A soil salinity inversion method based on multi-time scale remote sensing feature fusion

Abstract

This invention discloses a method for soil salinity inversion based on the fusion of remote sensing features at multiple time scales, comprising the following steps: A1 field sampling and soil conductivity measurement; A2 remote sensing data preprocessing and spectral index calculation; A3 environmental covariate acquisition and preprocessing; A4 determination of the optimal time scale feature combination; A5 feature variable screening and optimization; A6 soil salinity inversion model construction; fusing remote sensing features at different time scales, and constructing a soil salinity remote sensing inversion model based on the optimal feature combination using a random forest algorithm. By systematically integrating the instantaneous response features of single-period images with the features of seasonal composite images, the information of remote sensing data at different time scales is fully exploited, solving the problems of single time scale of remote sensing features, insufficient information utilization, and difficulty in coordinating multi-scale features in traditional modeling processes. This improves the model's ability to characterize the spatial distribution of soil salinity and its prediction accuracy, providing more scientific and reliable technical support for remote sensing monitoring of salinized areas.

Description

Technical Field

[0001] This invention belongs to the fields of big data, artificial intelligence, and quantitative remote sensing technology, and specifically relates to a method for soil salinity inversion based on the fusion of remote sensing features at multiple time scales. Background Technology

[0002] Soil salinization is a key type of desertification and land degradation, its formation depending on specific climatic, hydrogeological, and topographical conditions. Under the combined effects of irrational human use of land resources and the inherent vulnerability of ecosystems, large amounts of soluble salts accumulate in the soil, ultimately leading to soil salinization. High concentrations of salt in the soil directly affect the efficiency of plant roots in absorbing water and nutrients, thus inhibiting plant growth, damaging surface vegetation structure, and ultimately negatively impacting ecosystem stability and regional sustainable development. Therefore, developing efficient and reliable soil salinization monitoring technologies is of significant practical importance and application value for accurately understanding the spatiotemporal evolution of soil salinity, scientifically formulating soil salinization control strategies, and improving land resource utilization efficiency.

[0003] Current research widely utilizes optical remote sensing imagery for soil salinity monitoring. Visible, near-infrared, and short-wave infrared bands can reflect the chemical and mineral composition information related to soil salinity. Because soils with different salinity contents exhibit significantly different spectral response characteristics within specific ranges of these bands, this characteristic provides a physical basis for soil salinity remote sensing inversion and spatial mapping. Compared to traditional point-based monitoring, remote sensing monitoring methods have advantages such as wide coverage, high update frequency, and rich spectral information, making them particularly suitable for dynamic monitoring of salinization at regional scales. Currently, various medium- and high-resolution optical remote sensing images, such as Sentinel-2 and Landsat 8, have been widely used in soil salinity remote sensing inversion research.

[0004] Current research on remote sensing inversion of soil salinity mostly uses single-period remote sensing images corresponding to the time of field soil sampling. Inversion models are constructed by extracting soil salinity-related spectral features from these single-period images. However, single-scene remote sensing images are easily affected by short-term interference factors such as weather conditions (e.g., cloud cover, atmospheric scattering), surface disturbances (e.g., irrigation, rainfall), and sensor noise, resulting in insufficient stability and generalization ability of the constructed soil salinity inversion models. Multi-temporal remote sensing images, by extracting more stable temporal spectral features, can better reflect salinity characteristics. In particular, synthesizing multi-temporal remote sensing images at seasonal scales can not only effectively reflect seasonal dynamic changes in soil salinity but also reduce inversion model errors caused by occasional factors such as cloud cover and random meteorological disturbances, thus improving model reliability. However, time averaging may mask local differences, leading to insufficient prediction accuracy. Existing technologies mainly rely on feature extraction methods at a single time scale, which have the drawback of failing to balance model stability and prediction accuracy, making it difficult to meet the needs of dynamic salinization monitoring. Summary of the Invention

[0005] The purpose of this invention is to provide a soil salinity inversion method based on multi-scale remote sensing feature fusion, addressing the shortcomings of existing technologies and improving the accuracy of soil salinity remote sensing inversion. This method constructs a multi-temporal-scale feature set based on single-scene imagery and seasonal composite imagery, fully utilizing complementary information between different temporal scales. Through feature selection and optimization strategies, key spectral features are retained while redundant and noisy variables are removed. This method not only improves the model's robustness to occasional disturbances but also ensures sensitivity to local salinity differences, thereby significantly improving the accuracy of soil salinity monitoring.

[0006] The technical solution of the present invention is as follows:

[0007] A method for soil salinity inversion by fusing remote sensing features across multiple time scales includes the following steps:

[0008] A1 Field sampling and soil conductivity measurement;

[0009] A2 remote sensing data preprocessing and spectral index calculation; Sentinel-2 satellite imagery was obtained from the Copernicus open access platform as the basic data source;

[0010] A3 Environmental Covariate Acquisition and Preprocessing: Soil, climate, and topography were selected as three types of environmental factors as covariates;

[0011] The optimal time-scale feature combination for A4 was determined; the feature combination formed by fusing single-scene and seasonal composite images was determined as the optimal time-scale feature combination, and served as the basis for subsequent introduction of environmental factors and feature selection.

[0012] A5 Feature Variable Screening and Optimization: Based on the optimal combination of remote sensing features at different time scales determined in step four, environmental factors such as soil, climate, and topography are further introduced to enhance the model's response to the spatiotemporal variation of soil salinity. To optimize model performance and enhance its interpretability, a stepwise feature screening method based on SHAP values ​​is adopted. Finally, a feature subset that achieves the best balance between improving model accuracy and controlling complexity is selected, and this feature subset is used as the input variable for the final soil salinity inversion model.

[0013] Construction of A6 soil salinity inversion model;

[0014] In the method described above, step A2 involves selecting 10 bands for soil salinity inversion, including visible light B2, B3, and B4; near-infrared B8 and B8A; red-edge B5, B6, and B7; and short-wave infrared B11 and B12 regions. Images with cloud cover exceeding 20% ​​are excluded using a scene classification layer (SCL), and cloud pixels are masked. All bands are resampled to 10 bands using bilinear interpolation in Python. m; For missing areas, a time-series image interpolation strategy is used to fill in the gaps, i.e., a weighted composite of two consecutive cloudless images in the time series, to improve spatiotemporal continuity and avoid the impact of missing information on subsequent modeling; finally, single-scene images and seasonally averaged composite images corresponding to the field sampling dates are obtained; for the single-scene images and seasonally averaged composite images, salinity index and vegetation index are calculated respectively to characterize the surface spectral response; the calculated salinity index includes SI, SI1, SI2, SI3, SI4, S1, S2, S3, S4, S5, S6, S7, S8, S9, SI-T, NDSI and CRSI; the calculated vegetation index includes RVI, NDVI, ENDVI, DVI, GDVI, SAVI and EVI; the index calculated based on the single-scene image is denoted by the prefix "D_", and the index calculated based on the seasonally averaged composite image is denoted by the prefix "S_", to distinguish remote sensing features at different time scales.

[0015] In the method described above, in step A3, the soil property data is obtained from the "China Land Surface Simulation Soil Property Dataset" (CSDLv2). To obtain the soil property values ​​of the 0-20 cm soil layer, the data of the 0-5 cm, 5-15 cm, and 15-30 cm layers are first integrated using a weighted average method. Subsequently, the seven selected soil property indicators, including cation exchange capacity (CEC), organic carbon (OC), pH value (pH), bulk density (BD), sand content, silt content, and clay content, are uniformly resampled to a spatial resolution of 10 m.

[0016] In the method described above, in step A3, the climate data is sourced from the WorldClim database, and the data on maximum temperature (T_max), minimum temperature (T_min), average temperature (T_mean), precipitation (prec), solar radiation (srad), saturated vapor pressure (vapor), and wind speed (wind) are resampled to a spatial resolution of 10 m.

[0017] In the method described above, step A3 involves using a digital elevation model (DEM), which is an important topographic factor influencing the spatial pattern and dynamic changes of soil salinity. NASA's SRTM DEM data is used. ArcGIS is used to extract elevation (DEM), slope, aspect, plan curvature, profile curvature, and topographic moisture index (TWI), and the data is resampled to a resolution of 10 m.

[0018] In the method described above, step A4 involves using measured soil electrical conductivity samples as the basic data, randomly dividing them into a training set and a validation set at a ratio of 7:3. Then, based on remote sensing images, feature variables corresponding to different time scales are extracted to construct three types of random forest soil salinity inversion models. The first type extracts features from single-scene Sentinel-2 images corresponding to the sampling date and constructs a soil salinity inversion model. The second type extracts features from seasonal composite images and constructs a soil salinity inversion model. The third type integrates features extracted from single-scene images and seasonal composite images at the feature level to form a multi-time-scale feature combination, and constructs a soil salinity inversion model accordingly. To ensure comparability between models, all three types of random forest models use three types of features: band reflectance, vegetation index, and salinity index, without feature selection processing, to control the consistency of input variables and highlight the impact of time scale differences on model performance.

[0019] The specific operation flow of the stepwise feature selection method based on SHAP values ​​in step A5 of the method described above is as follows: First, a complete random forest model is constructed, containing all candidate variables including the optimal combination of remote sensing features at the optimal time scale, soil factor variables, climate factor variables, and topographic factor variables. The mean absolute SHAP value of each candidate feature variable is calculated, and they are sorted according to the importance of the mean absolute SHAP value. Subsequently, feature variables are gradually introduced into the model in descending order of importance. After each feature is added, the model is retrained, and the model performance is evaluated using the prediction accuracy index of the validation set.

[0020] In the method described above, in step A6, based on the optimal feature subset obtained in step five, a random forest regression method is used to construct an optimal soil salinity inversion model; let the input feature subset be a vector:

[0021] …, (1)

[0022] A random forest consists of N regression trees, and the prediction function of each tree is denoted as . The model's predicted value is...

[0023] (2)

[0024] Where: 𝑦 is the predicted soil salinity value (e.g., EC). 1:5 ).

[0025] The beneficial effects of this invention are: by integrating remote sensing features at different time scales, a soil salinity remote sensing inversion model based on optimal feature combination is constructed using the random forest algorithm. This method, by systematically integrating the instantaneous response features of single-period images with the features of seasonal composite images, fully mines the information from remote sensing data at different time scales. It effectively solves the problems of single-time-scale remote sensing features, insufficient information utilization, and difficulty in coordinating multi-scale features in traditional modeling processes. This improves the model's ability to characterize the spatial distribution of soil salinity and its prediction accuracy, providing more scientific and reliable technical support for remote sensing monitoring of salinized areas. Attached Figure Description

[0026] Figure 1 The flowchart shows the soil salinity inversion method based on the fusion of remote sensing features at multiple time scales.

[0027] Figure 2 To improve the accuracy of soil salinity inversion models based on remote sensing features at different time scales;

[0028] Figure 3 To determine the feature importance distribution and optimal model based on SHAP values;

[0029] Figure 4 To improve the accuracy of soil conductivity inversion, a spatial distribution map of soil conductivity is provided. Detailed Implementation

[0030] The present invention will be described in detail below with reference to specific embodiments.

[0031] refer to Figure 1 A method for soil salinity inversion by fusing remote sensing features across multiple time scales includes the following steps:

[0032] Step 1: Soil Sample Collection and Analysis

[0033] Soil field sampling was conducted in Kenli District, Dongying City, Shandong Province in October 2024. Sampling points were laid out in a 3 km × 3 km grid within the study area. The location of sampling points was optimized during the layout phase, taking into account crop type, uniformity of sample point distribution, and accessibility. A total of 214 surface (0-20 cm) soil samples were collected. Each sampling point was prepared by mixing five subsamples, and latitude, longitude, and environmental information were recorded. In the laboratory, the samples were air-dried, ground, and sieved through a 2 mm sieve. 20 g of soil sample was weighed and added to 100 mL of deionized water to prepare a soil solution at a water-to-soil ratio of 5:1. After thorough shaking and mixing, the electrical conductivity (EC) of the soil solution was measured. 1:5 (μS / m) is used as an indicator of soil salinity.

[0034] Step 2: Remote sensing data preprocessing and spectral index calculation

[0035] Sentinel-2 satellite imagery was obtained from the Copernicus Open Access Platform as the primary data source. This satellite has 13 spectral bands, covering the visible, near-infrared, red-edge, and short-wave infrared spectral regions, offering high spatiotemporal resolution. It has a revisit period of 5 days and spatial resolutions of 10 m (B2, B3, B4, B8), 20 m (B5, B6, B7, B8A, B11, B12), and 60 m (B1, B9, B10). Ten bands were selected for soil salinity retrieval, including the visible (B2, B3, and B4), near-infrared (B8 and B8A), red-edge (B5, B6, and B7), and short-wave infrared (B11 and B12) regions. Scene classification layers (SCL) were used to exclude images with cloud cover exceeding 20%, and cloud pixels were masked. To ensure consistency, all bands were resampled to 10 m using bilinear interpolation in Python. For missing areas, a time-series image interpolation strategy is used for filling in the gaps, i.e., a weighted composite of two consecutive cloud-free images in the time series, to improve spatiotemporal continuity and avoid the impact of missing information on subsequent modeling. Finally, single-scene images and seasonally averaged composite images corresponding to the field sampling dates are obtained. For the single-scene images and seasonally averaged composite images, salinity indices and vegetation indices are calculated to characterize the surface spectral response. The calculated salinity indices include SI, SI1, SI2, SI3, SI4, S1, S2, S3, S4, S5, S6, S7, S8, S9, SI-T, NDSI, and CRSI; the calculated vegetation indices include RVI, NDVI, ENDVI, DVI, GDVI, SAVI, and EVI. Indices calculated based on single-scene images are denoted with the prefix "D_" (e.g., D_NDVI), and indices calculated based on seasonally averaged composite images are denoted with the prefix "S_" (e.g., S_NDVI) to distinguish remote sensing features at different time scales.

[0036] Step 3: Environmental covariate acquisition and preprocessing

[0037] To characterize the spatial differences in soil salinity, this embodiment selects three environmental factors—soil, climate, and topography—as covariates. Soil attribute data are derived from the "China Land Surface Simulation Soil Properties Dataset" (CSDLv2), which provides various key soil parameters at stratification depths of 0-5 cm, 5-15 cm, 15-30 cm, and 30-60 cm, with a spatial resolution of 90 m. To obtain soil attribute values ​​for the 0-20 cm soil layer, a weighted average method was first used to integrate data from the 0-5 cm, 5-15 cm, and 15-30 cm layers. Subsequently, seven selected soil attribute indicators—cation exchange capacity (CEC), organic carbon (OC), pH value, bulk density (BD), sand content, silt content, and clay content—were uniformly resampled to a spatial resolution of 10 m. Climate data were sourced from the WorldClim database (https: / / worldclim.org / data / index.html), which provides long-term monthly climate data from 1970 to 2000 with a spatial resolution of 1 km. Maximum temperature (T_max), minimum temperature (T_min), mean temperature (T_mean), precipitation (prec), solar radiation (srad), saturated vapor pressure (vapor), and wind speed data were resampled to a 10 m spatial resolution. The Digital Elevation Model (DEM) is an important topographic factor influencing the spatial pattern and dynamic changes of soil salinity. NASA's SRTM DEM data (30 m resolution) was used. Elevation (DEM), slope, aspect, plan curvature, profile curvature, and topographic moisture index (TWI) were extracted using ArcGIS and resampled to a 10 m resolution. These climate, soil, and topographic factors were used as environmental covariates for subsequent soil salinity inversion modeling.

[0038] Step 4: Determining the optimal combination of time-scale features

[0039] To evaluate the impact of remote sensing features at different time scales on the accuracy of soil salinity retrieval, 214 measured soil electrical conductivity samples were used as the basic data and randomly divided into a training set (150 samples) and a validation set (64 samples) in a 7:3 ratio. Subsequently, feature variables (i.e., band reflectance, salinity index, and vegetation index) corresponding to different time scales were extracted from remote sensing images to construct three types of random forest soil salinity retrieval models. The first type extracted features from a single Sentinel-2 image corresponding to the sampling date and constructed a soil salinity retrieval model. The second type extracted features from seasonal composite images and constructed a soil salinity retrieval model. The third type fused features extracted from single-scene images and seasonal composite images at the feature level to form a multi-time-scale feature combination, and constructed a soil salinity retrieval model accordingly. To ensure comparability between models, all three types of random forest models used band reflectance, vegetation index, and salinity index features without feature selection processing to control the consistency of input variables and highlight the impact of time scale differences on model performance. By comparing the prediction accuracy on the validation set, the results show that the third type of model performs better in R... 2 In terms of both RMSE and other metrics, it outperforms the previous two types of models. Therefore, the feature combination formed by fusing single-scene and seasonal composite images is determined as the optimal time-scale feature combination, and serves as the basis for introducing environmental factors and feature selection in the subsequent process.

[0040] Step 5: Feature Variable Selection and Optimization

[0041] Based on the optimal time-scale remote sensing feature combination determined in step four, environmental factor variables such as soil, climate, and topography are further introduced to enhance the model's response to the spatiotemporal variability of soil salinity. Furthermore, to optimize model performance and enhance its interpretability, a stepwise feature selection method based on SHAP (SHapley Additive exPlanations) values ​​is adopted. The specific operation process is as follows: First, a complete random forest model containing all candidate variables including the optimal time-scale remote sensing feature combination, soil factor variables, climate factor variables, and topographic factor variables is constructed. The mean absolute SHAP value of each candidate feature variable is calculated and ranked according to the importance of the mean absolute SHAP value. Subsequently, feature variables are gradually introduced into the model in descending order of importance. After each feature is added, the model is retrained, and the model performance is evaluated using the prediction accuracy metrics (R² and RMSE) of the validation set. Finally, a feature subset that achieves the best balance between improving model accuracy and controlling complexity is selected, and this feature subset is used as the input variable for the final soil salinity inversion model.

[0042] In this embodiment, the stepwise feature selection method based on SHAP values ​​is used, and the final optimal feature subset includes: single-scene image features (such as D_B12, D_B6, D_ENDVI, D_S6, D_B11); seasonal composite image features (such as S_NDVI, S_DVI, S_CRSI, S_B2); and environmental factors (such as srad, wind, silt, BD). It should be noted that the specific combination of the optimal feature subset may differ depending on the study area and data conditions, and this method is not limited to the above example.

[0043] Step Six: Construction of Soil Salinity Inversion Model

[0044] Based on the optimal feature subset obtained in step five, the optimal soil salinity inversion model is constructed using the random forest regression method. Let the input feature subset be a vector:

[0045] …, (1)

[0046] A random forest consists of N regression trees, and the prediction function of each tree is denoted as . The model's predicted value is...

[0047] (2)

[0048] Where: 𝑦 is the predicted soil salinity value (e.g., EC). 1:5 ).

[0049] The model was trained on the training set and its accuracy was evaluated on the validation set to verify its stability and reliability. Based on the trained model, pixel-by-pixel predictions were made for the study area to generate a spatial distribution map of soil salinity.

[0050] Figure 2 The accuracy of soil salinity inversion models under remote sensing features at different time scales is shown. The model constructed using single-scene image features and the model constructed using seasonal composite image features show little difference in accuracy, both at a relatively low level. The model constructed based on the fusion of single-scene and seasonal features shows a significantly improved fit and a marked reduction in error, indicating that introducing multi-time-scale features can effectively enhance the reliability and applicability of the model. Figure 3 The distribution of feature importance in SHAP and the soil conductivity inversion accuracy of the optimal model are presented. The results show that the selected optimal features can effectively characterize the spectral response mechanism of soil salinity, with a validation set RMSE of 555.88 μS / cm and R0.05. 2 With a score of 0.66, the model demonstrates good interpretability and generalization ability. Figure 4This is a spatial distribution map of soil electrical conductivity derived from the optimal model. As shown in the figure, soil electrical conductivity exhibits significant spatial heterogeneity, generally decreasing from northeast to southwest. This result is largely consistent with the regional water and salt transport patterns, further demonstrating the reliability and application value of the method presented in this invention.

[0051] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for soil salinity inversion by fusing remote sensing features at multiple time scales, characterized in that, Includes the following steps: A1 Field sampling and soil conductivity measurement; A2 remote sensing data preprocessing and spectral index calculation; Sentinel-2 satellite imagery was obtained from the Copernicus open access platform as the basic data source; A3 Environmental Covariate Acquisition and Preprocessing: Soil, climate, and topography were selected as three types of environmental factors as covariates; A4 optimal time scale feature combination determined; Based on measured soil electrical conductivity samples, the data was randomly divided into training and validation sets in a 7:3 ratio. Subsequently, feature variables corresponding to different time scales were extracted from remote sensing images to construct three types of random forest soil salinity inversion models. The first type extracted features from single-scene Sentinel-2 images corresponding to the sampling date and constructed a soil salinity inversion model. The second type extracted features from seasonal composite images and constructed a soil salinity inversion model. The third type fused features extracted from single-scene and seasonal composite images at the feature level to form a multi-time-scale feature combination, and constructed a soil salinity inversion model accordingly. To ensure comparability between models, all three types of random forest models used three types of features: band reflectance, vegetation index, and salinity index, without feature selection processing, to control the consistency of input variables and highlight the impact of time scale differences on model performance. By comparing the prediction accuracy on the validation set, the results showed that the third type of model achieved higher R... 2 It outperforms the previous two models in terms of RMSE index. The feature combination formed by fusing single scene and seasonal composite images is determined as the optimal time scale feature combination, and serves as the basic variable set for subsequent introduction of environmental factors and feature selection. A5 Feature Variable Screening and Optimization: Based on the optimal combination of remote sensing features at different time scales determined in step A4, soil, climate, and topographic environmental factors are further introduced to enhance the model's response to the spatiotemporal variability of soil salinity. To optimize model performance and enhance its interpretability, a stepwise feature screening method based on SHAP values ​​is adopted. Finally, a feature subset that achieves the best balance between improving model accuracy and controlling complexity is selected and used as the input variable for the final soil salinity inversion model. The specific operation process of the stepwise feature selection method based on SHAP values ​​is as follows: First, a complete random forest model is constructed, containing all candidate variables including the optimal combination of remote sensing features at the optimal time scale, soil factor variables, climate factor variables, and topographic factor variables. The mean absolute SHAP value of each candidate feature variable is calculated and sorted according to the importance of the mean absolute SHAP value. Then, feature variables are gradually introduced into the model in descending order of importance. After each feature is added, the model is retrained, and the model performance is evaluated using the prediction accuracy index of the validation set. A6 Soil salinity inversion model construction; Based on the optimal feature subset obtained in step A5, the optimal soil salinity inversion model is constructed using the random forest regression method; Let the input feature subset be a vector: (1) A random forest consists of N regression trees, and the prediction function of each tree is denoted as . The model prediction value is （2） in: This represents the predicted soil salinity value.

2. The method according to claim 1, characterized in that, In step A2, 10 bands are selected for soil salinity inversion, including visible light B2, B3 and B4, near-infrared B8 and B8A, red edge B5, B6 and B7 and short-wave infrared B11 and B12 regions. Images with cloud coverage exceeding 20% ​​were excluded using a scene classification layer, and cloud pixels were masked. All bands were resampled to 10 using bilinear interpolation in Python. m; For missing areas, a time-series image interpolation strategy is used to fill in the gaps, i.e., a weighted composite of two consecutive cloudless images in the time series, to improve spatiotemporal continuity and avoid the impact of missing information on subsequent modeling; finally, single-scene images and seasonally averaged composite images corresponding to the field sampling dates are obtained; for the single-scene images and seasonally averaged composite images, salinity index and vegetation index are calculated respectively to characterize the surface spectral response; the calculated salinity index includes SI, SI1, SI2, SI3, SI4, S1, S2, S3, S4, S5, S6, S7, S8, S9, SI-T, NDSI and CRSI; the calculated vegetation index includes RVI, NDVI, ENDVI, DVI, GDVI, SAVI and EVI; the index calculated based on the single-scene image is denoted by the prefix "D_", and the index calculated based on the seasonally averaged composite image is denoted by the prefix "S_", to distinguish remote sensing features at different time scales.

3. The method according to claim 1, characterized in that, In step A3, the soil property data comes from the "Soil Property Dataset for Simulating Land Surface in China" CSDLv2. To obtain the soil property values ​​of the 0-20 cm soil layer, the data of the 0-5 cm, 5-15 cm and 15-30 cm layers are first integrated using a weighted average method. Then, the seven selected soil property indicators, including cation exchange capacity, organic carbon, pH value, bulk density, sand content, silt content and clay content, are uniformly resampled to a spatial resolution of 10 m.

4. The method according to claim 1, characterized in that, In step A3, the climate data is sourced from the WorldClim database, and the data on maximum temperature, minimum temperature, average temperature, precipitation, solar radiation, saturated vapor pressure, and wind speed are resampled to a spatial resolution of 10 m.

5. The method according to claim 1, characterized in that, In step A3, the digital elevation model is an important topographic factor affecting the spatial pattern and dynamic changes of soil salinity. NASA's SRTM DEM data is used. ArcGIS is used to extract elevation, slope, aspect, plane curvature, profile curvature and topographic humidity index, and resampled to 10 m resolution.

Images