fusion strategy is used to calculate the comprehensive importance score of each feature, and the fused feature score is obtained. Based on the fused feature scores, a feature importance weight vector is formed, and the original feature matrix is weighted to construct a weighted feature set.
[0012] Optionally, the residual correction model adopts the XGBoost model; The RF model and XGBoost model were trained using the 5-fold cross-validation method.
[0013] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform a salting-type quantitative inversion method based on feature weighting and residual correction.
[0014] A computer-readable storage medium storing instructions that, when executed, perform a salting-type quantitative inversion method based on feature weighting and residual correction.
[0015] The beneficial effects of the technical solution provided in this application are: This invention innovatively introduces a residual learning mechanism into the field of multispectral remote sensing salinization inversion. By selecting an appropriate machine learning algorithm, a two-stage residual learning regression model is constructed. The first stage predicts preliminary results, and the second stage performs in-depth modeling and correction on the predicted residuals, effectively capturing complex nonlinear relationships and potential error sources. Compared with traditional single models or simple ensemble methods, residual learning significantly reduces prediction errors and improves the model's accuracy and stability. Furthermore, this method enhances the model's adaptability to heterogeneous data and complex environmental changes, significantly improving the generalization performance and application reliability of salinization type inversion. Attached Figure Description
[0016] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a step diagram of an embodiment of this application; Figure 2 This is a schematic diagram of the model in the embodiments of this application; Figure 3 This is a diagram of the training and prediction process in the embodiments of this application; Figure 4 This is the XGBoost gain feature importance score map in the embodiments of this application; Figure 5 This is a Pearson correlation analysis feature importance score graph in the embodiments of this application; Figure 6 This is a MI feature importance score map in the embodiments of this application; Figure 7 This is the importance weight map of the fusion features in the embodiments of this application; Figure 8 This is the fusion weight bar chart in the embodiments of this application; Figure 9 This is an example of the embodiments in this application. Scatter plot of prediction results; Figure 10 This is a scatter plot of the prediction results of a single master model in the embodiments of this application; Figure 11 This is a scatter plot of the prediction results of a single master model using a feature-weighted dataset in the embodiments of this application; Figure 12 This is a scatter plot of the prediction results of the main model + residual correction model in the embodiments of this application; Figure 13 This is a scatter plot of the prediction results of the main model + residual correction model using the feature-weighted dataset in the embodiments of this application; Figure 14 This is a spatiotemporal distribution map of salinization types in the study area in this application embodiment; Figure 15This is a schematic diagram of the electronic device structure in the embodiments of this application; Figure 16 This is a scatter plot of the prediction results of the main model + residual correction model using the feature-weighted dataset in the embodiments of this application; Figure 17 This is a comparison chart of the accuracy of each model in the embodiments of this application. Detailed Implementation
[0017] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0018] Embodiments of this application provide a quantitative inversion method for salting type based on feature weighting and residual correction.
[0019] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a quantitative inversion method for salting type based on feature weighting and residual correction in an embodiment of this application, including: S1: Collect measured concentration data of soil salt ions in the target area; acquire remote sensing images of the target area and extract its multi-source feature information; S2: Construct a quantitative inversion dataset for salinization types using measured concentration data and multi-source feature information; S3: Use feature selection methods to evaluate the correlation between multi-source feature information and soil salt ion concentration in the quantitative inversion dataset of salinization types, and form a unified feature importance weight vector through standardization and weighted fusion; construct a weighted feature set through the feature importance weight vector; S4: Construct an RF model and train it using a weighted feature set to obtain the prediction residual value of the prediction result; S5: Based on the predicted residual values, construct a residual correction model and train it using a weighted feature set; As one example, a residual correction model is constructed based on a weighted feature set. The first stage is used to predict the concentration of various soil ions, and the second stage uses the prediction residuals of the first stage as the learning target to perform residual modeling in order to improve the overall prediction accuracy.
[0020] S6: Acquire the remote sensing image data to be tested; S7: Using the trained RF model and residual correction model and the remote sensing image data to be measured, predict and generate ion concentration data of the target area, and generate a two-dimensional spatiotemporal distribution map of soil salinization type for quantitative classification.
[0021] This application proposes a remote sensing feature importance assessment mechanism based on the weighted fusion of multiple feature selection algorithms, effectively integrating linear and nonlinear correlation information. A residual correction structure is designed to improve the accuracy and generalization of ion concentration inversion. A fully automated method is implemented from ion concentration inversion to standard land type classification, providing technical support for refined management of soil salinization. This application features high accuracy, strong generalization, and good operability, and is applicable to multiple application scenarios such as agricultural planning, salinization control, and land resource management. Specifically, this application covers key technical aspects such as remote sensing data acquisition and preprocessing, soil ion concentration analysis, feature selection and fusion, regression modeling and residual learning, standardized classification of salinization types, and geographic information system (GIS) integration and result visualization.
[0022] The multi-source feature information includes: vegetation index, salinity index, and multi-band reflectance.
[0023] Step S2 includes: S21: Preprocess the remote sensing images of the study area; S22: Extract multi-source feature information of the corresponding locations in the remote sensing image based on the geographic coordinates of the soil sampling points in the target area; S23: Spatial matching of extracted multi-source feature information with measured concentration data to form image and ground observation data pairs; S24: Based on image and ground observation data pairs, calculate various vegetation indices and salinity indices to construct a comprehensive dataset that can be used for quantitative inversion of salinization types, namely, the salinization type quantitative inversion dataset.
[0024] As one example, when collecting multispectral remote sensing image data required for soil salinization inversion in the target area, remote sensing images within 5 days before and after the sampling date are preferred. The time window can be flexibly adjusted according to specific environmental and application requirements. To ensure the quality of the inversion data, the selected images must meet high image quality standards. In addition to meeting the requirements for temporal and spatial coverage, they must also have conditions such as cloud cover below 10%, clear weather, and no clouds in the target area.
[0025] In the specific operation, the image acquisition indicators are first defined based on the geographical scope, time span, and required band configuration of the study area to narrow down the data selection range. Subsequently, using open remote sensing data platforms such as Google Earth Engine (GEE) and the European Space Agency's Copernicus program, the system searches for multispectral remote sensing image data that meets the requirements by setting filtering conditions such as keywords, geographic coordinates, time intervals, and cloud cover thresholds.
[0026] After selecting suitable data, use the preview tool provided by the data platform to perform a preliminary quality check on the images to ensure they meet research requirements. Once confirmed, set the download parameters (such as data format, spatial resolution, etc.) according to the requirements to complete the data acquisition.
[0027] Multiple soil sampling points were set up within the target area, collecting data covering the concentrations of eight soluble salt ions and their corresponding precise geographic coordinates. The collected soil data underwent preprocessing, including outlier removal, missing value imputation, and normalization transformation, to improve data quality and consistency, laying a solid foundation for subsequent analysis and model building.
[0028] Finally, through precise spatial and temporal matching, remote sensing image data and soil sampling data are fused to construct a comprehensive dataset for soil salinization inversion. Based on this dataset, combined with multi-model feature weighting and residual correction models, the cost of data acquisition in traditional inversion methods is effectively reduced, while significantly improving inversion accuracy and model generalization ability.
[0029] The preprocessing includes: radiometric calibration and atmospheric correction.
[0030] As one embodiment, after data acquisition, the remote sensing image data undergoes systematic preprocessing, including two key steps: radiometric calibration and atmospheric correction, to ensure the physical consistency and spectral accuracy of subsequent model input data. Specifically: Radiometric calibration: This process converts the raw digital values (DN values) of remote sensing images into physically meaningful radiance or reflectance values. This process is typically based on calibration coefficients in the image metadata or uses radiometric reference targets within the image to ensure the comparability of radiometric data across different time periods and sensors, providing support for multi-source data fusion and physical modeling.
[0031] Atmospheric correction: Based on atmospheric transport models and atmospheric information in imagery (such as water vapor content, aerosol concentration, etc.), atmospheric effects are inverted and corrected in remote sensing data, effectively reducing spectral distortion caused by atmospheric scattering and absorption, and improving the accuracy of ground reflectance extraction.
[0032] Step S3 includes: Feature selection methods include: XGBoost gain, Pearson correlation analysis, and mutual information; The gain index of the XGBoost model is used to calculate the contribution of each input feature in the quantitative inversion dataset of salinization type to the prediction of soil salt ion concentration in the regression model, and the gain score of each feature is obtained. Pearson correlation analysis was performed on the same batch of input features to calculate the linear correlation coefficient between each feature and the soil salt ion concentration, and the absolute value of the coefficient was taken as the correlation score. The mutual information method is used to evaluate the degree of nonlinear dependence between each input feature and soil salt ion concentration, and to obtain the mutual information score of each feature. The gain score, correlation score, and mutual information score were normalized separately. For the three normalized scores, a weighted average or weight fusion strategy is used to calculate the comprehensive importance score of each feature, and the fused feature score is obtained. Based on the fused feature scores, a feature importance weight vector is formed, and the original feature matrix is weighted to construct a weighted feature set.
[0033] The residual correction model adopts the XGBoost model; The RF model and XGBoost model were trained using the 5-fold cross-validation method.
[0034] As one example, a two-stage model based on weighted features is constructed. The first stage uses an RF model for preliminary prediction of soil ion concentration, and the second stage uses an XGBoost model for residual learning and error correction. (Refer to...) Figure 2 , Figure 2 This is a schematic diagram of a quantitative inversion method for salting type based on a multi-model feature weighting and residual correction model in an embodiment of this application.
[0035] As one implementation, the weighted feature set is divided into a training set and a test set; a 5-fold cross-validation is set, and the training set is randomly divided into 5 mutually exclusive subsets, which are used as the validation set in turn, while the remaining subsets are used for training; in each fold, the RF model is used to train the training subset to obtain the corresponding prediction result, and the residual is calculated; the prediction results of each fold are summarized and averaged to obtain a stable first-stage prediction output; based on the prediction residual of the first stage, a second-stage residual correction model is constructed, and 5-fold cross-validation is used for training; the first-stage and second-stage models are integrated into a complete two-stage residual learning structure for predicting soil ion concentration; the predicted value is compared with the true value to evaluate the prediction error, and the relevant parameters or feature weighting strategy can be adjusted through model iteration to further improve the model performance.
[0036] It should be noted that the residual correction model in this application is not limited to the XGBoost model. It can also be replaced with other regression models with good performance, such as random forest, LightGBM, and support vector regression, to optimize the accuracy of the residual learning stage, depending on the specific application requirements. Similarly, the feature weighting fusion method used is not limited to the three feature selection strategies of XGBoost Gain, Pearson correlation analysis, and mutual information exemplified in this application. Other feature selection methods with correlation evaluation capabilities can also be used, as long as they can measure and weight the importance of the correlation between the input features and the prediction target.
[0037] As one example, Figure 3 This diagram illustrates the training and prediction process of a quantitative inversion method for salting types based on a multi-model feature weighting and residual correction model, as described in this application. The main model and residual correction model are not limited to any particular machine learning model. In the experimental process description of this patent, the XGBoost model is selected for demonstration. Through feature weighting fusion in stage one, the features are input into the main model for initial prediction. The predicted values are then compared with the corresponding true values to obtain residual values. These residual values are then further input into the residual correction model for residual prediction. Finally, a new residual value is predicted, and related calculations are performed to obtain the final predicted value, which serves as the final output of the model.
[0038] As one example, Figure 3 The method of feature weighting (feature engineering method) is not limited to simple average weighting; it can also assign weights separately based on the performance of each feature engineering condition. The relevant formulas are as follows: (1) The input features are as follows:
[0039] in This represents the y-th input feature; (2) For each feature engineering method Calculate the importance score for each feature. : Taking three feature engineering methods as examples (XG Boostgain, Pearson correlation analysis, and mutual information), the importance scores of each feature engineering method are obtained. , where x represents the x-th feature and y represents the y-th input feature; (3) Normalization calculation of importance score: The scores for each method are min-max normalized to fall within the [0,1] interval:
[0040] (4) Calculate the importance score weights using softmax: For each feature The softmax values of the normalized scores under the three methods are used as the final importance score weights:
[0041] in, This represents the importance score weight of the y-th feature. The temperature parameter is n, which represents the number of feature engineering methods used (in this experiment, n=3, which represents the three feature engineering methods: XGBoostgain, Pearson correlation analysis, and MI). (5) Integration and Scale Standardization: Weighted importance of each feature:
[0042] in, This represents the weighted importance score of the y-th input feature. This represents the importance score weight of the y-th feature. This represents the normalized importance score of the y-th feature; (6) Feature weighting to generate final input features: Ultimate importance As a feature The weights are applied to the original training and test sets:
[0043] Obtain the weighted input feature matrix , which serves as the training input for the main model.
[0044] As one example, the residual correction process is as follows: (1) Main model training and prediction. After weighting the features of the first-stage model, the predicted value of the first-stage main model is output. And then proceed to the second phase.
[0045] (2) Residual calculation. Based on the true labels of the training data, calculate the residuals of the main model: (3) Residual modeling. Further, the calculated residuals are used as labels, and the first-stage predicted values are used as features, input into the second-stage model to obtain the predicted residual values. This serves as the basis for calculating the final predicted value.
[0046] (4) Prediction correction. The predicted values of the master model and the predicted values of the residuals are added together to obtain the final prediction result.
[0047] As one example, with For example, the experimental process and results are as follows (data acquisition and dataset creation process omitted): (1) The visualization results of the importance scores of each feature engineering (XGBoost gain, Pearson correlation analysis and MI) are as follows: Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 The results of the entire feature fusion are shown, in which Figure 7 This demonstrates the feature importance weights after feature fusion, which leads to the final weighted features used as model input: (2) The fitting results of the quantitative inversion method for salting type based on multi-model feature weighting and residual correction model mentioned in this experiment show the coefficient of determination. The value is 0.741, MAE is 0.249, and RMSE is 0.291. The corresponding scatter plot is shown below. Figure 9 As shown.
[0048] (3) To further verify the quantitative inversion method for salting type based on multi-model feature weighting and residual correction model mentioned in this application, the following comparative experimental results are given. The accuracy comparison results are shown in Table 1. Figure 10 , Figure 11 , Figure 12 and Figure 13 The following are scatter plots of the prediction results for the corresponding models, in order. Figure 14 This is a spatiotemporal distribution map of salinization types in the study area; Figure 16 This is a scatter plot of the prediction results from the main model + residual correction model using the feature-weighted dataset. Figure 17 The diagram shows a comparison of the accuracy of each model. It is clear that the quantitative inversion method for salting type based on multi-model feature weighting and residual correction model mentioned in this application performs best, with a determination coefficient of 0.741, making it the optimal model.
[0049] The results of the comparative experiment are as follows:
[0050] (4) This experiment takes the oasis areas of Alar City and Aksu City in Xinjiang as the research area. Combining the quantitative inversion method of salinization type based on multi-model feature weighting and residual correction model mentioned in this application, further, according to the classification standard of saline soil type in the third national land survey, a two-dimensional spatiotemporal map of soil salinization type distribution in this area is finally generated, as shown in Figure k. From the inversion results, the salinization type in the study area is mainly divided into three categories: chloride-sulfate type saline soil, sulfate-chloride type saline soil, and sulfate type saline soil.
[0051] This application also discloses an electronic device. (See reference...) Figure 15 , Figure 15 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.
[0052] The communication bus 502 is used to enable communication between these components.
[0053] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.
[0054] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).
[0055] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the above-described quantitative inversion method for salting-type based on feature weighting and residual correction.
[0056] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.
[0057] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.
Claims
1. A salinization type quantitative inversion method based on feature weighting and residual correction, characterized in that, The method comprises the following steps: S1: collecting measured concentration data of soil salt ions in the target area; obtaining remote sensing images of the target area and extracting multi-source feature information thereof; S2: constructing a salinization type quantitative inversion data set through the measured concentration data and the multi-source feature information; S3: evaluating the correlation between the multi-source feature information in the salinization type quantitative inversion data set and the soil salt ion concentration by using a feature selection method, forming a unified feature importance weight vector through standardization and weighted fusion, and constructing a weighted feature set through the feature importance weight vector; S4: constructing an RF model and training the same through the weighted feature set to obtain a prediction residual value of a prediction result; S5: constructing a residual correction model based on the prediction residual value and training the same through the weighted feature set; S6: obtaining remote sensing image data to be measured; S7: predicting and generating ion concentration data of the target area and generating a two-dimensional spatial-temporal distribution map of quantitative classification of soil salinization types by using the trained RF model and residual correction model and the remote sensing image data to be measured.
2. The method according to claim 1, wherein, The multi-source feature information comprises a vegetation index, a salt index and multi-band reflectivity.
3. The method of quantitative inversion of salinization types based on feature weighting and residual correction according to claim 1, characterized in that, Step S2 comprises: S21: pre-processing remote sensing images of the research area; S22: extracting multi-source feature information of corresponding positions in the remote sensing images according to geographical coordinates of soil sampling points in the target area; S23: spatially matching the extracted multi-source feature information with the measured concentration data to form an image and ground observation data pair; S24: calculating various vegetation indexes and salt indexes based on the image and ground observation data pair to construct a comprehensive data set that can be used for salinization type quantitative inversion, i.e., a salinization type quantitative inversion data set.
4. The method according to claim 3, wherein, The pre-processing comprises radiation calibration and atmospheric correction.
5. The method of quantitative inversion of salinization types based on feature weighting and residual correction according to claim 1, characterized in that, Step S3 comprises: The feature selection method comprises XGBoost gain, Pearson correlation analysis and mutual information; The gain index of the XGBoost model is used to calculate the contribution of each input feature in the salinization type quantitative inversion data set to the prediction of the soil salt ion concentration in the regression model to obtain a gain score of each feature; Pearson correlation analysis is performed on the same batch of input features to calculate the linear correlation coefficient between each feature and the soil salt ion concentration, and the absolute value thereof is taken as a correlation score; The mutual information method is used to evaluate the degree of nonlinear dependence between each input feature and the soil salt ion concentration to obtain a mutual information score of each feature; The gain score, the correlation score and the mutual information are normalized respectively; The normalized three groups of scores are subjected to weighted averaging or weight fusion strategy to calculate the comprehensive importance score of each feature to obtain the fused feature score; Based on the fused feature score, a feature importance weight vector is formed to perform weighted processing on the original feature matrix to construct a weighted feature set.
6. The method of quantitative inversion of salinization types based on feature weighting and residual correction according to claim 1, characterized in that, The residual correction model adopts the XGBoost model; The RF model and the XGBoost model are trained based on a 5-fold cross-validation method.
7. An electronic device, comprising: The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-6.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the method as described in any one of claims 1-6.