A method and system for constructing a surface soil moisture data set

Abstract

This invention discloses a method and system for constructing a surface soil moisture content dataset, belonging to the field of data processing technology. In each sub-region, surface soil moisture content from UAV imagery is acquired, and the normalized vegetation index (NDI) of the UAV imagery is extracted. The dispersion of the NDI of effective UAV pixels within the spatial coverage area of ​​each satellite pixel is determined and used as the pixel value of the corresponding satellite pixel. A raster layer is generated according to the spatial reference and resolution of the satellite imagery to obtain a sub-pixel heterogeneous raster. The presence or absence of valid values ​​in the sub-pixel heterogeneous raster is determined based on whether the number of effective UAV pixels within each satellite pixel is greater than or equal to a preset threshold. Heterogeneous and non-heterogeneous regression models are then constructed and trained. The prediction results corresponding to heterogeneous and non-heterogeneous regions in each sub-region are fused from the model output to construct a surface soil moisture content dataset. This method can effectively improve the model's cross-plot generalization ability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a method and system for constructing a dataset of surface soil moisture content. Background Technology

[0002] Soil moisture content is a key state variable in land surface hydrothermal coupling and ecological processes, and plays an important role in monitoring alpine grassland degradation, drought assessment, vegetation growth and ecological security evaluation.

[0003] Existing methods for acquiring surface soil moisture content data are mostly based on training satellite inversion models using ground-based or limited-range UAV observations. These models extract multi-dimensional environmental features such as normalized vegetation index and surface temperature from the satellite, train regression models, and extrapolate regional moisture content. Alternatively, they can improve resolution by downscaling and fusing multi-source images.

[0004] However, existing technologies have the following drawbacks: On the one hand, due to the sparse ground points and limited coverage of UAVs, it is impossible to form a continuous regional dataset and a large-area continuous label dataset covering different land surface types and humidity gradients, resulting in insufficient training samples; on the other hand, although satellite inversion covers a wide range, it is affected by spatial resolution, sensor response differences, atmospheric and cloud pollution, scale effects, etc. Directly applying the UAV domain model to the satellite domain will result in inconsistent input feature distribution due to scale / sensor differences, causing domain drift and system bias, resulting in poor model generalization ability across regions / time periods. Summary of the Invention

[0005] To address the problems existing in the above-mentioned fields, this invention proposes a method and system for constructing a surface soil moisture content dataset. It not only fully leverages the advantages of UAV high-resolution data in characterizing microscopic heterogeneity, but also utilizes satellite imagery to achieve coverage of the entire target area, effectively mitigating domain drift and system bias in cross-scale inversion. It solves the problem that traditional methods cannot simultaneously achieve high-resolution accuracy and large-scale coverage, and ultimately outputs a high-precision, fully covered, and quantifiable surface soil moisture content dataset.

[0006] To address the aforementioned technical problems, this invention discloses a method for constructing a surface soil moisture content dataset, comprising the following steps: Simultaneously acquire ground-measured soil moisture content data, UAV imagery, and satellite imagery of the target area, and divide the target area into multiple sub-regions; In each sub-region, the surface soil moisture content of the UAV imagery is acquired and aggregated to the satellite pixel scale as a label value, while the normalized vegetation index of the UAV imagery is extracted. Based on the normalized vegetation index of UAV imagery, the dispersion of the normalized vegetation index of effective UAV pixels within the spatial coverage area of ​​each satellite pixel is determined; the dispersion is used as the pixel value of the corresponding satellite pixel, and a raster layer is generated according to the spatial reference and resolution of the satellite imagery to obtain a sub-pixel heterogeneous raster. When the number of valid UAV pixels in each satellite pixel is greater than or equal to a preset threshold, the sub-pixel heterogeneous raster is determined to have a valid value. The multidimensional environmental features extracted from the satellite image and the pixel value are used as inputs, and the label value is used as output to construct and train a heterogeneous region regression model. When the number of valid UAV pixels is less than the preset threshold, the sub-pixel heterogeneous raster is determined to have no valid value. The multidimensional environmental features are used as inputs, and the label value is used as output to construct and train a non-heterogeneous region regression model. Based on the trained regression model, the prediction results of heterogeneous and non-heterogeneous areas in each sub-region are obtained and spliced ​​together to construct a surface soil moisture content dataset covering the entire target area.

[0007] Preferably, in each sub-region, the surface soil moisture content of the UAV image is acquired and aggregated to the satellite pixel scale as a label value, and the normalized vegetation index of the UAV image is extracted, specifically including: Extract the normalized vegetation index and surface temperature from the UAV imagery of the current sub-region; The temperature-vegetation drought index is determined by constructing a characteristic triangular space between the normalized vegetation index and the land surface temperature. Using the measured soil moisture content data of the current sub-region, the temperature-vegetation-drought index is regressed and calibrated to generate the surface soil moisture content of the current sub-region. The surface soil moisture content at the UAV pixel scale is aggregated to the satellite pixel scale and used as the label value at the satellite pixel scale.

[0008] Preferably, the normalized vegetation index based on UAV imagery, which determines the dispersion of the normalized vegetation index of the effective UAV pixels within the spatial coverage area of ​​each satellite pixel, specifically includes: The effective drone pixels are drone pixels whose normalized vegetation index (NVI) values ​​are not empty and whose NVI values ​​are within a preset range. Extract the normalized vegetation index values ​​of all valid UAV pixels within the spatial coverage area of ​​the satellite pixel to form the index set of the satellite pixel; Obtain the standard deviation of the index set as the degree of dispersion; For satellite pixels that do not contain valid drone pixels, assign an invalid identifier in the sub-pixel heterogeneity raster.

[0009] Preferably, when the number of valid UAV pixels within each satellite pixel is greater than or equal to a preset threshold, it is determined that the sub-pixel heterogeneous raster has a valid value. The heterogeneous region regression model is constructed and trained using multi-dimensional environmental features and pixel values ​​extracted from satellite imagery as input and the label value as output. Specifically, this includes: When the number of valid UAV pixels within a satellite pixel is greater than or equal to a preset threshold, it is determined that the sub-pixel heterogeneous raster corresponding to the satellite pixel has a valid value, indicating that the satellite pixel belongs to a heterogeneous region. The preset threshold is determined based on the spatial resolution ratio of satellite pixels and UAV pixels and the preset minimum spatial coverage, or based on the minimum statistical sample size required for the degree of dispersion. In heterogeneous regions, multiple candidate regression models for heterogeneous regions are constructed by using multidimensional environmental features extracted from satellite imagery and pixel values ​​of sub-pixel heterogeneous rasters as input features and the label values ​​as output. The multidimensional environmental features include normalized vegetation index, surface temperature, digital elevation model, slope, topographic humidity index and soil texture index, and the candidate regression models include random forest regression model, extreme random tree regression model and histogram gradient boosting regression model. Obtain the comprehensive score for each candidate regression model based on the median R² and stability penalty; Using sub-regions as the grouping unit, one sub-region is used as the test set and the remaining sub-regions are used as the training set. Leave-one-out sample cross-validation is used to obtain the leave-one-out R² of each candidate regression model. The surface soil moisture content of the current sub-region is divided into several blocks according to spatial proximity. Spatial block cross-validation is used to obtain the spatial block R² of each candidate regression model. Based on the leave-one-out R², spatial block R², and comprehensive score of each candidate regression model, the optimal regression model for the heterogeneous region is determined. The optimal regression model for heterogeneous regions is used as the regression model to be constructed and trained.

[0010] Preferably, when the number of effective UAV pixels is less than a preset threshold, it is determined that the sub-pixel heterogeneous raster has no valid value. Using multi-dimensional environmental features as input and the label value as output, a non-heterogeneous region regression model is constructed and trained, specifically including: When the number of valid UAV pixels in a satellite pixel is less than a preset threshold, it is determined that the sub-pixel heterogeneous grid corresponding to the satellite pixel has no valid value, indicating that the satellite pixel belongs to a non-heterogeneous area. In non-heterogeneous regions, multiple candidate regression models for non-heterogeneous regions are constructed using multidimensional environmental features as input features and the label values ​​as output. k-fold cross-validation or leave-one-out cross-validation was used to evaluate the predictive performance of candidate regression models in non-heterogeneous regions and obtain the corresponding coefficient of determination R² and root mean square error RMSE. The optimal regression model for the non-heterogeneous region is determined based on the coefficient of determination R² and the root mean square error RMSE. The optimal regression model in the non-heterogeneous region is used as the regression model to be constructed and trained.

[0011] Preferably, the process of constructing and training the regression model further includes equalization processing of the label value samples at the satellite pixel scale of the current sub-region, specifically including: The sample size of the label values ​​at the satellite pixel scale for each sub-region is counted, and sub-regions with sample sizes below the median are upsampled with replacement.

[0012] Preferably, the process of constructing and training the regression model further includes feature selection of multidimensional environmental features based on leave-one-out cross-validation, specifically including: The contribution of each environmental feature in the multidimensional environmental features is assessed by permutation importance evaluation. A subset of key environmental features is retained according to the cumulative contribution rate threshold, while redundant environmental features are eliminated.

[0013] Preferably, the splicing and fusion involves splicing the prediction results of regression models corresponding to heterogeneous and non-heterogeneous regions of each sub-region using a mask to generate a surface water content dataset covering the entire target region.

[0014] Preferably, the simultaneous acquisition of ground-measured soil moisture content data, UAV imagery, and satellite imagery of the target area further includes: performing atmospheric correction and cloud removal processing on the acquired data.

[0015] Preferably, it also includes a system for constructing a surface soil moisture content dataset, comprising: The data acquisition module is used to simultaneously acquire ground-measured soil moisture content data, UAV imagery, and satellite imagery of the target area, and to divide the target area into multiple sub-areas; The training sample generation module is used to acquire the surface soil moisture content of UAV images in each sub-region and aggregate it to the satellite pixel scale as a label value, while extracting the normalized vegetation index of the UAV images. Based on the normalized vegetation index of the UAV images, the dispersion of the normalized vegetation index of the effective UAV pixels contained within the spatial coverage of each satellite pixel is determined. The dispersion is used as the pixel value of the corresponding satellite pixel, and a raster layer is generated according to the spatial reference and resolution of the satellite images to obtain a sub-pixel heterogeneous raster. The regional modeling module is used to determine that sub-pixel heterogeneous gratings have valid values ​​when the number of valid UAV pixels in each satellite pixel is greater than or equal to a preset threshold. It uses multi-dimensional environmental features extracted from satellite images and the pixel values ​​as inputs, and the label values ​​as outputs to construct and train a heterogeneous region regression model. When the number of valid UAV pixels is less than the preset threshold, it determines that sub-pixel heterogeneous gratings have no valid values. It uses multi-dimensional environmental features as inputs and the label values ​​as outputs to construct and train a non-heterogeneous region regression model. The dataset construction module is used to obtain the prediction results of heterogeneous and non-heterogeneous areas of each sub-region based on the trained regression model, and then splice and merge them to construct a surface soil moisture content dataset covering the entire target area.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The proposed method for constructing a surface soil moisture content dataset aggregates surface soil moisture content from UAV imagery to satellite pixel scale as label values, achieving spatial alignment of training samples across scales. Simultaneously, based on the dispersion determined by the UAV's normalized vegetation index, it can accurately quantify the spatial heterogeneity within satellite pixels, fully leveraging the advantages of high-resolution UAV data in representing microscopic heterogeneity and compensating for the blind spots of single macroscopic satellite data. By objectively determining the validity of sub-pixel heterogeneity raster and implementing partitioning based on whether the number of effective UAV pixels within each satellite pixel is greater than or equal to a preset threshold, this mechanism ensures the reliability of dispersion statistics (avoiding interference from invalid pixels) and provides a scientific prior foundation for subsequent differentiated modeling. In heterogeneous regions, multidimensional environmental features extracted from satellite imagery and pixel values ​​from sub-pixel heterogeneous rasters are used as inputs. These pixel values ​​compensate for the inadequacy of macroscopic environmental features in representing the complex spatial structure within pixels, effectively improving the prediction accuracy for complex surface areas. In non-heterogeneous regions, only multidimensional environmental features are used as input, avoiding the introduction of redundant microscopic heterogeneous information and prediction distortion caused by forced extrapolation of local models, ensuring global deployability and generalization ability. This method fully leverages the advantages of high-resolution UAV data in representing microscopic heterogeneity while utilizing satellite imagery to achieve macroscopic coverage of the entire target area. Label aggregation and partitioned differential modeling effectively mitigate domain drift and system bias in cross-scale inversion, solving the problem of traditional methods struggling to balance high-resolution accuracy and large-scale coverage. Ultimately, it outputs a high-precision, fully covered, and quantifiable dataset of surface soil moisture content. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for constructing a surface soil moisture content dataset proposed in this invention; Figure 2 A detailed flowchart of the partitioning modeling strategy provided in this embodiment of the invention; Figure 3 This is a spatial map of land surface temperature (LST) and normalized difference vegetation index (NDVI) provided in an embodiment of the present invention. Figure 4 A multivariate georegression plot showing the relationship between the Temperature-Vegetation-Drought Index (TVDI) and surface soil moisture content, provided for an embodiment of the present invention. Figure 5 This is a map showing the distribution of surface soil moisture content within the UAV sample plot area provided in this embodiment of the invention. Figure 6 This is a schematic diagram of multidimensional environmental feature extraction provided in an embodiment of the present invention; Figure 7 A comparison chart of training and fitting results of different models provided in the embodiments of the present invention in heterogeneous partitions; Figure 8 This is a comparison chart of multi-model cross-validation scores provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the partitioning modeling strategy provided in an embodiment of the present invention; Figure 10 A spatial distribution diagram of total uncertainty provided for embodiments of the present invention; Figure 11 A reliability grading result diagram provided for an embodiment of the present invention; Figure 12 Spatial distribution characteristics of surface soil moisture content in the formal results area provided for embodiments of the present invention; Figure 13 The image shows the prediction results of the optimal soil surface water content across the entire region, as provided in this embodiment of the invention. Detailed Implementation

[0018] The following will refer to the appendices in the embodiments of the present invention. Figures 1-13 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0019] Example like Figure 1 As shown, this invention proposes a method for constructing a surface soil moisture content dataset, which includes the following steps: S1: Simultaneously acquire ground-measured soil moisture content data, UAV imagery, and satellite imagery of the target area, and divide the target area into multiple sub-regions; S2: In each sub-region, the surface soil moisture content of the UAV image is acquired and aggregated to the satellite pixel scale as a label value. At the same time, the normalized vegetation index of the UAV image is extracted. S3: Based on the normalized vegetation index of UAV imagery, determine the dispersion of the normalized vegetation index of effective UAV pixels within the spatial coverage area of ​​each satellite pixel; use the dispersion as the pixel value of the corresponding satellite pixel, generate a raster layer according to the spatial reference and resolution of the satellite imagery, and obtain a sub-pixel heterogeneous raster. S4: When the number of valid UAV pixels in each satellite pixel is greater than or equal to a preset threshold, the sub-pixel heterogeneous raster is determined to have a valid value. The multidimensional environmental features and pixel values ​​extracted from the satellite image are used as inputs, and the label values ​​are used as outputs to construct and train a heterogeneous region regression model. When the number of valid UAV pixels is less than the preset threshold, the sub-pixel heterogeneous raster is determined to have no valid values. The multidimensional environmental features are used as inputs, and the label values ​​are used as outputs to construct and train a non-heterogeneous region regression model. S5: Based on the trained regression model, obtain the prediction results of heterogeneous and non-heterogeneous areas of each sub-region and splice and merge them to construct a surface soil moisture content dataset covering the entire target area.

[0020] In step S1, at the ground reference layer: the present invention divides the target area into multiple sub-regions, selects the current sub-region as the sample plot, and obtains the measured soil moisture content data of the current sub-region.

[0021] At the UAV bridge layer: Simultaneously acquire high-resolution thermal infrared and multispectral images of the current sub-region, and perform radiometric calibration, geometric correction and resampling on them to extract the normalized vegetation index (NDVI) and land surface temperature (LST) at the UAV scale for the current sub-region.

[0022] In the satellite overlay: Satellite images of the current sub-region are acquired synchronously. After atmospheric correction and cloud removal, the satellite normalized vegetation index (NDVI), land surface temperature (LST), digital elevation model (DEM), slope, topographic humidity index (TWI), and soil texture index (STI) of the current sub-region are extracted as multidimensional environmental features.

[0023] In step S2, based on the extracted Normalized Difference Vegetation Index (NDVI) and Land Surface Temperature (LST) of the UAV imagery, a characteristic triangular space (NDVI-T) for the NDVI and LST of the UAV imagery is constructed. s (Characteristic triangular space), the temperature and vegetation drought index is determined as follows: ; Among them, TVDI is the Temperature-Vegetation Aridity Index, with a value ranging from 0 to 1; the closer the TVDI value is to 1, the higher the degree of drought; the closer the TVDI value is to 0, the higher the degree of surface moisture; T s The surface temperature of a pixel reflects the energy balance state of the Earth's surface. The minimum surface temperature under the Normalized Difference Vegetation Index (NDVI) condition corresponds to a surface state with sufficient moisture (humid endmember). The maximum surface temperature under the Normalized Difference Vegetation Index (NDVI) is the surface temperature corresponding to the water deficit (drought endmember) condition.

[0024] By using ground-measured soil moisture content data, the Temperature-Vegetation-Drought Index (TVDI) is regressed and calibrated to generate a surface soil moisture content distribution map of the current sub-region's UAV imagery. Based on the surface soil moisture content distribution map, the surface soil moisture content is discretized and used as a label value at the satellite pixel scale, i.e., a massive number of "pixel-level training samples" to replace the traditional sparse ground points for subsequent model construction and training processes.

[0025] Multidimensional environmental features are extracted from satellite imagery, including Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), Digital Elevation Model (DEM), Slope, Topographic Moisture Index (TWI), and Soil Texture Index (STI). Specifically, NDVI reflects vegetation cover, LST reflects surface energy balance, DEM and Slope reflect the influence of topography on water transport, TWI characterizes the impact of topography on soil moisture retention, and STI comprehensively reflects the interaction between soil type and topography. These six features constitute a multidimensional environmental profile, providing globally available satellite remote sensing data that can be used independently without relying on UAV data.

[0026] In step S3, partitioning modeling is performed based on the sub-pixel heterogeneity features of UAV imagery.

[0027] The process of generating heterogeneous gratings provided in this embodiment specifically includes: Based on the spatial reference and resolution of satellite imagery, for each satellite pixel, the normalized vegetation index (NDI) value of the effective UAV pixels contained within its spatial coverage area is statistically analyzed; the standard deviation, variance, or coefficient of variation of the NDI value of the effective UAV pixels is calculated as the degree of dispersion. Using the standard deviation as the pixel value of the corresponding satellite pixel, a raster layer is generated according to the spatial reference and resolution of the satellite image to obtain a sub-pixel heterogeneous raster. For satellite pixels that do not contain valid drone pixels, assign an invalid identifier in the sub-pixel heterogeneity raster.

[0028] Among them, effective drone pixels refer to the pixels selected after preprocessing the drone imagery before calculating the degree of dispersion.

[0029] In this embodiment, a valid UAV pixel refers to a UAV pixel whose Normalized Difference Vegetation Index (NDV) value is not empty and is within a preset range. Specifically, it excludes pixels with no data (i.e., an empty NDC) caused by sensor malfunction or gaps in flight path stitching, as well as abnormal noise pixels whose NDC values ​​exceed physically reasonable ranges (e.g., less than -0.1 or greater than 1) due to cloud and snow shadows, water bodies, or residual errors in atmospheric correction.

[0030] Only the normalized vegetation index of these valid pixels is included in the standard deviation calculation to ensure that the calculated dispersion can truly and objectively reflect the real spatial heterogeneity of the surface physical properties within the satellite pixels. This avoids serious distortion of statistics due to missing local data or extreme abnormal noise, thereby preventing pixels from being misclassified as heterogeneous areas and ensuring the reliability of the input features of the subsequent heterogeneous area regression model and the accuracy of the final soil moisture content prediction.

[0031] When the number of valid UAV pixels within a satellite pixel is less than a preset threshold, due to insufficient sample size, a reliable dispersion statistic cannot be generated. Therefore, it is determined that the heterogeneous raster corresponding to the satellite pixel has no valid value, indicating that the satellite pixel belongs to a non-heterogeneous region. When the number of valid UAV pixels within a satellite pixel is greater than or equal to a preset threshold, it is determined that the heterogeneous raster corresponding to the satellite pixel has a valid value, indicating that the satellite pixel belongs to a heterogeneous region.

[0032] The preset threshold is determined based on the spatial resolution ratio of satellite pixels and UAV pixels and the preset minimum spatial coverage, or based on the minimum statistical sample size required for the degree of dispersion.

[0033] In step S4, based on the validity of the heterogeneous raster, the current sub-region is divided into heterogeneous and non-heterogeneous regions, such as... Figure 2 As shown, where: (1) Heterogeneous region: The multidimensional environmental features extracted from satellite imagery and the pixel values ​​of sub-pixel heterogeneous raster are used as inputs and the label values ​​are used as outputs to construct and train a heterogeneous region regression model, making full use of the high-precision local information of UAVs to improve prediction accuracy. (2) Non-heterogeneous region: Using multi-dimensional environmental features extracted from satellite imagery as input and label values ​​as output, a regression model for non-heterogeneous region is constructed and trained to ensure full-domain deployability.

[0034] The partitioning process is fully automated: when the number of valid UAV pixels within a satellite pixel is below a preset threshold, reliable statistics cannot be generated, and the pixel is automatically assigned to a non-heterogeneous region without the need for manual specification of partition boundaries. This automatic rollback mechanism ensures the robustness of the method in regions with missing heterogeneous data.

[0035] Construction and training of regression models 1) Perform equalization processing on the label values ​​at the satellite pixel scale of the current sub-region, specifically including: The number of satellite pixel-scale label values ​​in each sub-region is counted. Sub-regions with a sample size below the median are upsampled with replacement to make the surface soil moisture content of each sub-region more balanced, thus avoiding the sub-regions with high surface soil moisture content from dominating model training.

[0036] 2) Using the random forest with default parameters as the baseline model for regression, the permutation importance of each feature is calculated based on leave-one-out cross-validation. A subset of key features is retained according to the cumulative contribution rate threshold (default 0.70), and redundant features with low contribution to the baseline model are removed.

[0037] 3) Construct three types of candidate regression models, including Random Forest Regression Model (RF), ExtraTrees Regression Model, and Histogram Gradient Boosting Regression Model (HistGBR).

[0038] Two cross-validation strategies, leave-one-out sample cross-validation and spatial block cross-validation, were used to evaluate the generalization performance of various regression models: (1) Leave-one-out cross-validation (LOSO): Using each sub-region as a group unit, one sub-region is used as the test set and the remaining sub-regions are used as the training set to obtain the leave-one-out R² of each candidate regression model and evaluate the predictive ability of each candidate regression model in the test area. (2) Spatial Block Cross-Validation (BlockCV): The surface soil moisture content of the current sub-region is divided into several blocks according to spatial proximity, and the spatial block R² of each candidate regression model is obtained; the surface soil moisture content of the current sub-region is divided into several blocks according to spatial proximity, and spatial block cross-validation is used to obtain the spatial block R² of each candidate regression model, and the spatial generalization ability of each candidate regression model is evaluated.

[0039] Finally, the comprehensive score of each candidate regression model is obtained based on the median R² and the stability penalty; based on the leave-one-out R², spatial block R² and comprehensive score of each candidate regression model, the optimal regression model is determined.

[0040] This embodiment employs a comprehensive scoring method based on a weighted average of leave-one-out R² (weight 0.55), sample median R² (weight 0.20), spatial partitioning R² (weight 0.20), out-of-bag score (weight 0.03), normalized RMSE (weight -0.01), and model stability penalty (weight -0.01) to determine the optimal regression model. The model stability penalty is determined by calculating the catastrophic failure rate (the proportion of R² < -0.5, weight 0.35) of the leave-one-out R² for each sub-region in the LOSO, and is used to penalize models that exhibit unstable performance across sub-regions.

[0041] In step S5, based on the trained regression model, the prediction results of heterogeneous and non-heterogeneous regions of each sub-region are obtained respectively. The prediction results are then cropped to be within a reasonable quantile range of surface soil moisture content in the training samples to prevent unreasonable extreme values ​​from being generated by model extrapolation.

[0042] Finally, by stitching together the prediction results of heterogeneous and non-heterogeneous areas in each sub-region using a mask, surface water content data covering the entire target area is generated. This zoning strategy avoids the prediction distortion problem caused by forcibly extrapolating local UAV heterogeneity variables to the entire region.

[0043] The method proposed in this invention further includes step S6: integrating model uncertainty, training domain distance and regional support to construct a reliability evaluation system for the prediction results.

[0044] Model uncertainty: For random forest models, the standard deviation between trees is calculated by predicting each tree and is used as the model uncertainty; for non-random forest models, a residual surrogate model (ExtraTrees model with absolute residuals as the training target) is constructed to predict the model uncertainty of each drone pixel.

[0045] Training domain distance: Based on the K-nearest neighbor algorithm, the average distance between each UAV pixel and the surface soil moisture content of the UAV image of the sub-region is calculated in the standardized feature space. The larger the distance, the more the UAV pixel deviates from the training domain and the lower the reliability.

[0046] Regional support: Based on the coverage and spatial proximity of surface soil moisture content in UAV imagery of sub-regions, a regional support proxy index is constructed using weighted K-nearest neighbor interpolation. The weights decay with distance; the farther away from the surface soil moisture content in the UAV imagery of the sub-region, the lower the support.

[0047] Total uncertainty calculation: After normalizing the above three indicators respectively, the total uncertainty is calculated by weighting them according to their respective weights. Total uncertainty = α ×Model uncertainty (normalization)+ β × Training domain distance (normalized) + γ ×Regional support risk (normalized), where regional support risk = 1 - regional support; α, β, γ For the preset weights, satisfy α + β+γ =1.

[0048] Reliability grading: Reliability index = 1 - Total uncertainty. Based on the reliability index threshold (threshold range is 0.40-0.67), the prediction results are divided into three levels: reliability index ≥ 0.67 is high reliability (outputs official result graph), 0.40 ≤ reliability index < 0.67 is medium reliability (outputs cautious result graph), and reliability index < 0.40 is low reliability (outputs high risk graph).

[0049] Based on the reliability assessment results, output the formal outcome diagram, the cautious outcome diagram, and the high-risk diagram. Figure 3 The classification results provide users with a reliable and quantifiable dataset of surface water content.

[0050] It should be noted that this invention replaces the traditional quantile mapping method with a partitioned modeling strategy to address the problem of cross-scale feature distribution differences. The partitioned modeling strategy divides the target area into heterogeneous effective areas and non-heterogeneous areas, using models with different feature sets for prediction in each. This ensures global deployability while fully utilizing the high-precision local information from UAVs. Simultaneously, by fusing multi-dimensional environmental features, digital elevation models, slope, topographic moisture index, and soil texture index are introduced, compensating for the shortcomings of relying solely on NDVI and LST, which are insufficient to fully characterize the influence of topographic and hydrological factors on the spatial distribution of soil moisture. The model and parameters have regional adaptability, and their applicability can be further improved by increasing the number of training plots, expanding the feature set, and using temporal grouping.

[0051] This invention also proposes a system for constructing a dataset of surface soil moisture content, comprising: The data acquisition module is used to simultaneously acquire ground-measured soil moisture content data, UAV imagery, and satellite imagery of the target area, and to divide the target area into multiple sub-areas; The training sample generation module is used to acquire the surface soil moisture content of UAV images in each sub-region and aggregate it to the satellite pixel scale as a label value, while extracting the normalized vegetation index of the UAV images. Based on the normalized vegetation index of the UAV images, the dispersion of the normalized vegetation index of the effective UAV pixels contained within the spatial coverage of each satellite pixel is determined. The dispersion is used as the pixel value of the corresponding satellite pixel, and a raster layer is generated according to the spatial reference and resolution of the satellite images to obtain a sub-pixel heterogeneous raster. The regional modeling module is used to determine that sub-pixel heterogeneous gratings have valid values ​​when the number of valid UAV pixels in each satellite pixel is greater than or equal to a preset threshold. It uses multi-dimensional environmental features extracted from satellite imagery and pixel values ​​as inputs and label values ​​as outputs to construct and train a heterogeneous region regression model. When the number of valid UAV pixels is less than the preset threshold, it determines that sub-pixel heterogeneous gratings have no valid values. It uses multi-dimensional environmental features as inputs and label values ​​as outputs to construct and train a non-heterogeneous region regression model. The dataset construction module is used to obtain the prediction results of heterogeneous and non-heterogeneous areas of each sub-region based on the trained regression model, and then splice and merge them to construct a surface soil moisture content dataset covering the entire target area.

[0052] The method proposed in this invention uses ground point moisture content to calibrate the wet / dry index / temperature-vegetation spatial distribution within the UAV domain, generating continuous surface soil moisture content within the UAV sample plot. Through the surface soil moisture content distribution map, the true continuous surface soil moisture content within the UAV sample plot (UAV-derived Soil Surface) can be obtained. MoistureContent (UAV-SSMC) raster is used to expand sparse points into pixel-level high-density training samples. Then, multi-dimensional environmental features such as normalized vegetation index, surface temperature, digital elevation model, slope, topographic humidity index, and soil texture index from satellite imagery are used as inputs, and UAV-SSMC is used as the output label. The regression model is trained by selecting the best model through multi-model cross-validation. Based on the heterogeneity features of UAV sub-pixel, regional modeling is carried out. In the heterogeneous effective area, local high-precision information is fully utilized to improve the prediction accuracy, while in the non-heterogeneous area, it is backed up to the basic model to ensure global deployability. Finally, a reliability evaluation system is constructed by combining model uncertainty, training domain distance, and regional support, and a reliability classification result map is output to realize the global extrapolation of the satellite domain and produce a regional scale SSMC dataset as a surface soil moisture content dataset covering the entire target area.

[0053] Example Simulation like Figure 3 The figure shows the spatial distribution of land surface temperature (LST) and normalized difference vegetation index (NDVI) provided in this embodiment of the invention, illustrating the spatial distribution of NDVI–LST characteristics and the fitting results of wet and dry edges for different plots. Overall, all plots exhibit a typical triangular distribution structure, indicating a significant negative correlation between land surface temperature and vegetation index. In areas with low NDVI, the land surface temperature is higher, reflecting characteristics of bare land or arid regions; as NDVI increases, the land surface temperature gradually decreases, demonstrating the cooling effect of vegetation transpiration.

[0054] The dry edge and wet edge represent the temperature boundaries under extremely dry and humid conditions, respectively. The dry edge has a high coefficient of determination (R²), indicating that the dry state has stable thermal response characteristics. The wet edge has a relatively low fitting accuracy, reflecting that the thermal environment under humid conditions is affected by multiple factors.

[0055] This feature space provides a reliable basis for calculating the Temperature Vegetation Drought Index (TVDI) and supports subsequent soil moisture content inversion based on UAV data.

[0056] like Figure 4 As shown, the linear regression results between TVDI and measured soil surface water content (SSMC) are presented. Overall, all plots exhibit a significant negative correlation, meaning that soil water content gradually decreases as TVDI increases, consistent with the physical meaning of the drought index. The R² value for most plots is greater than 0.7, indicating... TVDI The model demonstrates strong explanatory power for soil moisture content. However, the relationship varies under different vegetation cover conditions. Sample points in high TVDI areas are more concentrated, and the regression relationship is more stable, while low NDVI areas are significantly influenced by soil background and surface conditions, exhibiting relatively stronger dispersion. Furthermore, the model confidence interval increases significantly in high TVDI areas, indicating higher uncertainty in soil moisture estimation under drought conditions. These results validate the feasibility of the TVDI-based soil moisture retrieval method and provide a reliable foundation for subsequent UAV-scale continuous soil moisture content estimation.

[0057] like Figure 5The image shows the surface soil moisture content distribution within the UAV plot area. This example uses the upper reaches of the xx watershed in the Qinghai-Tibet Plateau as the study area. Based on the method proposed in this invention, a continuous surface soil moisture content distribution map of this area was generated through ground point calibration. Overall, each plot exhibits significant spatial heterogeneity, with soil moisture showing a distinct patchy and gradient distribution pattern. High-value areas (blue-green) are mainly concentrated in rivers, lakes, and their adjacent areas, indicating that water recharge has a significant effect on increasing soil moisture content; low-value areas (brown) are widely distributed in areas far from water sources, reflecting the rapid decline of soil moisture under drought conditions. Comparison of different plots shows that the SSMC values ​​vary considerably, with the highest value reaching approximately 80%–90%, while the lowest value is close to 0%, indicating significant differences in moisture conditions within the study area. In some sample plots, continuous high-value zones were observed along rivers or water systems, indicating clear water transport and replenishment pathways. However, in areas with relatively flat terrain or far from water sources, SSMC distribution was more scattered, exhibiting spatial distribution characteristics influenced by soil texture, land cover, and micro-topography. Furthermore, significant differences existed in the SSMC variation gradient within different sample plots. In areas with well-developed water systems, SSMC exhibited a zonal distribution structure that gradually decreased from the river channel towards both sides. In arid or semi-arid regions, the spatial distribution of SSMC was more fragmented, with localized high-value areas possibly related to micro-topographic depressions or localized waterlogging conditions. Overall, the UAV-scale inversion results effectively characterized the fine spatial distribution features of soil moisture, reflecting both the high-humidity environment of water bodies and their adjacent areas, and revealing the spatial decay patterns of soil moisture under arid conditions. This provides reliable high-resolution foundational data for subsequent transfer of UAV-scale soil moisture information to satellite scales.

[0058] like Figure 6 As shown in the diagram, the multidimensional environmental feature extraction provided in this embodiment of the invention includes Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), Digital Elevation Model (DEM), Slope, Topographic Moisture Index (TWI), and Soil Texture Index (STI). Each feature exhibits significant differences in spatial distribution, comprehensively characterizing surface environmental features from multiple dimensions such as vegetation, surface heat, topographic structure, and hydrological conditions. Specifically, NDVI and LST primarily reflect surface energy and vegetation status, DEM and Slope describe topographic structure features, TWI characterizes regional water accumulation capacity, and STI provides soil background information. These features exhibit good spatial complementarity. Together, these multidimensional environmental features construct a multidimensional feature space describing the surface environment, providing an important foundation of input variables for subsequent soil moisture content retrieval models.

[0059] like Figure 7The figure shows the training and fitting results of different models provided in the embodiments of the present invention in heterogeneous partitioning. Overall, all three models (ExtraTrees, HistGBR, and RF) showed good fitting ability, with a significant positive correlation between predicted and observed values, and the sample points were mainly distributed near the 1:1 line. In terms of evaluation metrics, the coefficients of determination (R²) of each model were between 0.68 and 0.71, indicating that the models had a strong explanatory power for soil moisture content; meanwhile, the RMSE was approximately 6.0–6.3%, and the MAE was approximately 4.6–4.9%, indicating that the overall error level was relatively stable. Among them, the Random Forest (RF) model performed best, with the highest R² (approximately 0.712) and a relatively low error index, indicating that it had certain advantages in terms of fitting accuracy and stability. After introducing heterogeneity features, the overall performance of the model changed little, but the distribution of fitting results became more concentrated, indicating that heterogeneity features optimized the sample representation to a certain extent, helping to improve the model's ability to characterize complex spatial structures. Overall, all models were able to fit the relationship of soil moisture content changes well during the training phase, with the RF model showing the best overall performance, providing a reliable foundation for subsequent regional predictions.

[0060] like Figure 8The figure shows a comparison of the multi-model cross-validation scores provided in this embodiment of the invention, illustrating the performance comparison results of the base model and the model after introducing heterogeneity features under different evaluation metrics. Overall, there are significant differences in generalization ability and stability among the different models, with the Random Forest (RF) model performing best in most evaluation metrics. Among the base models, the leave-one-out cross-validation (LOSO R²) scores for RF, ExtraTrees, and HistGBR are approximately 0.408, 0.376, and 0.404, respectively, indicating that all three models have a certain degree of cross-site generalization ability, with RF and HistGBR performing relatively better. In spatial block cross-validation (Block CV R²), the three models perform relatively similarly (approximately 0.46–0.47), indicating that the models have little difference in spatial extrapolation ability. However, from the perspective of the median site R², all models are close to 0 or even negative, indicating that the model's fitting ability is limited at a single site scale, and there is a certain degree of instability. The overall scoring results show that the RF model performed best (approximately 0.289), followed by HistGBR and ExtraTrees. Introducing heterogeneous features improved the overall model performance. LOSO R² generally improved, with RF increasing to approximately 0.441, ExtraTrees to approximately 0.411, while HistGBR slightly decreased (approximately 0.387). In terms of Block CV R², ExtraTrees achieved the highest value (approximately 0.509), demonstrating stronger spatial generalization ability, while RF and HistGBR were approximately 0.500 and 0.486, respectively. Nevertheless, Median Site R² remained negative overall, indicating that the model still suffers from stability issues at local scales. The overall scoring results show that the RF model still maintains the best performance (approximately 0.327), showing an improvement over the base model, followed by ExtraTrees, while HistGBR performed the weakest. Overall, introducing sub-pixel heterogeneity features can improve the model's cross-plot and spatial generalization ability to some extent, especially in terms of LOSO and Block CV indices. However, the fitting ability at the local scale (site scale) still needs further optimization. Based on a comprehensive evaluation of multiple indices, the random forest model achieves a good balance between stability and generalization ability, making it the best-performing model type in this study.

[0061] like Figure 9The diagram illustrates the zoning modeling strategy (heterogeneous effective zone and non-heterogeneous zone) provided in this embodiment of the invention. Based on the effectiveness of sub-pixel heterogeneous features, the study area is divided into heterogeneous effective zones and non-heterogeneous zones. The heterogeneous effective zones are discretely distributed, often corresponding to areas with significant variations in surface conditions; the non-heterogeneous zones are dominant, with relatively homogeneous internal structures. Based on this zoning, the invention employs different modeling strategies for the heterogeneous effective and non-heterogeneous zones: in the heterogeneous effective zones, multidimensional environmental features and variation features are introduced as input to construct a heterogeneous model, enhancing its ability to express complex spatial structures; in the non-heterogeneous zones, a model based solely on multidimensional environmental features is used for prediction to ensure model stability and computational efficiency. Finally, through zoning prediction and result stitching, a high-precision estimate of the surface soil moisture content of the entire region is achieved. This zoning modeling strategy effectively avoids the overfitting or underfitting problems caused by uniform modeling across the entire region, achieving a balance between model complexity and generalization ability, and is one of the important technical innovations of this invention.

[0062] like Figure 10 The figure shows the spatial distribution of total uncertainty provided by this embodiment of the invention. Overall, the uncertainty exhibits a clear spatial heterogeneity. High-value areas (purple) are mainly concentrated in regions with complex terrain or drastically changing environmental conditions, while low-value areas (green) are mostly distributed in regions with relatively flat terrain and stable characteristics. Spatially, uncertainty clusters along terrain transition zones and areas with local elevation differences, indicating that the model's predictive stability is relatively low under complex surface conditions. Conversely, in regions with more uniform environmental conditions, uncertainty is lower, and the prediction results are more reliable. In summary, this result effectively reflects the distribution characteristics of uncertainty in the model's predictions, providing a basis for subsequent reliability evaluation.

[0063] like Figure 11 As shown, this is a reliability grading diagram (formal / cautious / high-risk) provided by an embodiment of the present invention. The study area is divided into a formal area (reliability ≥ 0.67), a cautious area (0.40 ≤ reliability < 0.67), and a high-risk area (reliability < 0.40). Overall, the cautious area accounts for the majority and is widely distributed across most of the study area; the formal area is distributed in a banded or concentrated manner, mostly corresponding to areas with high model stability; the high-risk area is distributed in patches and mainly appears in areas with complex environmental conditions or unstable features. This grading result can intuitively reflect the reliability of the model prediction results and provide a basis for the application of results in different areas. The formal area is suitable for direct application, the cautious area needs to be used in conjunction with auxiliary information, and the high-risk area requires careful interpretation or further verification.

[0064] like Figure 12As shown, this paper illustrates the spatial distribution characteristics of soil surface moisture content (SSMC) in the formal results area provided by this invention embodiment. Overall, SSMC exhibits a clear spatial differentiation pattern. High-value areas (blue-green) are mainly distributed along rivers and their adjacent areas, forming a continuous banded structure, reflecting the significant impact of water recharge on soil moisture. Low-value areas (brown) are widely distributed in areas far from water sources, exhibiting low moisture characteristics under arid conditions. Spatially, SSMC shows a gradient change from high to low along the river system, reflecting a clear pattern of moisture diffusion and attenuation. Simultaneously, some patchy distributions exist in local areas, reflecting the regulatory effect of topographic conditions and surface cover differences on soil moisture. In summary, the SSMC distribution in the formal results area has good spatial continuity and physical rationality, effectively characterizing the spatial pattern of soil moisture in the study area, and can serve as a reliable data foundation for subsequent analysis and application.

[0065] like Figure 13 As shown, the spatial distribution characteristics of the optimal soil surface water content prediction results provided by this embodiment of the invention are as follows: Overall, the soil water content exhibits a clear spatial differentiation pattern. High-value areas (blue-green) are mainly distributed along rivers and their surrounding areas, forming a continuous zonal structure, reflecting the significant impact of water bodies and their recharge on soil moisture. Low-value areas (brown) are widely distributed in areas far from water sources and with large topographic relief, exhibiting low moisture characteristics under drought conditions. Spatially, the SSMC shows a gradient change from high to low along the river system, while also showing a certain patchy distribution in some local areas, reflecting the comprehensive regulatory effect of topographic conditions and surface cover differences on soil moisture. In general, the results maintain spatial continuity while taking into account the prediction reliability of different regions, and can relatively realistically reflect the spatial pattern of soil moisture in the study area.

[0066] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0067] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for constructing a dataset of surface soil moisture content, characterized in that, Includes the following steps: Simultaneously acquire ground-measured soil moisture content data, UAV imagery, and satellite imagery of the target area, and divide the target area into multiple sub-regions; In each sub-region, the surface soil moisture content of the UAV imagery is acquired and aggregated to the satellite pixel scale as a label value, while the normalized vegetation index of the UAV imagery is extracted. Based on the normalized vegetation index of UAV imagery, the dispersion of the normalized vegetation index of effective UAV pixels contained within the spatial coverage area of ​​each satellite pixel is determined. Using the degree of dispersion as the pixel value of the corresponding satellite pixel, a raster layer is generated according to the spatial reference and resolution of the satellite image to obtain a sub-pixel heterogeneous raster. When the number of valid UAV pixels in each satellite pixel is greater than or equal to a preset threshold, the sub-pixel heterogeneous raster is determined to have a valid value. The multidimensional environmental features extracted from the satellite image and the pixel value are used as inputs, and the label value is used as output to construct and train a heterogeneous region regression model. When the number of valid UAV pixels is less than the preset threshold, the sub-pixel heterogeneous raster is determined to have no valid value. The multidimensional environmental features are used as inputs, and the label value is used as output to construct and train a non-heterogeneous region regression model. Based on the trained regression model, the prediction results of heterogeneous and non-heterogeneous areas in each sub-region are obtained and spliced ​​together to construct a surface soil moisture content dataset covering the entire target area.

2. The method for constructing a surface soil moisture content dataset according to claim 1, characterized in that, In each sub-region, the surface soil moisture content of the UAV imagery is acquired and aggregated to the satellite pixel scale as a label value. Simultaneously, the normalized vegetation index (NDI) of the UAV imagery is extracted, specifically including: Extract the normalized vegetation index and surface temperature from the UAV imagery of the current sub-region; The temperature-vegetation drought index is determined by constructing a characteristic triangular space between the normalized vegetation index and the land surface temperature. Using the measured soil moisture content data of the current sub-region, the temperature-vegetation-drought index is regressed and calibrated to generate the surface soil moisture content of the current sub-region. The surface soil moisture content at the UAV pixel scale is aggregated to the satellite pixel scale and used as the label value at the satellite pixel scale.

3. The method for constructing a surface soil moisture content dataset according to claim 1, characterized in that, The normalized vegetation index (NVI) based on UAV imagery determines the dispersion of the NVI of effective UAV pixels within the spatial coverage area of ​​each satellite pixel, specifically including: The effective drone pixels are drone pixels whose normalized vegetation index (NVI) values ​​are not empty and whose NVI values ​​are within a preset range. Extract the normalized vegetation index values ​​of all valid UAV pixels within the spatial coverage area of ​​the satellite pixel to form the index set of the satellite pixel; Obtain the standard deviation of the index set as the degree of dispersion; For satellite pixels that do not contain valid drone pixels, assign an invalid identifier in the sub-pixel heterogeneity raster.

4. The method for constructing a surface soil moisture content dataset according to claim 1, characterized in that, When the number of valid UAV pixels within each satellite pixel is greater than or equal to a preset threshold, it is determined that the sub-pixel heterogeneous raster has a valid value. Using multidimensional environmental features and pixel values ​​extracted from satellite imagery as input, and the label value as output, a heterogeneous region regression model is constructed and trained, specifically including: When the number of valid UAV pixels within a satellite pixel is greater than or equal to a preset threshold, it is determined that the sub-pixel heterogeneous raster corresponding to the satellite pixel has a valid value, indicating that the satellite pixel belongs to a heterogeneous region. The preset threshold is determined based on the spatial resolution ratio of satellite pixels and UAV pixels and the preset minimum spatial coverage, or based on the minimum statistical sample size required for the degree of dispersion. In heterogeneous regions, multiple candidate regression models for heterogeneous regions are constructed by using multidimensional environmental features extracted from satellite imagery and pixel values ​​of sub-pixel heterogeneous rasters as input features and the label values ​​as output. The multidimensional environmental features include normalized vegetation index, surface temperature, digital elevation model, slope, topographic humidity index and soil texture index, and the candidate regression models include random forest regression model, extreme random tree regression model and histogram gradient boosting regression model. Obtain the comprehensive score for each candidate regression model based on the median R² and stability penalty; Using sub-regions as the grouping unit, one sub-region is used as the test set and the remaining sub-regions are used as the training set. Leave-one-out sample cross-validation is used to obtain the leave-one-out R² of each candidate regression model. The surface soil moisture content of the current sub-region is divided into several blocks according to spatial proximity. Spatial block cross-validation is used to obtain the spatial block R² of each candidate regression model. Based on the leave-one-out R², spatial block R², and comprehensive score of each candidate regression model, the optimal regression model for the heterogeneous region is determined. The optimal regression model for heterogeneous regions is used as the regression model to be constructed and trained.

5. The method for constructing a surface soil moisture content dataset according to claim 1, characterized in that, When the number of effective drone pixels is less than a preset threshold, it is determined that the sub-pixel heterogeneous raster has no valid value. Using multidimensional environmental features as input and the label value as output, a non-heterogeneous region regression model is constructed and trained, specifically including: When the number of valid UAV pixels in a satellite pixel is less than a preset threshold, it is determined that the sub-pixel heterogeneous grid corresponding to the satellite pixel has no valid value, indicating that the satellite pixel belongs to a non-heterogeneous area. In non-heterogeneous regions, multiple candidate regression models for non-heterogeneous regions are constructed using multidimensional environmental features as input features and the label values ​​as output. k-fold cross-validation or leave-one-out cross-validation was used to evaluate the predictive performance of candidate regression models in non-heterogeneous regions and obtain the corresponding coefficient of determination R² and root mean square error RMSE. The optimal regression model for the non-heterogeneous region is determined based on the coefficient of determination R² and the root mean square error RMSE. The optimal regression model in the non-heterogeneous region is used as the regression model to be constructed and trained.

6. The method for constructing a surface soil moisture content dataset according to claim 1, characterized in that, The process of constructing and training the regression model also includes equalization processing of the label value samples at the satellite pixel scale of the current sub-region, specifically including: The sample size of the label values ​​at the satellite pixel scale for each sub-region is counted, and sub-regions with sample sizes below the median are upsampled with replacement.

7. The method for constructing a surface soil moisture content dataset according to claim 1, characterized in that, The process of constructing and training the regression model also includes feature selection of multidimensional environmental features based on leave-one-out cross-validation, specifically including: The contribution of each environmental feature in the multidimensional environmental features is assessed by permutation importance evaluation. A subset of key environmental features is retained according to the cumulative contribution rate threshold, while redundant environmental features are eliminated.

8. The method for constructing a surface soil moisture content dataset according to claim 1, characterized in that, The stitching and fusion process involves stitching together the prediction results of regression models corresponding to heterogeneous and non-heterogeneous regions of each sub-region using a mask, thereby generating a surface water content dataset covering the entire target region.

9. The method for constructing a surface soil moisture content dataset according to claim 1, characterized in that, The simultaneous acquisition of ground-measured soil moisture content data, UAV imagery, and satellite imagery of the target area also includes atmospheric correction and cloud removal processing of the acquired data.

10. A system for constructing a dataset of surface soil moisture content, characterized in that, include: The data acquisition module is used to simultaneously acquire ground-measured soil moisture content data, UAV imagery, and satellite imagery of the target area, and to divide the target area into multiple sub-areas; The training sample generation module is used to acquire the surface soil moisture content of UAV images in each sub-region and aggregate it to the satellite pixel scale as a label value. At the same time, it extracts the normalized vegetation index of UAV images. Based on the normalized vegetation index of UAV images, it determines the dispersion of the normalized vegetation index of effective UAV pixels contained within the spatial coverage of each satellite pixel. Using the degree of dispersion as the pixel value of the corresponding satellite pixel, a raster layer is generated according to the spatial reference and resolution of the satellite image to obtain a sub-pixel heterogeneous raster. The regional modeling module is used to determine that sub-pixel heterogeneous gratings have valid values ​​when the number of valid UAV pixels in each satellite pixel is greater than or equal to a preset threshold. It uses multi-dimensional environmental features extracted from satellite images and the pixel values ​​as inputs, and the label values ​​as outputs to construct and train a heterogeneous region regression model. When the number of valid UAV pixels is less than the preset threshold, it determines that sub-pixel heterogeneous gratings have no valid values. It uses multi-dimensional environmental features as inputs and the label values ​​as outputs to construct and train a non-heterogeneous region regression model. The dataset construction module is used to obtain the prediction results of heterogeneous and non-heterogeneous areas of each sub-region based on the trained regression model, and then splice and merge them to construct a surface soil moisture content dataset covering the entire target area.

Images