Method for estimating soil salt content at regional scale of slope surface based on deep learning

Abstract

The present application relates to the technical field of digital soil mapping, in particular to a slope regional scale soil salt content estimation method based on deep learning; first, multispectral data of a target region is obtained by a UAV to construct multi-modal features such as spectral features, spatial texture features and terrain hydrological features; then, a slope salt accumulation prior index is constructed according to the laws of slope water salt migration and salt low-lying enrichment; thereafter, the key spectral features, spatial texture features and terrain hydrological features, and the slope salt accumulation prior index are input into a gated multi-branch fusion regression model to generate a soil salt spatial distribution map of the target region through the gated multi-branch fusion regression model; compared with the traditional estimation scheme, the present scheme fuses the spectral features, spatial texture features and terrain hydrological features, and generates the corresponding soil salt spatial distribution map of the target region through the gated multi-branch fusion regression model, greatly improving the accuracy of soil salt estimation.

Description

Technical Field

[0001] This invention relates to the field of digital soil mapping technology, specifically to a method for estimating soil salinity at the slope regional scale based on deep learning. Background Technology

[0002] Soil salinization is a common land degradation problem in arid and semi-arid regions, as well as in lakeside areas, river valleys, and irrigated areas. Influenced by factors such as strong evaporation, variations in groundwater depth, slope runoff, poor drainage, and human activities such as irrigation and drainage, salts often accumulate spatially at the toe of slopes, in low-lying runoff areas, along gully banks, lakeshores, and on exposed surfaces. Accurately obtaining the spatial distribution of soil salinity at the slope, small watershed, and regional scales is of great significance for saline-alkali land management, the design of salt drainage projects, the improvement of arable land quality, and ecological restoration.

[0003] Currently, obtaining the spatial distribution of soil salinity is mostly achieved through manual sampling or satellite remote sensing. Manual sampling and laboratory testing offer high accuracy, but the number of sampling points is limited, costs are high, and spatial continuity is insufficient, making it difficult to quickly reflect localized salt accumulation under slope micro-topographic conditions. While satellite remote sensing can achieve regional monitoring, it is limited by data spatial resolution and satellite revisit time, making it difficult to identify small-scale differences such as gully edges, slope toe depressions, and lakeside salt crust patches. UAV remote sensing, on the other hand, offers advantages such as high spatial resolution, high mobility, and flexible data acquisition. It can simultaneously acquire visible light, multispectral, and oblique photography data, thus providing a new technical means for precise estimation of soil salinity at the slope scale. However, the accuracy of existing UAV soil salinity estimation methods is insufficient. Summary of the Invention

[0004] The main objective of this invention is to provide a deep learning-based method for estimating soil salinity at the slope regional scale, aiming to address the inaccuracy of existing UAV-based soil salinity estimation methods.

[0005] The technical solution proposed in this invention is as follows: A deep learning-based method for estimating soil salinity at the slope regional scale includes: Multimodal data of the target area is acquired by drones, and the multimodal data is preprocessed to obtain a multimodal feature dataset. Slope layered soil sampling points were set up in the target area, and soil salinity labels were constructed. Feature variables related to the spatial distribution of soil salinity were extracted from the multimodal feature dataset. These feature variables included spectral features, spatial texture features, and topographic and hydrological features. Construct a priori index of slope salinity accumulation in the target area; Optimize the selection of feature variables based on salt content labels; A gated multi-branch fusion regression model based on the prior index of slope salt accumulation was constructed, and the gated multi-branch fusion regression model was trained and its accuracy was evaluated. Based on the gated multi-branch fusion regression model that has completed training and accuracy evaluation, and the optimized and screened feature variables, a spatial distribution map of soil salinity corresponding to the target area is generated.

[0006] Preferably, the UAV is equipped with a spectral sensor; the multimodal feature dataset includes multispectral imagery, a digital surface model, a digital elevation model, and a digital terrain model; the process of acquiring multimodal data for the target area via the UAV and preprocessing the multimodal data to obtain the multimodal feature dataset includes: The target area is acquired by using a drone, wherein the multispectral image includes two bands: red light and near-infrared light. A digital surface model of the target area is generated by reconstructing multispectral images; Non-ground point removal, vegetation / building masking, filtering, and smoothing are performed on the digital surface model to generate a digital elevation model or digital terrain model for hydrological analysis. Percentile normalization of multispectral images: (1) In the formula, These are the standardized eigenvalues; These are the original pixel feature values ​​of the multispectral image; This represents the 2nd percentile value of this feature within the target region. This represents the 98th percentile value of the feature within the target region. Then let ;like Then let .

[0007] Preferably, the step of setting up slope-layered soil sampling points in the target area and constructing soil salinity labels includes: Inversion is performed based on the digital surface model of the target area to obtain the slope, relative slope position, runoff area, distance from the gully or lake shore, bare soil patches, apparent brightness of salt crust, and vegetation sparseness of the target area. The target area is then divided into zones, including the top of the slope, the middle of the slope, the foot of the slope, the low-lying runoff area, the gully bank area, the lakeside area, and the relatively flat background area. Sampling points were evenly distributed within each zone. A rectangular quadrat was set up at each sampling point. Soil samples were collected at five locations, including the four corners and the center of the quadrat, and the real-time dynamic differential positioning coordinates of the sampling points were recorded. Soil samples were taken from different depths within quadrats to determine the salinity content of each soil layer and to construct a weighted salinity label. (2) In the formula, Y represents the soil salinity label; This refers to the soil salinity at a depth of 0-2cm. This refers to the soil salinity at a depth of 2-5 cm. This refers to the soil salinity at a depth of 5-10cm.

[0008] Preferably, the extraction of feature variables related to the spatial distribution of soil salinity from the multimodal feature dataset includes: Spectral features are extracted from multispectral images in a multimodal feature dataset, where spectral features are used to express the effects of salinity, soil background, and vegetation cover on surface reflectance. Normalized difference salinity index obtained based on spectral features: (3) Soil brightness index obtained based on spectral characteristics: (4) Normalized vegetation index obtained based on spectral features: (5) In the formula, The reflectance in the red band of the spectrum. This represents the near-infrared reflectance in the spectral characteristics.

[0009] Preferably, the step of obtaining the normalized vegetation index based on spectral features further includes: Based on the digital surface model of the target area, topographic and hydrological features at the slope scale of the target area are extracted. The topographic and hydrological features are divided into geomorphic indicators and hydrological indicators. The geomorphic indicators include: mountain shadow, slope, aspect, plane curvature, profile curvature, convergence index, topographic humidity index and relative slope position. The hydrological indicators include: closed depressions, total catchment area, LS factor, gully network reference surface, gully network distance and valley depth. The formula for calculating slope is: (6) In the formula, The slope is mentioned. The elevation change rate of the central pixel in the horizontal direction. The elevation change rate of the central pixel in the vertical direction; The formula for calculating the topographic humidity index is: (7) In the formula, TWI is the terrain humidity index; The upslope runoff area per unit contour line length; Slope; Based on multispectral imagery, the spatial texture features of the target region are extracted using the gray-level co-occurrence matrix method. The spatial texture features include statistical texture features, contrastive texture features, and ordered texture features. Statistical texture features include mean, variance, and correlation. Contrastive texture features include contrast, dissimilarity, and homogeneity. Ordered texture features include second moment and entropy.

[0010] Preferably, the formula for calculating the prior index of slope salt accumulation is: (8) In the formula, The prior index for salt accumulation on the slope; This is the percentile standardization function; The topographic humidity index; The valley is deep; The catchment area; Relative slope position; This serves as the reference surface for the trench. Slope; Distance from the channel; , , , , , and All are weighting coefficients.

[0011] Preferably, the optimized screening of feature variables based on salt labels includes: Feature table construction: Centered on the real-time dynamic differential positioning coordinates of soil sampling points in the target area, extract the effective pixel feature values ​​within the rectangular quadrat area, and take the average value of the effective pixels within the same rectangular quadrat as the feature value corresponding to that soil sampling point. (9) In the formula, For the first The first soil sampling point One eigenvalue; For the first The set of effective pixels within a rectangular quadrat of a soil sampling point; Number of effective pixels; For the first Features in pixels The value at that location; No. Initial feature vector of each soil sampling point Represented as: (10) In the formula, For the first spectral characteristics of each soil sampling point For the first Spatial texture features of each soil sampling point For the first Topographic and hydrological characteristics of each soil sampling point; Calculate each eigenvalue The Pearson correlation coefficient between soil salinity and the label, where the Pearson correlation coefficient is used to determine the eigenvalue. The response strength to changes in soil salinity, and the formula for calculating the Pearson correlation coefficient is as follows: (11) In the formula, For the first Pearson correlation coefficients between eigenvalues ​​and soil salinity labels; For the first The first soil sampling point One eigenvalue; For the first The average value of each feature across all soil sampling points; For the first Soil salinity labels at each soil sampling point; This represents the average soil salinity label from all soil sampling points; Let i be the total number of soil sampling points, 1 ≤ i ≤ n; Within the three modes of spectral features, spatial texture features, and topographic and hydrological features, respectively, according to Sort the features from largest to smallest, and in each modality, input the feature values ​​that strongly respond to the soil salinity label into the next step of filtering, and remove the feature values ​​that weakly respond to the soil salinity label from the feature set.

[0012] Preferably, within the three modalities of spectral features, spatial texture features, and topographic and hydrological features, respectively, according to... Sort the features from largest to smallest, and within each modality, input the feature values ​​that strongly respond to the soil salinity label into the next filtering step, while removing the feature values ​​that weakly respond to the soil salinity label from the feature set. This process also includes: Calculate the variance inflation factor (VIF) for each retained feature: (12) In the formula, For the first Variance inflation factor for each eigenvalue; For the first The coefficient of determination is obtained by performing linear regression with one feature as the dependent variable and the remaining features within the same mode as the independent variables. when At that time, delete the first one. One eigenvalue, and recalculate the remaining eigenvalues. Until all retained features within each modality satisfy ; A random forest regression model is used to repeatedly train each feature, and the SHAP contribution value of each feature in each training iteration is calculated. The SHAP contribution value of the feature in each training iteration is then calculated. In the training session, the first The average absolute SHAP contribution of each feature is: (13) In the formula, For the first In the training session, the first The average absolute SHAP contribution value of each feature; For the first In the training session, the first The feature is related to the first The SHAP contribution value of the predicted values ​​from the soil sampling points, the ... The SHAP stability contribution of each feature is defined as follows: (14) In the formula, For the first The SHAP stability contribution of each feature; For the first The average SHAP contribution value of each feature across multiple training iterations; For the first The standard deviation of the SHAP contribution value of each feature in multiple training sessions; the stable SHAP contribution value is used to view the degree of stable response of each feature variable to changes in soil salinity. The remaining spectral features, spatial texture features, and topographic and hydrological features are retained and used as the final key multimodal feature set: (15) In the formula, For key multimodal feature sets; This indicates the spectral features retained after screening; This indicates the spatial texture features retained after filtering; This indicates the topographic and hydrological features that are retained after screening.

[0013] Preferably, the gated multi-branch fusion regression model includes a spectral coding branch, a texture coding branch, a topographic and hydrological coding branch, a gated weight allocation layer, and a salinity regression output layer; the construction of a gated multi-branch fusion regression model based on a slope salinity accumulation prior index, and the training and accuracy evaluation of the gated multi-branch fusion regression model, include: The multi-branch encoding of the gated multi-branch fusion regression model is expressed as follows: (16) In the formula, It is a spectral encoder and a multilayer perceptron; It is a spatial texture encoder and a multilayer perceptron; This is a topographic and hydrological encoder used to learn the nonlinear relationship between topographic and hydrological features and salinity accumulation. These are implicit features of spectral modes. For spatial texture modal latent features, These are implicit features of topographic and hydrological modes; Construct a gating weight allocation layer: (17) In the formula, For adaptive weights of spectral modes, For adaptive weights of spatial texture modality, Adaptive weights for topographic and hydrological modes; and All of these are trainable parameters; These are the latent features of the fused pixels; The fused latent features are input into the salinity regression output layer to obtain an estimate of soil salinity content: (18) In the formula, For the first Estimated soil salinity at each pixel or soil sampling point location; and All are parameters from the salt regression output layer; The function is used to ensure that the output value is non-negative; Introducing slope prior constraint loss The expression is as follows: (19) In the formula, For pixels The estimated value of soil salinity, For pixels Estimated soil salinity; For pixels Prior indices for slope salt accumulation. For pixels Prior indices for slope salt accumulation; The comprehensive loss function of the gated multi-branch fusion regression model Set as: (20) In the formula, Loss of estimated soil salinity; For slope prior constraint loss; These are the weighting coefficients.

[0014] Preferably, the generation of the soil salinity spatial distribution map corresponding to the target area based on the gated multi-branch fusion regression model that has completed training and accuracy evaluation, and the optimized and screened feature variables includes: For each valid soil pixel within the target area, determine the key multimodal feature set to extract the corresponding model input features: (25) In the formula, For the first The model input features for each effective soil pixel; For the first Key spectral features of each effective soil pixel For the first Key spatial texture features of an effective soil pixel For the first Key topographic and hydrological features of each effective soil pixel; For the first Prior index of slope salt accumulation for each effective soil cell; By inputting the model features of effective soil pixels into the trained gated multi-branch fusion regression model, the estimated soil salinity content of the effective soil pixels is obtained: (26) In the formula, For the first Estimated soil salinity content for each effective soil cell; The completed soil salinity estimation model; Set up training The first gated multi-branch fusion regression model, the... The first gated multi-branch fusion regression model for the first The predicted value of each effective soil cell is The final soil salinity mapping value is: (26) In the formula, For the first The final soil salinity mapping value for each effective soil pixel; The uncertainty in the graph is represented by the standard deviation of multiple prediction results: (Equation 1) In the formula, For the first Mapping uncertainty of each effective soil pixel; The soil salinity content of all effective soil pixels in the target area is estimated and the mapping uncertainty is expressed to output a spatial distribution map of soil salinity in the target area. The spatial distribution map of soil salinity includes a continuous distribution map of soil salinity, a soil salinity uncertainty map, and a mask map of effective soil pixels.

[0015] The above technical solution can achieve the following beneficial effects: The proposed deep learning-based method for estimating soil salinity at the slope regional scale addresses the inaccuracy of existing UAV-based soil salinity estimation methods. The core steps of this application are as follows: First, multispectral data of the target area is acquired using a UAV to construct multimodal features, including spectral characteristics, spatial texture features, and topographic and hydrological features. Then, a priori index for slope salinity accumulation is constructed based on slope water and salt migration and salinity enrichment patterns in low-lying areas. Next, key spectral features, spatial texture features, topographic and hydrological features, and the priori index for slope salinity accumulation are input into a gated multi-branch fusion regression model to generate a spatial distribution map of soil salinity in the target area. Compared to traditional estimation schemes, this method integrates spectral features, spatial texture features, and topographic and hydrological features, and generates a spatial distribution map of soil salinity in the target area through a gated multi-branch fusion regression model, significantly improving the accuracy of soil salinity estimation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the first embodiment of a deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0019] This invention proposes a method for estimating soil salinity at the regional scale of slopes based on deep learning.

[0020] As attached Figure 1 As shown, in the first embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, the embodiment includes the following steps: Step S110: Acquire multimodal data for the target area using a drone, and preprocess the multimodal data to obtain a multimodal feature dataset.

[0021] Step S120: Set up slope layered soil sampling points in the target area and construct soil salinity labels.

[0022] Step S130: Extract feature variables related to the spatial distribution of soil salinity from the multimodal feature dataset, wherein the feature variables include spectral features, spatial texture features and topographic and hydrological features.

[0023] Step S140: Construct the Prior Index of Slope Salt Accumulation (SAI) for the target area.

[0024] Step S150: Optimize and filter feature variables based on salt content labels.

[0025] Step S160: Construct a gated multi-branch fusion regression model based on the prior index of slope salt accumulation, and train and evaluate the accuracy of the gated multi-branch fusion regression model.

[0026] Step S170: Based on the gated multi-branch fusion regression model that has completed training and accuracy evaluation, and the optimized and screened feature variables, generate a spatial distribution map of soil salinity corresponding to the target area.

[0027] The proposed deep learning-based method for estimating soil salinity at the slope regional scale addresses the inaccuracy of existing UAV-based soil salinity estimation methods. The core steps of this application are as follows: First, multispectral data of the target area is acquired using a UAV to construct multimodal features, including spectral characteristics, spatial texture features, and topographic and hydrological features. Then, a priori index for slope salinity accumulation is constructed based on slope water and salt migration and salinity enrichment patterns in low-lying areas. Next, key spectral features, spatial texture features, topographic and hydrological features, and the priori index for slope salinity accumulation are input into a gated multi-branch fusion regression model to generate a spatial distribution map of soil salinity in the target area. Compared to traditional estimation schemes, this method integrates spectral features, spatial texture features, and topographic and hydrological features, and generates a spatial distribution map of soil salinity in the target area through a gated multi-branch fusion regression model, significantly improving the accuracy of soil salinity estimation.

[0028] In the second embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the first embodiment, the UAV is equipped with a spectral sensor; the multimodal feature dataset includes multispectral imagery, a digital surface model, a digital elevation model, and a digital terrain model; step S110 includes the following steps: Step S210: Acquire multispectral images of the target area using a drone, wherein the multispectral images include two bands: red (R) and near-infrared (NIR).

[0029] Specifically, the five-band multispectral images can be used to extract spectral features such as salinity index, soil index, and vegetation index.

[0030] Step S220: Reconstruct a digital surface model (DSM) of the target area using multispectral images.

[0031] Step S230: Perform non-ground point removal, vegetation / building masking, filtering, and smoothing on the digital surface model to generate a digital elevation model (DEM) or digital terrain model (DTM) for hydrological analysis.

[0032] Specifically, the UAV flight path is planned according to the study area and the distribution of sample points, and image acquisition and soil sampling are carried out simultaneously. After the acquisition is completed, five-band multispectral images and corresponding positioning data are obtained, which together constitute the raw data basis of UAV multimodal imaging for this method.

[0033] Specifically, data preprocessing is used to convert raw UAV multispectral imagery into data that can be used for feature extraction and model building. After image acquisition, DJI Terra software is used to post-process and radiometrically correct the UAV multispectral imagery, generating five-band multispectral image data, and the two-dimensional multispectral reconstruction function is used to generate a digital surface model (DSM). Step S240: Perform percentile normalization on the multispectral image: (1) In the formula, These are the standardized eigenvalues; These are the original pixel feature values ​​of the multispectral image; This represents the 2nd percentile value of this feature within the target region. This represents the 98th percentile value of the feature within the target region. Then let ;like Then let .

[0034] Specifically, percentile standardization reduces the impact of different feature dimensions and extreme values ​​on model training. After the above preprocessing, five-band multispectral imagery, DSM and its derived topographic and hydrological layers, texture feature layers, and standardized multimodal feature datasets are obtained, providing a unified data foundation for subsequent soil sample feature extraction, model training, and soil salinity spatial mapping.

[0035] In the third embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the first embodiment, step S120 includes the following steps: Step S310: Invert the digital surface model of the target area to obtain the slope, relative slope position, runoff area, distance from the ditch or lake shore, bare soil patches, apparent brightness of salt crust, and vegetation sparseness of the target area, thereby dividing the target area into zones. The zones of the target area include the top of the slope, the middle of the slope, the foot of the slope, the low-lying runoff area, the ditch bank area, the lake shore area, and the relatively flat background area.

[0036] Specifically, soil samples are used to establish model training labels for UAV remote sensing features, and the samples are arranged in layers according to the migration of water and salt on the slope and the spatial heterogeneity of salinity.

[0037] Step S320: Sampling points are set up in each zone using a combination of uniform and representative layout. A rectangular quadrat (e.g., 1m×1m) is set up for each sampling point. Soil samples are collected at five locations, including the four corners and the center of the quadrat, and the real-time dynamic differential positioning coordinates (Real-Time Kinematic, RTK) of the sampling points are recorded.

[0038] Step S330: Soil samples were taken from different depths in the quadrat to determine the salinity content of each soil layer, and a weighted salinity label was constructed based on the response intensity of the shallow surface layer by UAV remote sensing. (2) In the formula, Y is the soil salinity label used for model training; This refers to the soil salinity at a depth of 0-2cm. This refers to the soil salinity at a depth of 2-5 cm. This refers to the soil salinity at a depth of 5-10cm.

[0039] Specifically, this method sets the surface soil sampling depth to 0-10 cm to enhance the expression of vertical migration of salt on slopes and the shallow response of the topsoil layer in saline-alkali land management. Furthermore, based on the 0-10 cm soil sample, the soil sample is further divided into three soil layers: 0-2 cm, 2-5 cm, and 5-10 cm.

[0040] In the fourth embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the first embodiment, step S130 includes the following steps: Step S410: Extract spectral features from the multispectral images in the multimodal feature dataset, where the spectral features are used to express the effects of salinity, soil background, and vegetation cover on surface reflectance.

[0041] Step S420: Obtain the normalized difference salinity index based on spectral features: (3) Step S430: Obtain soil brightness index based on spectral characteristics: (4) Step S440: Obtain the normalized vegetation index based on spectral features: (5) In the formula, The reflectance in the red band of the spectrum. This represents the near-infrared reflectance in the spectral characteristics.

[0042] Specifically, the constructed spectral features include three categories: Normalized Difference Salinity Index (NDSI), Soil Brightness Index, and Normalized Difference Vegetation Index (NLI). The salinity indices include: Normalized Difference Salinity Index (NDSI), Salinity Index (SI), Salinity Index 1 (SI1), Salinity Index 2 (SI2), Salinity Index 3 (SI3), Salinity Index I (S1), Salinity Index II (S2), Salinity Index III (S3), Salinity Index V (S5), and Salinity Index VI (S6). Soil indices include: Color Index (CI), Carbonate Index (CAEX), and Brightness Index (BI). Vegetation indices include: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Nonlinear Vegetation Index (NLI).

[0043] In the fifth embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the fourth embodiment, step S440 is followed by the following steps: Step S510: Extract the slope-scale topographic and hydrological features of the target area based on the digital surface model of the target area. The topographic and hydrological features are divided into geomorphic indicators and hydrological indicators. The geomorphic indicators include: analysis of mountain shadow (AH), slope, aspect, plan curvature, profile curvature, convergence index (CI), topographic humidity index (TWI), and relative slope position (RSP). The hydrological indicators include: closed depression (CD), total catchment area (TCA), LS factor (LSF), channel network reference surface (CNBL), channel network distance (CND), and valley depth (VD).

[0044] The formula for calculating slope is: (6) In the formula, The slope is mentioned. The elevation change rate of the central pixel in the horizontal direction. The elevation change rate of the central pixel in the vertical direction.

[0045] The formula for calculating the topographic humidity index is: (7) In the formula, TWI is the terrain humidity index; The upslope runoff area per unit contour line length; For slope, The larger the value, the easier it is for the pixel to accumulate water and redistribute salt.

[0046] Step S520: Based on multispectral imagery, extract spatial texture features of the target region using the Gray-Level Co-occurrence Matrix (GLCM) method. The spatial texture features include statistical texture features, contrastive texture features, and ordered texture features. Statistical texture features include mean (ME), variance (VA), and correlation (COR). Contrastive texture features include contrast (CO), dissimilarity (DS), and homogeneity (HO). Ordered texture features include angular second moment (ASM) and entropy (EN).

[0047] Specifically, spatial texture features are used to characterize the spatial distribution of grayscale on the surface of saline soil and its local structural differences. They can reflect the spatial texture information formed by salt crust, bare soil particles, surface roughness, and salt patch boundaries. GLCM quantifies the spatial dependence of image grayscale by statistically analyzing the frequency of pixel grayscale value combinations under specific spatial relationships.

[0048] In the sixth embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the first embodiment, the calculation formula for the slope salinity accumulation prior index is as follows: (8) In the formula, The prior index for salt accumulation on the slope; This is the percentile standardization function; The topographic humidity index; The valley is deep; The catchment area; Relative slope position; This serves as the reference surface for the trench. Slope; Distance from the channel; , , , , , and All are weighting coefficients, and satisfy the following: ρ1+ρ2+ρ3+ρ4+ρ5+ρ6+ρ7=1, In the formula, , , , , , , The higher the SAI value, the higher the potential for salt accumulation on the slope.

[0049] Specifically, the slope salt accumulation prior index (SAI) is a topographic prior variable composed of topographic moisture, depression degree, runoff capacity, slope position, gully reference level, slope gradient, and distance from the gully. SAI does not represent the measured soil salinity content, but rather the relative potential for salt accumulation in a given pixel under the same climate, parent material, and land cover conditions due to topographic runoff, poor drainage, and evaporative concentration. This method optimizes variables for spectral features, textural features, and topographic hydrological candidate features. SAI, as a prior variable for slope water and salt migration, is not deleted and is forcibly retained in the final key feature set.

[0050] The principle behind constructing SAI is as follows: salt migrates with water. Water stays for a long time at the foot of slopes, along gully banks, lakeshores, and in low-lying areas. After evaporation, salt accumulates on the surface and in shallow soil. At the top of slopes, steep slopes, and in well-drained areas, water stays for a short time, resulting in low salt accumulation potential. Therefore, TWI, VD, and CA are positively correlated with salt accumulation, while RSP, CNBL, slope θ, and CND, after their directions are reversed, are also positively correlated with salt accumulation.

[0051] This method uses SAI for three purposes: first, as an input variable for the topographic and hydrological branch; second, as an input for the SAI-gated multi-branch fusion regression model, enabling the model to increase the weight of topographic and hydrological features in low-lying confluence areas; and third, as a slope water and salt migration constraint in the loss function, ensuring that the predicted salinity of higher SAI pixels is not lower than the predicted salinity of lower SAI pixels in the local neighborhood.

[0052] In the seventh embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the fifth embodiment, step S150 includes the following steps: Step S710: Feature Table Construction: Using the real-time dynamic differential positioning coordinates of the soil sampling points in the target area as the center, extract the effective pixel feature values ​​within the rectangular quadrat range (1m×1m). Take the average value of the effective pixels within the same rectangular quadrat as the feature value corresponding to that soil sampling point. (9) In the formula, For the first The first soil sampling point One eigenvalue; For the first The set of effective pixels within a rectangular quadrat of a soil sampling point; Number of effective pixels; For the first Features in pixels The value at that location.

[0053] Specifically, to reduce redundancy among multimodal feature datasets and improve subsequent model training efficiency, model generalization ability, and feature interpretability, this method optimizes and selects variables for spectral features, spatial texture features, and topographic and hydrological features. Feature variable optimization and selection includes: feature table construction, feature-label correlation screening, modal internal collinearity removal, variance inflation factor screening, SHAP stability contribution assessment, and construction of key multimodal feature sets.

[0054] Step S720: Initial feature vector of each soil sampling point Represented as: (10) In the formula, For the first spectral characteristics of each soil sampling point For the first Spatial texture features of each soil sampling point For the first Topographic and hydrological characteristics of the first soil sampling point, the first Weighted salinity labels at each soil sampling point The salt content was obtained using the above-described stratified weighted salt labeling method.

[0055] Step S730: Calculate each eigenvalue The Pearson correlation coefficient between soil salinity and the label, where the Pearson correlation coefficient is used to determine the eigenvalue. The response strength to changes in soil salinity, and the formula for calculating the Pearson correlation coefficient is as follows: (11) In the formula, For the first Pearson correlation coefficients between eigenvalues ​​and soil salinity labels; For the first The first soil sampling point One eigenvalue; For the first The average value of each feature across all soil sampling points; For the first Soil salinity labels at each soil sampling point; This represents the average soil salinity label from all soil sampling points; Let i be the total number of soil sampling points, where 1 ≤ i ≤ n.

[0056] Step S740: Within the three modes of spectral features, spatial texture features, and topographic and hydrological features, respectively, according to... Sort the features from largest to smallest, and in each modality, input the feature values ​​that strongly respond to the soil salinity label into the next step of filtering, and remove the feature values ​​that weakly respond to the soil salinity label from the feature set.

[0057] Specifically, when the Pearson correlation coefficient between the feature value and the soil salinity label is greater than or equal to 0.85, it indicates that the feature value responds strongly to the soil salinity label; conversely, when the Pearson correlation coefficient between the feature value and the soil salinity label is less than 0.85, it indicates that the feature value responds weakly to the soil salinity label.

[0058] Specifically, SAI is a prior variable for slope water and salt migration and is not included in the elimination process.

[0059] In the eighth embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the seventh embodiment, step S740 is followed by the following steps: Specifically, within each of the three modalities—spectral features, spatial texture features, and topographic and hydrological features—the Pearson correlation coefficient between any two candidate feature values ​​is calculated. When the absolute value of the correlation coefficient between two feature values ​​reaches 0.85, it is determined that both exhibit strong collinearity, and only the one with the soil salinity label is retained. Larger eigenvalues.

[0060] Step S810: After completing the removal of strong collinearity within modalities, continue to calculate the variance inflation factor (VIF) for each retained feature: (12) In the formula, For the first Variance inflation factor for each eigenvalue; For the first The coefficient of determination is obtained by performing linear regression with one feature as the dependent variable and the remaining features within the same mode as the independent variables.

[0061] Step S820: When When, explain the first If the eigenvalue exhibits multicollinearity with other eigenvalues ​​within the same mode, then the eigenvalue should be deleted. One eigenvalue, and recalculate the remaining eigenvalues. Until all retained features within each modality satisfy .

[0062] Step S830: To select features that contribute stably across multiple training iterations, this method uses a random forest regression model to repeatedly train each feature multiple times (100 times), and calculates the SHAP contribution value of each feature in each training iteration. The SHAP contribution value of the feature in each iteration is then calculated. In the training session, the first The average absolute SHAP contribution of each feature is: (13) In the formula, For the first In the training session, the first The average absolute SHAP contribution value of each feature; For the first In the training session, the first The feature is related to the first The SHAP contribution value of the predicted values ​​from the soil sampling points, the ... The SHAP stability contribution of each feature is defined as follows: (14) In the formula, For the first The SHAP stability contribution of each feature; For the first The average SHAP contribution of each feature across multiple (100) training iterations; For the first The standard deviation of the SHAP contribution value of each feature over multiple (100) training iterations. A higher value indicates that the feature contributes more and fluctuates less during multiple training iterations. Spectral features, spatial texture features, and topographic / hydrological features are respectively categorized according to... Sort by size from largest to smallest, the SHAP stability contribution value is used to view the degree of stable response of each feature variable to changes in soil salinity.

[0063] Specifically, the SHAP contribution value is used to precisely quantify the magnitude and direction of the influence of each feature in a machine learning model on a specific prediction result.

[0064] Step S840: After completing correlation screening, collinearity removal, and variance inflation factor screening, retain the remaining spectral features, spatial texture features, and topographic and hydrological features as the final key multimodal feature set. (15) In the formula, For key multimodal feature sets; This indicates the spectral features retained after screening; This indicates the spatial texture features retained after filtering; This indicates the topographic and hydrological features that are retained after screening.

[0065] Specifically, the spectral features, spatial texture features, and topographic and hydrological features obtained after feature variable optimization and screening together constitute the key multimodal feature set used as input to the gated multi-branch fusion regression model.

[0066] In the ninth embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the first embodiment, the gated multi-branch fusion regression model includes a spectral coding branch, a texture coding branch, a topographic and hydrological coding branch, a gated weight allocation layer, and a salinity regression output layer; step S160 includes the following steps: Step S910: The multi-branch encoding of the gated multi-branch fusion regression model is expressed as: (16) In the formula, It is a spectral encoder and a multilayer perceptron; It is a spatial texture encoder and a multilayer perceptron; This is a topographic and hydrological encoder used to learn the nonlinear relationship between topographic and hydrological features and salinity accumulation. These are implicit features of spectral modes. For spatial texture modal latent features, These are implicit features of topographic and hydrological modes.

[0067] Step S920: To enable the gated multi-branch fusion regression model to adaptively adjust the contributions of each modality according to the characteristics of different regions, a gated weight allocation layer is constructed: (17) In the formula, For adaptive weights of spectral modes, For adaptive weights of spatial texture modality, Adaptive weights for topographic and hydrological modes; and All of these are trainable parameters; To ensure that the output dimensions of the three coding branches remain consistent for the latent features of the fused pixels, the following conditions must be met.

[0068] When a cell is located in a low-lying catchment area or at the foot of a slope, the gated layer tends to increase the contribution of the topographic and hydrological branch based on the SAI and topographic and hydrological characteristics; when a cell has a significant salt crust spectral response, the contribution of the spectral branch can be enhanced; when a cell has patchy salinization texture, the contribution of the spatial texture branch can be enhanced; the SAI ranking constraint only takes effect in the local neighborhood, the same land cover type, or similar geomorphic units.

[0069] Step S930: Input the fused latent features into the salinity regression output layer to obtain the estimated soil salinity content: (18) In the formula, For the first Estimated soil salinity at each pixel or soil sampling point location; and All are parameters from the salt regression output layer; The function is used to ensure that the output value is non-negative.

[0070] Specifically, to reduce the discrepancies between the model output and the slope water and salt migration patterns, this method introduces the SAI ranking prior constraint. The basic idea is: within a local neighborhood, if the pixel... SAI is higher than pixel Then like a pixel The predicted salt content should not be significantly lower than that of a pixel. .

[0071] Step S940: Introduce slope prior constraint loss The expression is as follows: (19) In the formula, For pixels The estimated value of soil salinity, For pixels Estimated soil salinity; For pixels Prior indices for slope salt accumulation. For pixels The slope salt accumulation prior index is used, and the slope prior constraint loss is used to penalize the case where the predicted salt difference direction is opposite to the SAI difference direction.

[0072] Step S950: Combine the loss function of the gated multi-branch fusion regression model. Set as: (20) In the formula, The Huber loss is used to represent the loss in the estimated soil salinity. For slope prior constraint loss; These are the weighting coefficients.

[0073] Specifically, through the above structure, this model can retain the nonlinear fitting ability of deep learning while introducing prior knowledge of slope water and salt migration, thereby improving the spatial rationality and interpretability of the salinity estimation results.

[0074] The data samples were divided into training, validation and test sets according to spatial partitioning. Samples within the same spatial block were included in the same data subset. The gated multi-branch fusion regression model was trained using the training set. The spatial partitioning, model initialization and parameter updates were performed again for each training session. The final accuracy was the average of the evaluation results of 100 test sets.

[0075] Step S950: Use the coefficient of determination The root mean square error (RMSE), mean absolute error (MAE), and interquartile range (RPIQ) are used to evaluate the accuracy of gated multi-branch fusion regression models.

[0076] The model is used to evaluate its ability to explain the variation in measured salinity; RMSE is used to evaluate the magnitude of the estimation error; MAE is used to evaluate the mean absolute deviation; RPIQ is used to evaluate the model's predictive ability relative to the dispersion of salinity samples; among these, The formula for calculating the coefficient of determination is: ,(twenty one) Root mean square error: ,(twenty two) Mean absolute error: ,(twenty three) Interquartile range performance ratio: ,(twenty four) In the formula, To measure the soil salinity; To estimate soil salinity; This is the average of the measured values; This represents the total number of soil sampling points; This represents the difference between the upper and lower quartiles of the measured salt content. and The larger, and The smaller the value, the better the model estimation.

[0077] In the tenth embodiment of the deep learning-based method for estimating soil salinity at the slope regional scale proposed in this invention, based on the first embodiment, step S170 includes the following steps: Step S1010: For each valid soil pixel within the target area, determine the key multimodal feature set to extract the corresponding model input features: (25) In the formula, For the first The model input features for each effective soil pixel; For the first Key spectral features of each effective soil pixel For the first Key spatial texture features of an effective soil pixel For the first Key topographic and hydrological features of each effective soil pixel; For the first Prior index of slope salt accumulation for each effective soil cell.

[0078] Specifically, effective soil pixels refer to bare soil, salt crust, low vegetation cover, or sparse vegetation cover pixels that are retained after masking. Water bodies, buildings, hardened road surfaces, deep shadows, highly reflective non-soil targets, high vegetation cover areas, and invalid areas at flight boundaries are not included in pixel-level salinity estimation. The feature types, feature order, standardization methods, and masking rules in the mapping stage are consistent with those in the training stage.

[0079] Step S1020: Input the model input features of the effective soil pixels into the trained gated multi-branch fusion regression model to obtain the estimated soil salinity content of the effective soil pixels: (26) In the formula, For the first Estimated soil salinity content for each effective soil cell; This is the completed soil salinity estimation model.

[0080] Step S1030: To reduce the impact of sample partitioning, gated multi-branch fusion regression model initialization, and training perturbations on the mapping results, multiple gated multi-branch fusion regression models can be repeatedly trained, and predictions can be made multiple times for the same pixel. Let the total training time be... The first gated multi-branch fusion regression model, the... The first gated multi-branch fusion regression model for the first The predicted value of each effective soil cell is The final soil salinity mapping value is: (26) In the formula, For the first The final soil salinity mapping value for each effective soil cell.

[0081] Step S1040: Use the standard deviation of multiple prediction results to represent the uncertainty in the graph: (Equation 1) In the formula, For the first Mapping uncertainty of each effective soil pixel The larger the value, the greater the fluctuation of the pixel in multiple model predictions, and the lower the mapping stability. The smaller the value, the more stable the prediction result for that pixel.

[0082] Step S1050: Calculate the estimated soil salinity content for all effective soil pixels in the target area and represent the mapping uncertainty to output a spatial distribution map of soil salinity in the target area. The spatial distribution map of soil salinity includes a continuous distribution map of soil salinity content, a soil salinity uncertainty map, and a mask map of effective soil pixels.

[0083] Specifically, a continuous distribution map of soil salinity is used to represent the soil salinity estimation result for each effective soil cell in the target area; a soil salinity uncertainty map (made based on mapping uncertainty) is used to represent the stability of the salinity estimation result for each effective soil cell; and an effective soil cell mask map is used to mark the effective soil cells involved in soil salinity modeling and mapping, and to mark water bodies, buildings, hardened road surfaces, deep shadows, strongly reflective non-soil targets, high vegetation cover areas, and invalid areas of flight boundaries as no data.

[0084] Specifically, through the above spatial mapping process, this method can generate continuous, quantitative, and uncertain spatial distribution results of soil salinity in the target area, providing a basis for the zoning of saline-alkali land management, the layout of salt drainage projects, and ecological restoration.

[0085] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0086] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for estimating soil salinity at the slope regional scale based on deep learning, characterized in that, include: Multimodal data of the target area is acquired by drones, and the multimodal data is preprocessed to obtain a multimodal feature dataset. Slope layered soil sampling points were set up in the target area, and soil salinity labels were constructed. Feature variables related to the spatial distribution of soil salinity were extracted from the multimodal feature dataset. These feature variables included spectral features, spatial texture features, and topographic and hydrological features. Construct a priori index of slope salinity accumulation in the target area; Optimize the selection of feature variables based on salt content labels; A gated multi-branch fusion regression model based on the prior index of slope salt accumulation was constructed, and the gated multi-branch fusion regression model was trained and its accuracy was evaluated. Based on the gated multi-branch fusion regression model that has completed training and accuracy evaluation, and the optimized and screened feature variables, a spatial distribution map of soil salinity corresponding to the target area is generated.

2. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 1, characterized in that, The drone is equipped with a spectral sensor; the multimodal feature dataset includes multispectral imagery, digital surface models, digital elevation models, and digital terrain models; the process of acquiring multimodal data for a target area via the drone and preprocessing the multimodal data to obtain the multimodal feature dataset includes: The target area is acquired by using a drone, wherein the multispectral image includes two bands: red light and near-infrared light. A digital surface model of the target area is generated by reconstructing multispectral images; Non-ground point removal, vegetation / building masking, filtering, and smoothing are performed on the digital surface model to generate a digital elevation model or digital terrain model for hydrological analysis. Percentile normalization of multispectral images: ，（1） In the formula, These are the standardized eigenvalues; These are the original pixel feature values ​​of the multispectral image; This represents the 2nd percentile value of this feature within the target region. This represents the 98th percentile value of the feature within the target region. Then let ;like Then let .

3. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 1, characterized in that, The step of setting up stratified soil sampling points on the slope in the target area and constructing soil salinity labels includes: Inversion is performed based on the digital surface model of the target area to obtain the slope, relative slope position, runoff area, distance from the gully or lake shore, bare soil patches, apparent brightness of salt crust, and vegetation sparseness of the target area. The target area is then divided into zones, including the top of the slope, the middle of the slope, the foot of the slope, the low-lying runoff area, the gully bank area, the lakeside area, and the relatively flat background area. Sampling points were evenly distributed within each zone. A rectangular quadrat was set up at each sampling point. Soil samples were collected at five locations, including the four corners and the center of the quadrat, and the real-time dynamic differential positioning coordinates of the sampling points were recorded. Soil samples were taken from different depths within quadrats to determine the salinity content of each soil layer and to construct a weighted salinity label. ，（2） In the formula, Y represents the soil salinity label; This refers to the soil salinity at a depth of 0-2cm. This refers to the soil salinity at a depth of 2-5 cm. This refers to the soil salinity at a depth of 5-10cm.

4. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 1, characterized in that, The extraction of feature variables related to the spatial distribution of soil salinity from the multimodal feature dataset includes: Spectral features are extracted from multispectral images in a multimodal feature dataset, where spectral features are used to express the effects of salinity, soil background, and vegetation cover on surface reflectance. Normalized difference salinity index obtained based on spectral features: ，（3） Soil brightness index obtained based on spectral characteristics: ，（4） Normalized vegetation index obtained based on spectral features: ，（5） In the formula, The reflectance in the red band of the spectrum. This represents the near-infrared reflectance in the spectral characteristics.

5. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 4, characterized in that, The process of obtaining the normalized vegetation index based on spectral features further includes: Based on the digital surface model of the target area, topographic and hydrological features at the slope scale of the target area are extracted. The topographic and hydrological features are divided into geomorphic indicators and hydrological indicators. The geomorphic indicators include: mountain shadow, slope, aspect, plane curvature, profile curvature, convergence index, topographic humidity index and relative slope position. The hydrological indicators include: closed depressions, total catchment area, LS factor, gully network reference surface, gully network distance and valley depth. The formula for calculating slope is: ，（6） In the formula, The slope is mentioned. The elevation change rate of the central pixel in the horizontal direction. The elevation change rate of the central pixel in the vertical direction; The formula for calculating the topographic humidity index is: ，（7） In the formula, TWI is the terrain humidity index; The upslope runoff area per unit contour line length; Slope; Based on multispectral imagery, the spatial texture features of the target region are extracted using the gray-level co-occurrence matrix method. The spatial texture features include statistical texture features, contrastive texture features, and ordered texture features. Statistical texture features include mean, variance, and correlation. Contrastive texture features include contrast, dissimilarity, and homogeneity. Ordered texture features include second moment and entropy.

6. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 1, characterized in that, The formula for calculating the prior index of slope salt accumulation is as follows: ，（8） In the formula, The prior index for salt accumulation on the slope; This is the percentile standardization function; The topographic humidity index; The valley is deep; The catchment area; Relative slope position; This serves as the reference surface for the trench. Slope; Distance from the channel; , , , , , and All are weighting coefficients.

7. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 5, characterized in that, The optimization and screening of feature variables based on salt labels includes: Feature table construction: Centered on the real-time dynamic differential positioning coordinates of soil sampling points in the target area, extract the effective pixel feature values ​​within the rectangular quadrat area, and take the average value of the effective pixels within the same rectangular quadrat as the feature value corresponding to that soil sampling point. ，（9） In the formula, For the first The first soil sampling point One eigenvalue; For the first The set of effective pixels within a rectangular quadrat of a soil sampling point; Number of effective pixels; For the first Features in pixels The value at that location; No. Initial feature vector of each soil sampling point Represented as: ，（10） In the formula, For the first spectral characteristics of each soil sampling point For the first Spatial texture features of each soil sampling point For the first Topographic and hydrological characteristics of each soil sampling point; Calculate each eigenvalue The Pearson correlation coefficient between soil salinity and the label, where the Pearson correlation coefficient is used to determine the eigenvalue. The response strength to changes in soil salinity, and the formula for calculating the Pearson correlation coefficient is as follows: ，（11） In the formula, For the first Pearson correlation coefficients between eigenvalues ​​and soil salinity labels; For the first The first soil sampling point One eigenvalue; For the first The average value of each feature across all soil sampling points; For the first Soil salinity labels at each soil sampling point; This represents the average soil salinity label from all soil sampling points; Let i be the total number of soil sampling points, 1 ≤ i ≤ n; Within the three modes of spectral features, spatial texture features, and topographic and hydrological features, respectively, according to Sort the features from largest to smallest, and in each modality, input the feature values ​​that strongly respond to the soil salinity label into the next step of filtering, and remove the feature values ​​that weakly respond to the soil salinity label from the feature set.

8. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 7, characterized in that, Within the three modalities of spectral features, spatial texture features, and topographic and hydrological features, respectively, according to... Sort the features from largest to smallest, and within each modality, input the feature values ​​that strongly respond to the soil salinity label into the next filtering step, while removing the feature values ​​that weakly respond to the soil salinity label from the feature set. This process also includes: Calculate the variance inflation factor (VIF) for each retained feature: ，（12） In the formula, For the first Variance inflation factor for each eigenvalue; For the first The coefficient of determination is obtained by performing linear regression with one feature as the dependent variable and the remaining features within the same mode as the independent variables. when At that time, delete the first one. One eigenvalue, and recalculate the remaining eigenvalues. Until all retained features within each modality satisfy ; A random forest regression model is used to repeatedly train each feature, and the SHAP contribution value of each feature in each training iteration is calculated. The SHAP contribution value of the feature in each training iteration is then calculated. In the training session, the first The average absolute SHAP contribution of each feature is: ，（13） In the formula, For the first In the training session, the first The average absolute SHAP contribution value of each feature; For the first In the training session, the first The feature is related to the first The SHAP contribution value of the predicted values ​​from the soil sampling points, the ... The SHAP stability contribution of each feature is defined as follows: ，（14） In the formula, For the first The SHAP stability contribution of each feature; For the first The average SHAP contribution value of each feature across multiple training iterations; For the first The standard deviation of the SHAP contribution value of each feature in multiple training sessions; the stable SHAP contribution value is used to view the degree of stable response of each feature variable to changes in soil salinity. The remaining spectral features, spatial texture features, and topographic and hydrological features are retained and used as the final key multimodal feature set: ，（15） In the formula, For key multimodal feature sets; This indicates the spectral features retained after screening; This indicates the spatial texture features retained after filtering; This indicates the topographic and hydrological features that are retained after screening.

9. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 1, characterized in that, The gated multi-branch fusion regression model includes a spectral coding branch, a texture coding branch, a topographic and hydrological coding branch, a gated weight allocation layer, and a salinity regression output layer. The construction of a gated multi-branch fusion regression model based on a priori index of slope salt accumulation, and the training and accuracy evaluation of the gated multi-branch fusion regression model, include: The multi-branch encoding of the gated multi-branch fusion regression model is expressed as follows: ，（16） In the formula, It is a spectral encoder and a multilayer perceptron; It is a spatial texture encoder and a multilayer perceptron; This is a topographic and hydrological encoder used to learn the nonlinear relationship between topographic and hydrological features and salinity accumulation. These are implicit features of spectral modes. For spatial texture modal latent features, These are implicit features of topographic and hydrological modes; Construct a gating weight allocation layer: ，（17） In the formula, For adaptive weights of spectral modes, For adaptive weights of spatial texture modality, Adaptive weights for topographic and hydrological modes; and All of these are trainable parameters; These are the latent features of the fused pixels; The fused latent features are input into the salinity regression output layer to obtain an estimate of soil salinity content: ，（18） In the formula, For the first Estimated soil salinity at each pixel or soil sampling point location; and All are parameters from the salt regression output layer; The function is used to ensure that the output value is non-negative; Introducing slope prior constraint loss The expression is as follows: ，（19） In the formula, For pixels The estimated value of soil salinity, For pixels Estimated soil salinity; For pixels Prior indices for slope salt accumulation. For pixels Prior indices for slope salt accumulation; The comprehensive loss function of the gated multi-branch fusion regression model Set as: ，（20） In the formula, Loss of estimated soil salinity; For slope prior constraint loss; These are the weighting coefficients.

10. The method for estimating soil salinity at the slope regional scale based on deep learning according to claim 1, characterized in that, The method for generating a spatial distribution map of soil salinity in the target region based on a gated multi-branch fusion regression model that has undergone training and accuracy evaluation, and optimized and screened feature variables, includes: For each valid soil pixel within the target area, determine the key multimodal feature set to extract the corresponding model input features: ，（25） In the formula, For the first The model input features for each effective soil pixel; For the first Key spectral features of each effective soil pixel For the first Key spatial texture features of an effective soil pixel For the first Key topographic and hydrological features of each effective soil pixel; For the first Prior index of slope salt accumulation for each effective soil cell; By inputting the model features of effective soil pixels into the trained gated multi-branch fusion regression model, the estimated soil salinity content of the effective soil pixels is obtained: ，（26） In the formula, For the first Estimated soil salinity content for each effective soil cell; The completed soil salinity estimation model; Set up training The first gated multi-branch fusion regression model, the... The first gated multi-branch fusion regression model for the first The predicted value of each effective soil cell is The final soil salinity mapping value is: ，（26） In the formula, For the first The final soil salinity mapping value for each effective soil pixel; The uncertainty in the graph is represented by the standard deviation of multiple prediction results: (Equation 1) In the formula, For the first Mapping uncertainty of each effective soil pixel; The soil salinity content of all effective soil pixels in the target area is estimated and the mapping uncertainty is expressed to output a spatial distribution map of soil salinity in the target area. The spatial distribution map of soil salinity includes a continuous distribution map of soil salinity, a soil salinity uncertainty map, and a mask map of effective soil pixels.

Images