A soda saline soil salt content three-dimensional prediction method based on unmanned aerial vehicle texture constraint

Abstract

The application belongs to the field of soil environment remote sensing monitoring, and discloses a soda saline soil salt three-dimensional prediction method based on unmanned aerial vehicle texture constraint, which comprises the following steps: collecting satellite multispectral remote sensing images, unmanned aerial vehicle multi-source images, terrain, weather and soil salt data of different depths, and integrating the data to a reference grid after pretreatment; extracting unmanned aerial vehicle wide band and gray level co-occurrence matrix texture features, combining regional conservation and physical extreme value constraint with extreme random tree surface to point downscaling to reconstruct the downscaling salt index; constructing basic environmental variables and interactive features with soil depth to form a three-dimensional prediction covariant data set; training and optimizing the extreme random tree model, and layer by layer predicting and outputting high-precision three-dimensional salt distribution results. The method solves the problems of insufficient spatial resolution and unreasonable data fusion in the prior art, takes into account the advantages of multi-source data, realizes accurate prediction and uncertainty quantification of salt at different depths, provides reliable support for saline-alkali soil treatment, and has significant application value.

Description

Technical Field

[0001] This invention relates to the field of soil environmental remote sensing monitoring technology, specifically to a method, system, electronic device, and storage medium for three-dimensional prediction of salinity in soda saline-alkali soil based on UAV texture constraints. Background Technology

[0002] Soil salinization is a common and typical ecological and environmental problem in arid and semi-arid regions, and it is also a key factor restricting the healthy development of regional agriculture and the sustainable use of land resources. Accurately obtaining spatial distribution information of soil salinity at the field scale and understanding the distribution characteristics of salinity at different soil depths is of great practical significance for carrying out precise improvement of saline-alkali land, optimized allocation of water resources, adjustment of agricultural industrial structure, and regional ecological restoration. Traditional soil salinity monitoring methods mainly rely on field sampling combined with laboratory physicochemical testing. The overall workflow is cumbersome, with high manpower and time costs, and the discrete sampling data cannot fully reflect the continuous spatial variation of soil salinity across the entire area.

[0003] The Songnen Plain in Northeast my country is a core area of ​​concentrated soda saline-alkali soil. This region has a unique soil salinity composition, resulting in a pattern of interspersed vegetation communities and exposed saline-alkali land. When using conventional medium-resolution satellite remote sensing imagery for monitoring, pixel mixing is easily encountered, making accurate identification of small-scale saline-alkali areas difficult. Current methods for soil salinity prediction based on remote sensing technology still have many technical shortcomings. The overall monitoring precision is insufficient, the methods for integrating multi-source remote sensing data are relatively limited, traditional downscaling modeling techniques cannot adapt to the nonlinear correlation characteristics of complex subsurface salinity distribution, and most studies only focus on surface soil salinity monitoring, lacking a coupled modeling approach for vertical salinity distribution characteristics. Furthermore, existing prediction results lack uncertainty assessment criteria, making it difficult to effectively distinguish the reliability range of prediction results and meet the needs of practical production and remediation.

[0004] Current remote sensing technologies for soil salinity prediction cannot simultaneously leverage the advantages of satellite remote sensing's large-scale, comprehensive monitoring and UAV remote sensing's fine-scale spatial representation. Consequently, the overall technological system is ill-suited for the application scenario of three-dimensional, refined prediction of soil salinity in soda-saline-alkali soil regions. Therefore, developing an intelligent soil salinity prediction method that integrates multi-source remote sensing data and considers both spatial precision and vertical soil layer distribution characteristics can effectively address the shortcomings of existing technologies, achieving efficient and accurate prediction of soil salinity at different depths. This will provide reliable data support for ecological governance and agricultural production planning in saline-alkali land.

[0005] Therefore, we designed a three-dimensional prediction method for salinity of soda-alkali soil based on UAV texture constraints to solve the above problems. Summary of the Invention

[0006] To address the above technical problems, this invention provides a three-dimensional prediction method for salinity of soda-alkali soil based on UAV texture constraints, the method comprising the following steps: S100 collects satellite multispectral remote sensing images, UAV RGB orthophoto images, UAV multispectral images including near-infrared and red edge bands, digital terrain data, meteorological data, and measured soil profile salinity data at different depths in the target study area. After preprocessing all types of data, they are integrated into a benchmark grid with a preset accuracy. S200: Based on the extraction of wideband features and texture features of gray-level co-occurrence matrix from UAV images, these are used as point-scale fine spatial covariates. The surface-to-point spatial downscaling method based on extreme random trees is used, combined with regional conservation constraints and physical extreme value constraints, to perform spatial downscaling on the core salinity index. Finally, a downscaled salinity index with the same accuracy as the benchmark grid is reconstructed. S300: Based on the downscaled salinity index, the basic environmental variables are constructed by combining topographic factors, meteorological factors and gray-level co-occurrence matrix texture features. Soil depth is used as a continuous variable to construct the interaction features between environmental variables and soil depth, forming a covariate dataset for three-dimensional prediction of soil salinity. S400. Using the covariate dataset as the independent variable and the measured sample data of salinity in the stratified soil profile as the dependent variable, an extreme random tree nonlinear regression model is constructed, and model training and parameter optimization are completed through cross-validation. S500 uses an optimized extreme random tree nonlinear regression model to predict soil salinity at different depths within the target study area, outputting a three-dimensional spatial distribution of soil salinity consistent with a preset high-precision benchmark grid.

[0007] As a further improvement of the present invention, in step S200, the surface-to-point spatial downscaling method based on extreme random trees specifically includes the following steps: Using satellite remote sensing image pixels as regional support units and a reference grid with preset precision as point support units, the spatial correspondence between regional support units and point support units is clearly defined. Using the core salinity index as the target variable and the broadband features and GLCM texture features extracted from UAV images as covariates, an extreme random tree nonlinear regression model was established to obtain point-scale trend prediction values. The trend residual at the region scale is calculated, and bilinear interpolation is used to smoothly distribute the region-scale residual to the point-scale grid cells under the region conservation constraint. The predicted point-scale trend value is added to the allocated point-scale residual, and the synthesis result is constrained based on the physical extrema of the original region support unit to obtain a downscaled salinity index with the same resolution as the reference grid.

[0008] As a further improvement of the present invention, the construction of the core salinity index in step S200 specifically includes the following steps: Select characteristic bands in satellite multispectral remote sensing images that are sensitive to the salinity of soda saline-alkali soil; Based on the spectral response characteristics of soda saline-alkali soil, multiple salt sensitivity indices were constructed. The salt sensitivity indices include the improved normalized salt index, the canopy-response salt index, and the optimized soda saline soil salt index.

[0009] As a further improvement of the present invention, in step S200, the extraction of texture features based on the gray-level co-occurrence matrix specifically includes the following steps: Select bands in UAV imagery that are sensitive to soil salinity distribution; Set the sliding window to the appropriate size; Texture features are extracted within the sliding window using the gray-level co-occurrence matrix algorithm. The texture features include contrast, entropy, homogeneity, second moment of angle, correlation, and heterogeneity.

[0010] As a further improvement of the present invention, the extraction of broadband features in step S200 specifically includes the following steps: Extract multi-band reflectance features from UAV images, the multi-band including RGB band and near-infrared band; Vegetation-related indices are constructed by combining and calculating reflectance from different wavebands. The broadband features include the multi-band reflectance features and vegetation-related indices.

[0011] As a further improvement of the present invention, the construction of the interactive features in step S300 specifically includes the following steps: Clarify the coupling relationship between basic environmental variables and soil depth; Through mathematical correlation operations, the interaction characteristics between basic environmental variables and soil depth are constructed; The interactive features include the interaction between downscaled salinity index and soil depth, the interaction between gray-level co-occurrence matrix texture features and soil depth, the interaction between meteorological factors and soil depth, and the interaction between topographic factors and soil depth.

[0012] As a further improvement of the present invention, the construction and optimization of the extreme random tree nonlinear regression model in step S400 specifically includes the following steps: Set model configuration strategies, including non-bootstrapping resampling sample sampling strategies, adaptive feature selection strategies based on the total number of covariates, and the number of models ensembled to balance prediction accuracy and computational efficiency. Using the covariate dataset as the independent variable and the measured soil salinity data at different depths as the dependent variable, an extreme random tree nonlinear regression model was initialized. Cross-validation was used to train and optimize the model.

[0013] As a further improvement of the present invention, a step of predicting uncertainty quantification is also included: Based on the covariate dataset constructed in step S300 and the measured sample data of soil salinity at different depths, a quantile regression forest model was constructed. Set the corresponding quantile level; The prediction results of soil salinity at different depths are output in the prediction intervals at each quantile level, thereby achieving a quantitative characterization of prediction uncertainty.

[0014] Based on this, the present invention also provides a three-dimensional prediction system for salinity of soda-alkali soil based on UAV texture constraints. The system is used in the above-described method and includes... The data acquisition and preprocessing module is used to collect multi-source data and measured sample data of soil salinity at different depths in the target study area, complete geometric registration, projection transformation and spatial resolution unification processing, and integrate them into a reference grid with preset accuracy. The spatial downscaling module is used to extract wideband features and gray-level co-occurrence matrix texture features of UAVs, perform surface-to-point spatial downscaling, and output the downscaling salinity index with a preset accuracy of the reference grid resolution. The 3D covariate construction module is used to fuse downscaled salinity index, topographic factors, meteorological factors and gray-level co-occurrence matrix texture features to generate environmental variables and soil depth interaction features, and output a 3D prediction covariate dataset of soil salinity at different depths. The 3D prediction module is used to build and train an extreme random tree nonlinear regression model to predict soil salinity at different depths layer by layer, and output the 3D spatial distribution results of soil salinity at different depths with a preset accuracy benchmark grid resolution. The uncertainty quantification module is used to construct a quantile regression forest model based on covariate datasets and measured sample data of soil salinity at different depths, output the salinity prediction quantile interval, and quantify the prediction uncertainty.

[0015] Based on this, the present invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed, the following steps are performed: acquiring satellite multispectral remote sensing images, UAV RGB orthophoto images, UAV multispectral images including near-infrared and red-edge bands, digital terrain data, meteorological data, and measured salinity sample data of soil profiles at different depths for the target study area; preprocessing the various data types and integrating them into a reference grid with a preset precision; extracting wideband features and texture features of the gray-level co-occurrence matrix based on UAV images, using them as point-scale fine spatial covariates; employing a surface-to-point spatial downscaling method based on extreme random trees, combined with regional conservation constraints and physical extreme value constraints, to spatially downscale the core salinity index. After processing, a downscaled salinity index with the same accuracy as the benchmark grid was reconstructed. Based on the downscaled salinity index, basic environmental variables were constructed by combining topographic factors, meteorological factors, and gray-level co-occurrence matrix texture features. Soil depth was used as a continuous variable to construct the interaction features between environmental variables and soil depth, forming a covariate dataset for three-dimensional prediction of soil salinity. Using the covariate dataset as the independent variable and the measured salinity sample data of the layered soil profile as the dependent variable, an extreme random tree nonlinear regression model was constructed. The model was trained and the parameters were optimized through cross-validation. Using the trained and optimized extreme random tree nonlinear regression model, the soil salinity at each soil layer depth within different depth ranges in the target study area was predicted layer by layer, and the three-dimensional spatial distribution results of soil salinity consistent with the preset high-precision benchmark grid were output.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively solves many problems existing in soil salinity remote sensing prediction technology, such as insufficient spatial resolution, unreasonable multi-source remote sensing data fusion methods, difficulty in adapting to nonlinear surface distribution relationships, inability to achieve integrated prediction of soil salinity at different depths, and lack of prediction reliability assessment. By fusing macroscopic spectral information from satellite remote sensing with fine-scale spatial features from UAVs, and relying on extreme random trees to complete surface-to-point spatial downscaling, it effectively reduces the monitoring error caused by mixed pixels in satellite imagery, fully leverages the advantages of UAV data in fine-scale spatial structure representation, accurately identifies the distribution of small-scale saline-alkali patches, and simultaneously constructs environmental variables and soil layers... This technology leverages deep interactive features to establish a predictive dataset adapted to the vertical distribution patterns of salinity, overcoming the limitations of traditional surface salinity prediction. It enables accurate prediction of salinity at different soil depths and employs nonlinear modeling to fit the complex spatial distribution characteristics of soda saline-alkali soil regions. This simplifies the computation process for massive grid data, balancing prediction accuracy with overall computational efficiency. Furthermore, it can perform quantitative analysis of the uncertainty of prediction results, clarifying the credibility range of the prediction results. This provides comprehensive and reliable data for practical work such as precise improvement of saline-alkali land, optimized allocation of water resources, and regional ecological restoration, demonstrating significant practical application and engineering value. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are 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 these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the three-dimensional prediction method for salinity in soda-alkali soil in an embodiment of the present invention; Figure 2 This is a schematic diagram of the spatial downscaling of the core salinity index surface to point in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1 As the background technology indicates, current remote sensing prediction technologies for soil salinity cannot simultaneously leverage the advantages of satellite remote sensing's large-scale, comprehensive monitoring and UAV remote sensing's fine-scale spatial representation. Consequently, the overall technical system is ill-suited for the application scenario of three-dimensional, refined prediction of soil salinity in soda-alkali soil regions. Therefore, developing an intelligent soil salinity prediction method that integrates multi-source remote sensing data and considers both spatial precision and vertical soil layer distribution characteristics can effectively address the shortcomings of existing technologies, achieving efficient and accurate prediction of soil salinity at different depths, and providing reliable data support for ecological governance and agricultural production planning in saline-alkali land.

[0023] Please see Figure 1 and Figure 2 This invention provides a three-dimensional prediction method for salinity of soda-alkali soil based on UAV texture constraints. The method includes the following steps: S100 collects satellite multispectral remote sensing images, UAV RGB orthophoto images, UAV multispectral images including near-infrared and red edge bands, digital terrain data, meteorological data, and measured soil profile salinity data at different depths in the target study area. After preprocessing all types of data, they are integrated into a benchmark grid with a preset accuracy. In the above steps, by completing the unified integration and processing of multi-source remote sensing data, geographic environment data and field measurement data, the spatial coordinates and spatial scales of all data sources can be unified, spatial matching errors between different types of data can be eliminated, and a standardized and unified data foundation can be built for subsequent feature extraction, spatial downscaling and intelligent modeling.

[0024] S200: Based on the extraction of wideband features and texture features of gray-level co-occurrence matrix from UAV imagery, these are used as point-scale fine spatial covariates. A surface-to-point spatial downscaling method based on extreme random trees is used, combined with regional conservation constraints and physical extremum constraints, to perform spatial downscaling on the core salinity index. Finally, a downscaled salinity index with the same accuracy as the baseline grid is reconstructed.

[0025] Furthermore, in step S200, the surface-to-point spatial downscaling method based on extreme random trees specifically includes the following steps: Using satellite remote sensing image pixels as regional support units and a benchmark grid with preset precision as point support units, the spatial correspondence between regional support units and point support units is clarified. Using the core salinity index as the target variable and wideband features extracted from UAV imagery and gray-level co-occurrence matrix texture features as covariates, an extreme random tree nonlinear regression model is established to obtain point-scale trend prediction values. The regional-scale trend residuals are calculated, and bilinear interpolation is used to smoothly distribute the regional-scale residuals to point-scale grid units under regional conservation constraints. The point-scale trend prediction values ​​are added to the distributed point-scale residuals, and boundary constraints are applied to the synthetic results based on the physical extrema of the original regional support units to obtain a downscaled salinity index with the same resolution as the benchmark grid.

[0026] In the above steps, by establishing the correspondence between regional units and fine grid units and using nonlinear regression modeling, it is possible to fit the complex spatial distribution relationship of saline-alkali land. Relying on residual allocation and extreme value boundary constraints, the overall consistency of salinity values ​​before and after downscaling is ensured, thereby effectively reducing the mixed pixel problem in satellite remote sensing images and achieving accurate conversion of satellite salinity index to fine grid scale.

[0027] Furthermore, in step S200, the construction of the core salinity index specifically includes the following steps: Feature bands sensitive to the salinity of soda saline-alkali soil in satellite multispectral remote sensing images were selected; multiple salinity sensitivity indices were constructed based on the spectral response characteristics of soda saline-alkali soil; the salinity sensitivity indices include the improved normalized salinity index, the canopy response salinity index, and the optimized soda saline soil salinity index.

[0028] In the above steps, by relying on the unique spectral characteristics of soda saline-alkali soil to screen characteristic bands and build a unique salinity index, it is possible to accurately capture spectral change information related to soil salinity content in satellite images, thereby improving the effectiveness and relevance of basic salinity inversion indicators.

[0029] More specifically, in step S200, the extraction of texture features based on the gray-level co-occurrence matrix includes the following steps: Select bands from UAV imagery that are suitable for characterizing soil salinity distribution; Set the sliding window to the appropriate size; Texture features are extracted within a sliding window using the gray-level co-occurrence matrix algorithm; Texture features include contrast, entropy, homogeneity, second moment of angle, correlation, and heterogeneity.

[0030] In addition, the extraction of broadband features in step S200 specifically includes the following steps: extracting multi-band reflectance features of UAV images, including RGB bands and near-infrared bands; performing combined calculations based on the reflectance of different bands to construct vegetation-related indices; broadband features include multi-band reflectance features and vegetation-related indices.

[0031] In the above steps, by simultaneously extracting the spectral reflectance features and surface texture structure features of UAV images, we can give full play to the high-resolution observation advantages of UAVs, comprehensively characterize the fine-scale spatial differences in surface vegetation cover and saline-alkali exposed areas, and thus provide sufficient and refined modeling covariates for the spatial downscaling process.

[0032] S300: Based on the downscaled salinity index, the basic environmental variables are constructed by combining topographic factors, meteorological factors and gray-level co-occurrence matrix texture features. Soil depth is used as a continuous variable to construct the interaction features between environmental variables and soil depth, forming a covariate dataset for three-dimensional prediction of soil salinity. Furthermore, the construction of interactive features in step S300 specifically includes the following steps: clarifying the coupling relationship between basic environmental variables and soil depth; constructing interactive features between basic environmental variables and soil depth through mathematical correlation operations; the interactive features include the interactive features between downscaled salinity index and soil depth, the interactive features between gray-level co-occurrence matrix texture features and soil depth, the interactive features between meteorological factors and soil depth, and the interactive features between topographic factors and soil depth.

[0033] In the above steps, by coupling and associating various spatial environmental features with soil depth, we can explore the intrinsic evolution law of soil salinity with soil depth, break the limitations of single planar feature modeling, and then construct a three-dimensional feature dataset that can be adapted to the prediction of soil salinity at different depths.

[0034] S400. Using the covariate dataset as the independent variable and the measured sample data of salinity in the stratified soil profile as the dependent variable, an extreme random tree nonlinear regression model is constructed, and model training and parameter optimization are completed through cross-validation. Specifically, the construction and optimization of the extreme random tree nonlinear regression model in step S400 includes the following steps: setting a model configuration strategy, including a sample sampling strategy without bootstrapping resampling, an adaptive feature selection strategy based on the total number of covariates, and a model ensemble number that balances prediction accuracy and computational efficiency; initializing the extreme random tree nonlinear regression model with the covariate dataset as the independent variable and the measured sample data of soil salinity at different depths as the dependent variable; and training and optimizing the model using cross-validation.

[0035] In the above steps, by standardizing the model's operating parameters and using cross-validation training, the computational efficiency and fitting effect of the model can be reasonably balanced, effectively reducing the risk of model overfitting, and thus significantly improving the model's ability to generalize and predict soil salinity in different regions and at different depths.

[0036] S500 uses an optimized extreme random tree nonlinear regression model to predict soil salinity at different depths within the target study area, outputting a three-dimensional spatial distribution of soil salinity consistent with a preset high-precision benchmark grid.

[0037] Furthermore, in a preferred embodiment of the present invention, a step of predicting uncertainty quantification is also included: Based on the covariate dataset constructed in step S300 and the measured sample data of soil salinity at different depths, a quantile regression forest model is constructed; corresponding quantile levels are set; and the prediction intervals of soil salinity prediction results at different depths at each quantile level are output, thereby realizing the quantitative characterization of prediction uncertainty.

[0038] In the above steps, by building a quantile regression forest model to carry out interval prediction, the credibility of the salinity prediction results in different regions can be intuitively determined, and regions with high and low confidence in the prediction can be divided, thus providing a scientific reference for field soil sampling planning and graded management of saline-alkali land.

[0039] In summary, this invention sequentially completes the entire process of unified normalization of multi-source heterogeneous data, fine remote sensing feature extraction, high-precision spatial downscaling of satellite salinity index, construction of a three-dimensional coupled feature dataset, intelligent prediction model training and optimization, and full-domain hierarchical salinity prediction. It fully integrates the advantages of large-scale observation by satellite remote sensing and the advantages of fine-scale spatial representation by UAVs. It not only solves the problems of insufficient spatial precision and severe interference from mixed pixels in traditional remote sensing salinity monitoring, but also makes up for the shortcomings of existing technologies that cannot take into account both the planar distribution and vertical evolution of soil salinity. At the same time, with the addition of uncertainty quantification analysis methods, the three-dimensional prediction results of soil salinity are high-precision, comprehensive, and practical, which can efficiently meet the needs of various practical applications such as agricultural production planning, ecological restoration, and water and soil resource allocation in saline-alkali land.

[0040] Example 2 In this embodiment 2, specific parameters are introduced for calculation. Preferably, the reference grid accuracy is 1m and the soil depth is 0-180cm.

[0041] The purpose of this invention is to overcome the problems of insufficient spatial resolution of satellite remote sensing, insufficient utilization of high-resolution information from UAVs, limited nonlinear fitting ability of downscaling algorithms, and insufficient vertical prediction ability of soil salinity in existing technologies. This invention provides a method and system for three-dimensional prediction of salinity in soda-alkali soil based on UAV texture constraints and satellite remote sensing downscaling. This method uses wideband features and gray-level co-occurrence matrix texture features extracted from UAV images as point-scale spatial covariates. It performs area-to-point spatial downscaling (ATP-ET) based on extreme random trees on the core salinity index from satellite remote sensing, obtaining a 1m resolution downscaled salinity index. This improves spatial resolution while ensuring computational efficiency for large-scale high-resolution grids. Furthermore, by combining topographic factors, meteorological factors, and soil depth interaction features, a three-dimensional prediction covariate dataset is constructed, and an extreme random tree nonlinear regression model is used to predict the three-dimensional spatial distribution of salinity in soda-alkali soil from 0 to 180 cm depth.

[0042] The method in the preferred embodiment of the present invention specifically includes the following steps: S100 acquires satellite multispectral remote sensing images, UAV RGB orthophoto images and UAV images containing near-infrared and red-edge bands, topographic data, meteorological data and measured sample data of salinity in layered soil profiles of the target study area, and performs registration, projection conversion and spatial resolution unification processing on various types of data to unify them to the target 1m reference grid.

[0043] Furthermore, the data preprocessing in step S100 includes geometric registration, projection transformation, and spatial resolution unification processing. The purpose is to unify satellite multispectral remote sensing images, UAV RGB orthophotos, UAV images containing near-infrared and red-edge bands, topographic data, meteorological data, and soil profile sample data into the same coordinate reference system and grid system to ensure the spatial consistency of subsequent surface-to-point spatial downscaling and three-dimensional prediction modeling.

[0044] Furthermore, geometric registration is achieved using existing feature matching and homography transformation methods. The geometric registration relationship is expressed as: In the formula, u i and v i Let u' represent the x and y coordinates of the i-th feature point in the source image, respectively; i and v' iLet x and y represent the x and y coordinates of the i-th corresponding point in the target image, respectively; H represents the homography transformation matrix from the source image to the target image. The homography transformation matrix can be obtained using existing ORB or SIFT feature matching methods combined with the RANSAC method; if the input data is already a standard orthorectified product, its existing geometric correction results can be directly used.

[0045] Furthermore, projection transformation is used to transform various types of data from the original coordinate reference system to the target coordinate reference system, and its expression is: In the formula, (x s ,y s (x) represents the planar coordinates in the original coordinate reference frame; t ,y t ) represents the planar coordinates in the target coordinate reference frame; T proj This represents the projection transformation function. This step can be implemented using existing projection transformation methods in GIS or remote sensing software.

[0046] Furthermore, the spatial resolution unification process employs bilinear interpolation for continuous raster variables such as satellite reflectivity, DEM, terrain factors, and meteorological raster. The pixel value for the desired location (x, y) is: In the formula, f(x,y) represents the interpolation result at the position (x,y) to be interpolated; x and y represent the x-coordinate and y-coordinate of the position to be interpolated, respectively; f 00 f 10 f 01 f 11 These represent the cell values ​​of the four adjacent raster nodes surrounding the location to be determined; This represents the normalized distance of the location to be determined in the horizontal direction; This represents the normalized distance of the location to be determined along the vertical axis.

[0047] S200: Based on the extraction of wideband features and gray-level co-occurrence matrix (GLCM) texture features from UAV images, and using the wideband features and GLCM texture features as point-scale spatial covariates, the core salinity index calculated from satellite multispectral remote sensing images is subjected to surface-to-point spatial downscaling (ATP-ET) based on extreme random trees. Under the dual constraints of regional conservation and physical extrema, a 1m resolution downscaled salinity index is obtained.

[0048] Furthermore, in step S200, a surface-to-point spatial downscaling method based on extreme random tree regression and bilinear interpolation is employed. This process mainly includes: point-scale covariate construction, point-scale trend prediction, regional-scale residual calculation, residual smoothing allocation, and physical extreme value constraint reconstruction. The core salinity index includes the improved normalized salinity index NDSI.Alkali The canopy response salinity index (CRSI) and the optimized soda salinity index (OSIAS) were used.

[0049] Furthermore, satellite pixels are used as regional support units, and the target 1m grid is used as point support units. Let the regional support unit be A, and the set of point support units contained within it be G(A).

[0050] Let X(g) be the scale covariate vector at position g∈G(A), and let the prediction value of the k-th extreme random tree for that point be... If the model contains K regression trees, then the point-scale trend prediction value is: In the formula, denoted by , X(g) represents the trend prediction value at point support unit g; X(g) represents the covariate vector composed of the broadband feature and GLCM texture feature at point support unit g; h k (⋅) represents the regression function of the k-th extreme random tree; K represents the number of regression trees.

[0051] Furthermore, the point-scale trend forecast values ​​are aggregated into regional support cells to obtain regional-scale trend values: In the formula, This represents the trend aggregation value of region support unit A; This represents the number of point support units within region support unit A. Let the observed core salinity index at the satellite pixel scale be z(A), then the region-scale trend residual r(A) is: Furthermore, bilinear interpolation is used to smooth the point-scale distribution of the regional-scale residuals. Let the residuals of the four adjacent regional support units surrounding point support unit g be r(A1), r(A2), r(A3), and r(A4), respectively. The bilinear distribution result is as follows: In the formula, The residual obtained by bilinear interpolation at point support cell g represents the result. This represents the weight of the point support unit g relative to the m-th neighboring region support unit. This represents the trend residual of the m-th adjacent support unit. To satisfy the region conservation constraint, the point-scale residuals within the same support unit are mean-corrected to ensure that their average value is consistent with the region residuals.

[0052] Furthermore, the point-scale trend prediction is added to the allocated point-scale residual, and a physical extremum constraint is applied to obtain the final 1m resolution downscaling salinity index: In the formula, The final downscaling salinity index at point support cell g; This represents the point-scale residual after correction for region conservation constraints; and These represent the upper and lower physical limits of the region support unit A, respectively, to avoid overshoot or oscillation in the reconstruction results.

[0053] Furthermore, the broadband features include the red, green, and blue band reflectance of the UAV RGB orthophotos, as well as the Normalized Difference Vegetation Index (NDVI) calculated based on the near-infrared and red bands.

[0054] Furthermore, the GLCM texture features include contrast, entropy, homogeneity, second moment of angle, correlation, and heterogeneity. The texture features are calculated within a 5×5 sliding window based on the green band of UAV imagery.

[0055] Step S300: Construction of the 3D Prediction Covariate Dataset. Based on the 1m resolution downscaled salinity index obtained in step S200, basic environmental variables are constructed by combining topographic factors, meteorological factors, and GLCM texture features. Soil depth is used as a continuous variable to construct the interaction features between environmental variables and depth, forming a covariate dataset for 3D prediction of soil salinity.

[0056] Furthermore, let the basic environmental covariate vector at position (x,y) be E(x,y), then we have: In the formula, E(x,y) represents the basic environmental covariate vector at location (x,y); Z(x,y) represents the downscaled salinity index vector; T(x,y) represents the topographic factor vector; M(x,y) represents the meteorological factor vector; and G(x,y) represents the GLCM texture feature vector.

[0057] Furthermore, let the soil depth variable be d, and the m-th feature in the foundation environmental covariates be E. m (x,y), then the corresponding deep interaction term feature I m (x,y,d) can be represented as: Furthermore, the interactive features include the interaction between downscaled salinity index and soil depth, the interaction between GLCM texture features and soil depth, and the interaction between meteorological and topographic factors and soil depth.

[0058] Furthermore, assuming a total of N sample records, the three-dimensional predictive covariate dataset can be represented as matrix X: In the formula, X represents the covariate dataset used for three-dimensional prediction of soil salinity; E i I represents the basic environmental covariate vector of the i-th sample; i Let N represent the deep interaction feature vector of the i-th sample; N represents the total number of samples.

[0059] Step S400: Construction of a three-dimensional prediction model for soil salinity. Using the covariate dataset constructed in step S300 as the independent variable and the measured sample data of salinity from stratified soil profiles as the dependent variable, an ExtraTrees regression model is constructed.

[0060] Furthermore, suppose the ExtraTrees model contains K regression trees, and the k-th regression tree predicts h for the input covariate vector x. k (x), then the overall prediction result of the model for: In the formula, This represents the predicted salt content for the input sample x; K represents the number of regression trees; h k (x) represents the prediction result of the k-th regression tree for sample x.

[0061] Furthermore, the extreme random tree nonlinear regression model does not perform bootstrap resampling, adopts the sqrt strategy for the number of split features, has 200-300 decision trees, and is trained and optimized through cross-validation.

[0062] S500, Three-dimensional prediction of soil salinity. Using the extreme random tree nonlinear regression model trained in step S400, the soil salinity at each soil depth from 0 to 180 cm in the target study area is predicted layer by layer, and the three-dimensional spatial distribution of soil salinity is obtained.

[0063] Furthermore, based on the covariate dataset constructed in step S300 and the measured sample data of salinity in the stratified soil profile, the present invention can also construct a quantile regression forest model and output the quantile interval of the soil salinity prediction results to quantify the prediction uncertainty.

[0064] Let the input covariate vector be X, and the quantile level be... The corresponding conditional quantile predicted value is defined as: In the formula, This indicates that the input sample X is at the quantile level. The conditional quantile prediction value; q represents the candidate prediction value; N represents the number of training samples; The y represents the forest weight of sample i with respect to input sample X; 1(⋅) represents the indicator function, which takes the value 1 when the condition in parentheses is true, and 0 otherwise; iThis represents the measured salt content of the i-th sample.

[0065] Example 3 This embodiment focuses on a typical soda saline-alkali soil region. The soil salinity in this area is predominantly sodium carbonate and sodium bicarbonate, with a pH value ranging from 9.2 to 10.8. Surface cover types include Suaeda salsa communities, exposed soda saline-alkali patches, trees, farmland, artificial features, and water bodies, covering an area of ​​approximately 4.56 km². 2 The surface of this area exhibits a typical mosaic pattern of "Suaeda salsa-salt patches," with the diameter of the salt patches mostly ranging from 1 to 5 meters. The mixed pixel effect is significant in 10-meter resolution satellite images, making it suitable as a verification area for the method of this invention.

[0066] Step S1: Data Acquisition and Preprocessing (1) Satellite remote sensing data: Sentinel-2A multispectral remote sensing images were acquired from July to September 2025, with cloud and snow coverage of less than 5%, and containing 10 spectral bands. Radiometric calibration and atmospheric correction were performed on the original images, and the spatial resolution was unified to 10 meters.

[0067] (2) UAV imagery data: UAV images of the study area were acquired in July 2025, including RGB three-band orthophotos with a spatial resolution of 0.05m, as well as UAV images containing near-infrared and red-edge bands. The UAV images were radiometrically calibrated and orthorectified, and then resampled to a resolution of 1m as the target reference grid.

[0068] (3) Topographic data: The digital elevation model of the study area was obtained, and after resampling, it was unified to a resolution of 1m, and the slope, aspect and topographic humidity index TWI were derived.

[0069] (4) Meteorological data: The average annual temperature and annual precipitation data of the study area were obtained and spatially interpolated to generate 1m resolution raster data.

[0070] (5) Soil profile sample data: Six soil profile sampling points were set up in the study area. At each sampling point, stratified sampling was carried out at depths of 0–10 cm, 10–20 cm and up to 170–180 cm, for a total of 18 depth layers, resulting in 108 soil samples. The soil electrical conductivity of each sample was measured in the laboratory and used as an indicator of salinity. Outliers were removed.

[0071] Step S2: Spatial downscaling of the core salinity index (1) Calculation of satellite core salinity index. Various spectral indices, including MTCI, NDWI, MSI, OSAVI, and NDSI, were calculated from Sentinel-2A multispectral remote sensing images. Alkali OSI, CRSI, and OSIAS. Among them, NDSI... AlkaliCRSI and OSIAS are used as core salinity indices for subsequent surface-to-point spatial downscaling.

[0072] (2) UAV covariate extraction. Two types of covariates were extracted from UAV imagery resampled to 1m resolution: The first category is broadband features, including RGB three-band reflectivity and NDVI calculated based on near-infrared and red bands; The second category is GLCM texture features, based on the green band, which calculates contrast, entropy, homogeneity, second angular moment, correlation, and heterogeneity within a 5×5 pixel sliding window. Salt-alkali patch areas typically exhibit high homogeneity and low entropy, while Suaeda salsa areas typically exhibit high contrast and high entropy.

[0073] (3) Spatial downscaling from surface to point based on extreme random trees. 10m×10m pixels of Sentinel-2A imagery are used as regional support units, and 1m×1m grids are used as point support units. First, an extreme random tree nonlinear regression model is established between the core salinity index and the wideband features of the UAV and the GLCM texture features to obtain the 1m point-scale trend prediction value. Second, the 1m trend prediction value is aggregated to the 10m scale to calculate the prediction residual of the satellite pixel. Then, bilinear interpolation is used to robustly and smoothly distribute the 10m prediction residual, and under the constraint of regional conservation, it is redistributed to 100 1m×1m grid units within the pixel. Finally, the point-scale trend prediction value is added to the redistributed point-scale residual, and a physical extreme value clamping mechanism is introduced. The extreme value of the original 10m satellite image is used as the basis for boundary constraints to eliminate the numerical overshoot and oscillation phenomenon that is prone to occur during spatial downscaling. Finally, a high-fidelity 1m resolution downscaled salinity index with rigorous physical meaning is obtained.

[0074] In this embodiment, NDSI Alkali CRSI and OSIAS respectively performed the above downscaling process to form a 1m resolution downscaled salinity index dataset.

[0075] Step S3: Construction of the 3D Predictive Covariate Dataset Based on a downscaled salinity index at 1m resolution, a set of basic environmental variables was constructed by combining topographic factors, meteorological factors, and GLCM texture features at the same resolution. Soil depth was set to 0–180cm, with each layer lasting 10cm. For each depth, interactive features such as downscaled salinity index × depth, GLCM texture features × depth, and topographic factors × depth were constructed to form a covariate dataset for three-dimensional prediction of soil salinity. The final covariate dataset contained approximately 60 features.

[0076] Step S4: Construction of a three-dimensional prediction model for soil salinity One hundred and eighty soil samples were randomly divided into a training set and a validation set at a ratio of 80%:20%, with 86 samples in the training set and 22 samples in the validation set. Using the covariate dataset constructed in step S3 as the independent variable and soil electrical conductivity as the dependent variable, an extreme random tree nonlinear regression algorithm was used to construct a three-dimensional prediction model for soil salinity. Model parameters were set as follows: 200 decision trees, sqrt maximum feature count, and no bootstrapping resampling. Ten-fold cross-validation was used for model training and evaluation, and the coefficient of determination R0 was used as the statistical measure. 2 The root mean square error (RMSE) and Lin's coherent correlation coefficient (LCCC) were used as evaluation indicators.

[0077] Step S5: Three-dimensional prediction of soil salinity Using the trained extreme random tree nonlinear regression model, the depth of each soil layer in the 0-180cm range of the study area was predicted layer by layer, and the three-dimensional spatial distribution of soil salinity was obtained.

[0078] In this embodiment, the surface salt content ranges from 2.5 to 8.6 g / kg, with a peak salt accumulation at 40 to 60 cm depth followed by a gradual decrease to 0.8 to 2.2 g / kg at the bottom layer. This generally conforms to the vertical distribution pattern of soda saline-alkali soil: "surface accumulation – mid-layer sedimentation – bottom stability." The 1m resolution prediction results can clearly depict the boundaries and spatial distribution characteristics of saline-alkali patches with diameters ranging from 1 to 5 m.

[0079] To verify the effectiveness of the surface-to-point spatial downscaling method of this invention, the following comparative experiment was conducted: Control group: A covariate dataset was constructed using the 10m resolution satellite core salinity index without downscaling, and the same extreme random tree nonlinear regression model was used for three-dimensional prediction; Experimental group: The covariate dataset was constructed using the 1m resolution core salinity index downscaled according to the present invention, and the same extreme random tree nonlinear regression model was used for three-dimensional prediction.

[0080] Experimental results show that the coefficient of determination R of the experimental group on the validation set is... 2 The R value reached 0.73, the root mean square error (RMSE) was 1.48 g / kg, and Lin's concordance coefficient (LCCC) was 0.85; while the control group R value was 0.73. 2 The experimental group had a concentration of 0.61, an RMSE of 1.96 g / kg, and an LCCC of 0.74. Compared to the control group, the experimental group had a lower R value. 2 The accuracy was improved by 0.12, and the RMSE decreased by 0.48 g / kg. This indicates that the present invention effectively improves the accuracy of three-dimensional prediction of salinity in soda-alkali soil by using surface-to-point spatial downscaling constrained by UAV texture constraints.

[0081] Furthermore, feature importance analysis revealed that downscaling CRSI exponent × depth interaction term, GLCM entropy texture × depth interaction term, and NDSI... Alkali The index × depth interaction term ranks among the top three in importance, indicating that the downscaling salinity index and its coupling characteristics with depth play an important role in 3D prediction.

[0082] Building upon this embodiment, a quantile regression forest model can be further established based on the covariate dataset and stratified soil profile sample data constructed in step S3, with quantiles set to 0.05 and 0.95 to output the quantile intervals for soil salinity prediction. Areas with wider intervals indicate higher prediction uncertainty and can be prioritized for subsequent supplementary sampling.

[0083] This invention integrates macroscopic spectral information from satellite remote sensing with fine-scale spatial features from UAVs, and utilizes extreme random trees to perform surface-to-point spatial downscaling, effectively reducing monitoring errors caused by mixed pixels in satellite imagery. It fully leverages the advantages of UAV data in representing fine-scale spatial structures, accurately identifying the distribution of small-scale saline-alkali patches. Simultaneously, it constructs interactive features between environmental variables and soil depth, establishing a predictive dataset adapted to vertical salinity distribution patterns. This overcomes the limitations of traditional methods that only predict surface salinity, achieving accurate prediction of salinity at different soil depths. Furthermore, the nonlinear modeling approach closely matches the complex spatial distribution characteristics of soda saline-alkali soil regions, simplifying the computation process for massive grid data. It balances prediction accuracy with overall computational efficiency and also performs quantitative analysis of the uncertainty of prediction results, clarifying the credibility range of the prediction results. This provides comprehensive and reliable data for practical work such as precise improvement of saline-alkali land, optimized allocation of water resources, and regional ecological restoration, possessing significant practical application and engineering value.

[0084] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0085] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, it should be understood that the sequence number of each step in the above embodiments does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention. The actions or steps recorded in the claims can be performed in a different order than that in the above embodiments and can still achieve the desired result. In addition, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0086] Example 4 Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides a three-dimensional prediction system for salinity of soda-alkali soil based on UAV texture constraints, the system comprising: The data acquisition and preprocessing module is used to collect multi-source data and measured sample data of soil salinity at different depths in the target study area, complete geometric registration, projection transformation and spatial resolution unification processing, and integrate them into a reference grid with preset accuracy. The spatial downscaling module is used to extract wideband features and gray-level co-occurrence matrix texture features of UAVs, perform surface-to-point spatial downscaling, and output the downscaling salinity index with a preset accuracy of the reference grid resolution. The 3D covariate construction module is used to fuse downscaled salinity index, topographic factors, meteorological factors and gray-level co-occurrence matrix texture features to generate environmental variables and soil depth interaction features, and output a 3D prediction covariate dataset of soil salinity at different depths. The 3D prediction module is used to build and train an extreme random tree nonlinear regression model to predict soil salinity at different depths layer by layer, and output the 3D spatial distribution results of soil salinity at different depths with a preset accuracy benchmark grid resolution. The uncertainty quantification module is used to construct a quantile regression forest model based on covariate datasets and measured soil salinity samples at different depths, outputting the predicted quantile intervals for salinity and quantifying the prediction uncertainty. The system described in the above embodiments is used to implement a three-dimensional prediction method for salinity of soda saline-alkali soil based on UAV texture constraints in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0087] It should be noted that the above-mentioned three-dimensional prediction method for salinity of soda-alkali soil based on UAV texture constraints is embodied in the form of functional units. The term "module" here can be implemented in software and / or hardware, without specific limitations.

[0088] Example 5 Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a computer-readable storage medium storing computer instructions for causing the computer to execute a three-dimensional prediction method for salinity of soda-alkali soil based on UAV texture constraints as described in any of the above embodiments.

[0089] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0090] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute a three-dimensional prediction method for salinity of soda saline-alkali soil based on UAV texture constraints as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0091] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0092] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuitry) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this disclosure may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0093] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0094] Therefore, the units of the various examples described in the embodiments of this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0095] This disclosure is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A three-dimensional prediction method for salinity of soda-alkali soil based on UAV texture constraints, characterized in that, The method includes the following steps: S100 collects satellite multispectral remote sensing images, UAV RGB orthophoto images, UAV multispectral images including near-infrared and red edge bands, digital terrain data, meteorological data, and measured soil profile salinity data at different depths in the target study area. After preprocessing all types of data, they are integrated into a benchmark grid with a preset accuracy. S200: Based on the extraction of wideband features and texture features of gray-level co-occurrence matrix from UAV images, these are used as point-scale fine spatial covariates. The surface-to-point spatial downscaling method based on extreme random trees is used, combined with regional conservation constraints and physical extreme value constraints, to perform spatial downscaling on the core salinity index. Finally, a downscaled salinity index with the same accuracy as the benchmark grid is reconstructed. S300: Based on the downscaled salinity index, the basic environmental variables are constructed by combining topographic factors, meteorological factors and gray-level co-occurrence matrix texture features. Soil depth is used as a continuous variable to construct the interaction features between environmental variables and soil depth, forming a covariate dataset for three-dimensional prediction of soil salinity. S400. Using the covariate dataset as the independent variable and the measured sample data of salinity in the stratified soil profile as the dependent variable, an extreme random tree nonlinear regression model is constructed, and model training and parameter optimization are completed through cross-validation. S500 uses an optimized extreme random tree nonlinear regression model to predict soil salinity at different depths within the target study area, outputting a three-dimensional spatial distribution of soil salinity consistent with a preset high-precision benchmark grid.

2. The method according to claim 1, characterized in that, In step S200, the surface-to-point spatial downscaling method based on extreme random trees specifically includes the following steps: Using satellite remote sensing image pixels as regional support units and a reference grid with preset precision as point support units, the spatial correspondence between regional support units and point support units is clearly defined. Using the core salinity index as the target variable and the broadband features and GLCM texture features extracted from UAV images as covariates, an extreme random tree nonlinear regression model was established to obtain point-scale trend prediction values. The trend residual at the region scale is calculated, and bilinear interpolation is used to smoothly distribute the region-scale residual to the point-scale grid cells under the region conservation constraint. The point-scale trend prediction value is added to the allocated point-scale residual, and the synthesis result is bounded based on the physical extrema of the original region support cell to obtain a downscaled salinity index with the same resolution as the baseline grid.

3. The method according to claim 2, characterized in that, In step S200, the construction of the core salinity index specifically includes the following steps: Select characteristic bands in satellite multispectral remote sensing images that are sensitive to the salinity of soda saline-alkali soil; Based on the spectral response characteristics of soda saline-alkali soil, multiple salt sensitivity indices were constructed. The salt sensitivity indices include the improved normalized salt index, the canopy-response salt index, and the optimized soda saline soil salt index.

4. The method according to claim 2, characterized in that, In step S200, the extraction of texture features based on the gray-level co-occurrence matrix specifically includes the following steps: Select bands in UAV imagery that are sensitive to soil salinity distribution; Set the sliding window to the appropriate size; Texture features are extracted within the sliding window using the gray-level co-occurrence matrix algorithm. The texture features include contrast, entropy, homogeneity, second moment of angle, correlation, and heterogeneity.

5. The method according to claim 2, characterized in that, The extraction of broadband features in step S200 specifically includes the following steps: Extract multi-band reflectance features from UAV images, the multi-band including RGB band and near-infrared band; Vegetation-related indices are constructed by combining and calculating reflectance from different wavebands. The broadband features include the multi-band reflectance features and vegetation-related indices.

6. The method according to claim 1, characterized in that, The construction of interactive features in step S300 specifically includes the following steps: Clarify the coupling relationship between basic environmental variables and soil depth; Through mathematical correlation operations, the interaction characteristics between basic environmental variables and soil depth are constructed; The interactive features include the interaction between downscaled salinity index and soil depth, the interaction between gray-level co-occurrence matrix texture features and soil depth, the interaction between meteorological factors and soil depth, and the interaction between topographic factors and soil depth.

7. The method according to claim 1, characterized in that, The construction and optimization of the extreme random tree nonlinear regression model in step S400 specifically includes the following steps: Set model configuration strategies, including non-bootstrapping resampling sample sampling strategies, adaptive feature selection strategies based on the total number of covariates, and the number of models ensembled to balance prediction accuracy and computational efficiency. Using the covariate dataset as the independent variable and the measured soil salinity data at different depths as the dependent variable, an extreme random tree nonlinear regression model was initialized. Cross-validation was used to train and optimize the model.

8. The method according to claim 1, characterized in that, It also includes the step of quantifying the uncertainty of prediction: Based on the covariate dataset constructed in step S300 and the measured sample data of soil salinity at different depths, a quantile regression forest model is constructed. Set the corresponding quantile level; The prediction results of soil salinity at different depths are output in the prediction intervals at each quantile level, thereby achieving a quantitative characterization of prediction uncertainty.

9. A three-dimensional prediction system for salinity of soda-alkali soil based on UAV texture constraints, the system being used to implement the method described in any one of claims 1-8, characterized in that, The system includes The data acquisition and preprocessing module is used to collect multi-source data and measured sample data of soil salinity at different depths in the target study area, complete geometric registration, projection transformation and spatial resolution unification processing, and integrate them into a reference grid with preset accuracy. The spatial downscaling module is used to extract wideband features and gray-level co-occurrence matrix texture features of UAVs, perform surface-to-point spatial downscaling, and output the downscaling salinity index with a preset accuracy of the reference grid resolution. The 3D covariate construction module is used to fuse downscaled salinity index, topographic factors, meteorological factors and gray-level co-occurrence matrix texture features to generate environmental variables and soil depth interaction features, and output a 3D prediction covariate dataset of soil salinity at different depths. The 3D prediction module is used to build and train an extreme random tree nonlinear regression model to predict soil salinity at different depths layer by layer, and output the 3D spatial distribution results of soil salinity at different depths with a preset accuracy benchmark grid resolution. The uncertainty quantification module is used to construct a quantile regression forest model based on covariate datasets and measured sample data of soil salinity at different depths, output the salinity prediction quantile interval, and quantify the prediction uncertainty.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the method as described in any one of claims 1 to 8.

Images