A soil moisture downscaling method based on dynamic and static factors

Abstract

The application discloses a soil humidity downscaling method based on dynamic and static factors, which comprises the following steps: DEM data, GPM satellite precipitation product, and original satellite soil humidity data preprocessing; soil humidity multi-branch feature fusion based on dynamic and static factors; Swin Transformer feature reconstruction and boundary constraint; through double-scale DEM feature extraction, precipitation dynamic feature construction, and multi-branch Swin Transformer architecture, the physical rationality of the downscaling result is ensured, and the high-precision reconstruction problem of soil humidity spatial heterogeneity in complex terrain areas is solved; combined with dynamic and static factors, automatic high-resolution soil humidity generation is realized, which is beneficial to the direct calling of the soil humidity downscaling method based on deep learning, fast obtaining of fine and personalized soil moisture data, serving as the input of hydrological models and drought monitoring, and further promoting the in-depth development of precision agriculture and water resource management.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing and artificial intelligence, specifically to a method for downscaling soil moisture based on dynamic and static factors. Background Technology

[0002] Topographic features not only influence the form and distribution of precipitation but also have a profound impact on the spatial pattern of soil moisture. As a key variable in the terrestrial water cycle, accurate measurement of soil moisture is of great significance for agricultural irrigation, drought monitoring, and water resource management. However, existing satellite remote sensing soil moisture products (such as SMAP) are limited by sensor resolution (36km), making it difficult to meet the needs of refined applications at the regional scale. This contradiction is particularly prominent in areas with complex terrain.

[0003] Downscaling techniques serve as a crucial bridge connecting large-scale remote sensing observations with regional hydrological applications, aiming to improve the spatial resolution and application accuracy of soil moisture data at local scales. With the development of precision agriculture and eco-hydrological research, the demand for high-resolution soil moisture data is increasingly urgent, making soil moisture downscaling a research hotspot in the field of remote sensing hydrology. Existing soil moisture downscaling methods mainly include physical modeling methods and statistical learning methods. Physical modeling methods establish soil-vegetation-atmosphere transport models (SVAT) and combine them with surface energy balance equations for downscaling calculations. While these methods have clear physical mechanisms, they require a large number of surface parameter inputs and exhibit significant uncertainties in simulating turbulent exchange processes under complex terrain conditions. Another approach is statistical learning methods, which utilize machine learning algorithms (such as random forests and neural networks) to establish statistical relationships between low-resolution soil moisture and auxiliary variables (such as NDVI and surface temperature). Although these methods are computationally efficient, they have two inherent drawbacks: first, they rely on the quality and representativeness of the training data, and their performance drops significantly in remote areas with few observations; second, they are difficult to effectively integrate multi-source data with different spatiotemporal characteristics, especially in terms of insufficient consideration of the synergistic effect between precipitation dynamics and static topographic features.

[0004] To address the aforementioned issues, this paper describes the dynamic changes in soil moisture by quantifying the spatiotemporal evolution characteristics of precipitation fields. Simultaneously, it combines a digital elevation model (DEM) to extract topographic parameters, thereby constructing a SwinTransformer downscaling method for soil moisture based on dynamic and static factors to achieve high-resolution soil moisture reconstruction. Summary of the Invention

[0005] The purpose of this invention is to provide a soil moisture downscaling method based on dynamic and static factors. This method considers the synergistic influence of static topographic features and dynamic precipitation processes on the spatial distribution of soil moisture. It utilizes high- and low-resolution DEM data, GPM precipitation time series, and original SMAP soil moisture data to achieve high-precision downscaling of soil moisture through a multi-branch Swing Transformer architecture. This method has advantages such as sufficient feature fusion and good boundary preservation, which is beneficial for the rapid generation and application of soil moisture products in complex terrain areas.

[0006] To achieve the above functions, this invention designs a soil moisture downscaling method based on dynamic and static factors. For the target area, the following steps S1-S4 are performed to generate high-resolution soil moisture for the target area:

[0007] Step S1: Collect DEM data, GPM satellite precipitation products, and raw satellite soil moisture data of the target area; extract the dual-scale static DEM topographic features of the target area; construct the dynamic precipitation GPM features of the target area; resample the raw satellite soil moisture data of the target area to obtain the original mask.

[0008] Step S2: Load the dual-scale static data DEM topographic features of the target area as static topographic features; use an encoder network to encode the dynamic precipitation GPM features of the target area and the raw satellite soil moisture data of the target area to obtain dynamic coded features and observation coded features respectively;

[0009] Step S3: Concatenate and recombine the static terrain features, dynamic coding features, and observation coding features to obtain a fused feature map, input it into the Swing Transformer block for processing, and output a soil moisture prediction map.

[0010] Step S4: Based on the original mask obtained in step S1 and the soil moisture prediction map obtained in step S3, numerical physical constraints are applied to obtain the final downscaling result and store it, thus completing the high-resolution soil moisture generation of the target area.

[0011] As a preferred embodiment of the present invention, the specific steps for extracting the dual-scale static data (DEM) terrain features of the target area in step S1 are as follows:

[0012] Step S1.1.1: Resample the DEM data of the target area to the target resolution and the original satellite soil moisture data resolution respectively to obtain the target high-resolution DEM data and the original low-resolution DEM data respectively. Convert the target high-resolution DEM data and the original low-resolution DEM data into PyTorch tensor format and add preset batch and channel dimensions to generate input tensors with shapes of [batch size, number of channels, high-resolution height, high-resolution width] and [batch size, number of channels, low-resolution height, low-resolution width] respectively, where the batch size and number of channels are 1.

[0013] Step S1.1.2: Input the target high-resolution DEM data into the high-resolution branch network. The high-resolution branch network includes two 3×3 convolutional layers and outputs a 32-channel local terrain detail feature map. Upsample the original low-resolution DEM data to the high-resolution size through bilinear interpolation and input it into the low-resolution branch network. The low-resolution branch network includes two 7×7 convolutional layers and outputs a 32-channel macro-terrain trend feature map.

[0014] The 32-channel feature maps output from the two branch networks are concatenated along the channel dimension to form a 64-channel fused feature map; the fused feature map is then passed through a 1×1 convolutional layer to obtain a 64-channel dual-scale static data DEM terrain feature map with dimensions [1, 64, high-resolution height, high-resolution width].

[0015] Step S1.1.3: Save the dual-scale static data DEM terrain feature map as an HDF5 format file.

[0016] As a preferred embodiment of the present invention: the dynamic precipitation GPM characteristics of the target area in step S1 include current precipitation characteristics, optical flow amplitude characteristics, and precipitation change characteristics; the time window is set to 2 days, including the target day T and the day before the target day T-1; the specific steps for constructing the dynamic precipitation GPM characteristics of the target area are as follows:

[0017] Step S1.2.1: Generate current precipitation characteristics:

[0018] Load daily precipitation data from GPM satellites for the target area;

[0019] The bilinear interpolation method was used to resample the GPM satellite daily precipitation product data of the target area to the target resolution;

[0020] Fill missing data areas with zero values;

[0021] Step S1.2.2: Generate optical flow amplitude characteristics:

[0022] The precipitation fields on day T-1 and day T were normalized.

[0023] Calculation of precipitation motion field based on Farneback optical flow algorithm;

[0024] Calculate the optical flow amplitude characteristic matrix;

[0025] Step S1.2.3: Generate precipitation variation characteristics:

[0026] Calculate the pixel-by-pixel difference between the precipitation fields on day T and day T-1, and generate a difference matrix;

[0027] Fill in the missing data areas with zero values ​​to obtain the characteristics of precipitation variation.

[0028] As a preferred embodiment of the present invention, the method for obtaining the original mask in step S1 is as follows:

[0029] The original satellite soil moisture data of the target area is resampled to the target resolution. Linear interpolation is performed on the resampled data. The nearest neighbor interpolation is used to fill the edges of the target area. Non-NaN regions are extracted from the resampled data as the original mask.

[0030] As a preferred technical solution of the present invention, step S2 specifically involves the following steps:

[0031] Step S2.1: Load the HDF5 format dual-scale static data DEM terrain feature map obtained in step S1 as the static terrain feature;

[0032] Step S2.2: Input the three-channel target area dynamic precipitation GPM features into the encoder network. The first convolutional layer of the encoder network uses 32 3×3 convolutional kernels to extract spatial patterns and outputs a 32-channel feature map; the second convolutional layer uses 64 3×3 convolutional kernels to perform nonlinear transformation and high-order abstraction, and finally outputs 64-channel dynamic encoded features.

[0033] Step S2.3: Input the original satellite soil moisture data of the target area of ​​a single channel into the encoder network. The first layer of the encoder network uses 32 3×3 convolutional kernels to extract local features and outputs a 32-channel feature map. The second layer uses 64 3×3 convolutional kernels to expand the channel dimension and extract the spatial distribution pattern of a medium range, and finally outputs 64-channel observation coding features.

[0034] As a preferred technical solution of the present invention, step S3 is as follows:

[0035] Step S3.1: Concatenate the static terrain features, dynamic coding features, and observation coding features along the channel dimension to generate a 192-channel fused feature map. Then, convert the format from [B, C, H, W] to [B, H, W, C] through a dimension permutation operation, where B is the batch size, C is the total number of channels, and H and W are the target space height and width.

[0036] Step S3.2: Input the fused feature map into two consecutive Swing Transformer blocks, each block containing a window attention mechanism and a shift window mechanism;

[0037] Step S3.3: Replace the feature dimensions after processing by the Swin Transformer block back to the [B, C, H, W] format, and then perform dimensionality reduction through two 1×1 convolutional layers. The first layer compresses 192 channels to 64 channels and activates them with ReLU. The second layer maps the 64 channels to a single-channel soil moisture prediction map while maintaining the target resolution.

[0038] As a preferred technical solution of the present invention, step S4 specifically involves the following steps:

[0039] Step S4.1: Based on the original mask obtained in step S1, force the predicted values ​​of the area outside the mask in the soil moisture prediction map output by the Swin Transformer block in step S3 to be NaN;

[0040] Step S4.2: Extract the maximum and minimum resolution values ​​of the original satellite soil moisture data as the numerical range of numerical physics constraints. Use the truncation function to limit the resolution of the soil moisture prediction map output by the Swin Transformer block to this numerical range to obtain the final downscaling result.

[0041] Step S4.3: Store the final downscaling results in HDF5 format, create a dataset to save the final downscaling results, and record the data source, resolution, and numerical range information in the file attributes.

[0042] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0043] Compared with existing technologies, this invention proposes a soil moisture downscaling method based on dynamic and static factors. By fusing static topographic features, dynamic precipitation processes, and raw observation data, it quantifies the coupling relationship between elevation features, precipitation motion fields, and the spatial distribution of soil moisture. A multi-branch Swin Transformer architecture is employed for collaborative modeling, leading to a Swin Transformer downscaling method for soil moisture based on dynamic and static factors. This approach ensures both the physical rationality and spatial accuracy of the downscaling results while addressing the challenge of downscaling the strong spatial heterogeneity of soil moisture in complex topographic regions. This method primarily relies on remote sensing and topographic data, ensuring stable and reliable data sources. The feature extraction and fusion mechanisms in the model have clear physical meanings, facilitating the automatic learning of optimal feature representations and reconstruction strategies in different geographical regions. The end-to-end deep learning architecture simplifies the complex process of traditional downscaling while ensuring the objectivity and consistency of the results. This promotes the standardized application of soil moisture downscaling methods based on dynamic and static factors, rapidly generating high-resolution and physically consistent soil moisture products that can be used as input for land surface process models and hydrological models, further advancing the development of ecological environment monitoring and hydro-meteorological forecasting. Attached Figure Description

[0044] Figure 1 This is a flowchart of a soil moisture downscaling method based on dynamic and static factors according to an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the target area DEM data extracted according to an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the original satellite soil moisture data of the target area extracted according to an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of a downscaled soil moisture product provided according to an embodiment of the present invention. Detailed Implementation

[0048] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0049] This invention provides a soil moisture downscaling method based on dynamic and static factors, targeting a specific area, and referring to... Figure 1 Perform the following steps S1-S4 to generate high-resolution soil moisture for the target area:

[0050] Step S1: Collect DEM data, GPM satellite precipitation products, and raw satellite soil moisture data of the target area. Use the DEM data of the target area as a static factor to extract the dual-scale static DEM topographic features of the target area. Use the GPM satellite precipitation products of the target area as a dynamic factor to construct the dynamic precipitation GPM features of the target area. Resample the raw satellite soil moisture data of the target area to obtain the original mask.

[0051] In this embodiment, Anhui Province is taken as the target area. Located at the intersection of the eastern coastal economic belt and the Yangtze River economic belt, Anhui is situated in the transitional zone between the warm temperate and subtropical zones and has a typical monsoon climate. The province's terrain is high in the southwest and low in the northeast, with complex and diverse topography, encompassing plains, hills, mountains, and other landforms.

[0052] The raw DEM data for the target area was obtained from 1km resolution SRTM (ShuttleRadar Topography Mission) data provided by the Geospatial Data Cloud. Dynamic precipitation data was obtained from 0.1° resolution GPM (Global Precipitation Measurement). Raw satellite soil moisture data was obtained from 9km resolution SMAP L4 (Soil Moisture Active and Passive).

[0053] The specific steps for extracting the dual-scale static data DEM terrain features of the target area in step S1 are as follows:

[0054] Step S1.1.1: Resample the DEM data of the target area to the target resolution and the original satellite soil moisture data resolution, respectively, to obtain the target high-resolution DEM data and the original low-resolution DEM data. See the schematic diagram of the extracted target area DEM data for reference. Figure 2 (a) represents the target high-resolution DEM data, and (b) represents the original low-resolution DEM data. The target high-resolution DEM data and the original low-resolution DEM data are converted into PyTorch tensor format and preset batch and channel dimensions are added to generate input tensors with shapes of [batch size, number of channels, high-resolution height, high-resolution width] and [batch size, number of channels, low-resolution height, low-resolution width], respectively, where the batch size and number of channels are 1.

[0055] Step S1.1.2: Input the target high-resolution DEM data into the high-resolution branch network. The high-resolution branch network includes two 3×3 convolutional layers. The two 3×3 convolutional layers convert the 1-channel target high-resolution DEM data into 16 channels, and then into 32 channels, outputting a 32-channel local terrain detail feature map. After upsampling the original low-resolution DEM data to a high-resolution size through bilinear interpolation, input it into the low-resolution branch network. The low-resolution branch network includes two 7×7 convolutional layers. The two 7×7 convolutional layers convert the 1-channel original low-resolution DEM data into 16 channels, and then into 32 channels, outputting a 32-channel macro-terrain trend feature map.

[0056] The 32-channel feature maps output from the two branch networks are concatenated along the channel dimension to form a 64-channel fused feature map; the fused feature map is then passed through a 1×1 convolutional layer to obtain a 64-channel dual-scale static data DEM terrain feature map with dimensions [1, 64, high-resolution height, high-resolution width].

[0057] Step S1.1.3: Save the dual-scale static data DEM terrain feature map as an HDF5 format file, and create a dataset named "features" to store 64-channel feature maps for use in subsequent downscaling models.

[0058] In step S1, the dynamic precipitation GPM characteristics of the target area include current precipitation characteristics, optical flow amplitude characteristics, and precipitation change characteristics. Since soil moisture response to precipitation has a lag, the time window is set to 2 days, including the target day T and the day before the target day T-1. The specific steps for constructing the dynamic precipitation GPM characteristics of the target area are as follows:

[0059] Step S1.2.1: Generate current precipitation characteristics:

[0060] Load daily precipitation data from GPM satellites for the target area;

[0061] The bilinear interpolation method was used to resample the GPM satellite daily precipitation product data of the target area to the target resolution;

[0062] Fill missing data areas with zero values; this feature can provide the spatial distribution of precipitation intensity on the day, preserving the original spatial pattern of precipitation.

[0063] Step S1.2.2: Generate optical flow amplitude characteristics:

[0064] The precipitation fields on day T-1 and day T were normalized.

[0065] Calculation of precipitation motion field based on Farneback optical flow algorithm;

[0066] Calculate the optical flow amplitude feature matrix; this feature can quantify the moving speed and directional intensity of the precipitation system and capture the migration dynamics of the precipitation belt; this feature can capture the short-term precipitation increase and decrease trend and enhance the model's sensitivity to sudden heavy precipitation.

[0067] Step S1.2.3: Generate precipitation variation characteristics:

[0068] Calculate the pixel-by-pixel difference between the precipitation fields on day T and day T-1, and generate a difference matrix;

[0069] Fill in the missing data areas with zero values ​​to obtain the characteristics of precipitation variation;

[0070] The method for obtaining the original mask in step S1 is as follows:

[0071] The original satellite soil moisture data of the target area was resampled to the target resolution, with the original satellite soil moisture data of the target area referenced... Figure 3 Linear interpolation is performed on the resampled data to preserve its spatial continuity. Nearest-neighbor interpolation is used to fill the edges of the target region, effectively balancing smoothness and boundary integrity. Non-NaN regions are extracted from the resampled data as the original mask, where NaN regions represent undefined areas formed due to missing data. This step provides a reference for the target region boundaries and soil moisture range for subsequent model learning.

[0072] Step S2: Load the dual-scale static data DEM topographic features of the target area as static topographic features; use an encoder network to encode the dynamic precipitation GPM features of the target area and the raw satellite soil moisture data of the target area to obtain dynamic coded features and observation coded features respectively;

[0073] The specific steps of step S2 are as follows:

[0074] Step S2.1: Load the HDF5 format dual-scale static data DEM terrain feature map obtained in step S1. This feature map already contains the fusion information of high-resolution local terrain details and low-resolution macro terrain trends. It can be directly used as static terrain feature input to the subsequent multi-branch feature fusion stage without additional encoding processing.

[0075] Step S2.2: Input the three-channel target area dynamic precipitation GPM features into the encoder network. The first convolutional layer of the encoder network uses 32 3×3 convolutional kernels to extract spatial patterns and outputs a 32-channel feature map; the second convolutional layer uses 64 3×3 convolutional kernels to perform nonlinear transformation and high-order abstraction, and finally outputs 64-channel dynamic encoded features.

[0076] Step S2.3: Input the original satellite soil moisture data of the target area of ​​a single channel into the encoder network. The first layer of the encoder network uses 32 3×3 convolutional kernels to extract local features and outputs a 32-channel feature map. The second layer uses 64 3×3 convolutional kernels to expand the channel dimension and extract the spatial distribution pattern of a medium range, and finally outputs 64-channel observation coding features.

[0077] Step S3: Concatenate and recombine the static terrain features, dynamic coding features, and observation coding features to obtain a fused feature map, input it into the Swing Transformer block for processing, and output a soil moisture prediction map.

[0078] The specific steps of step S3 are as follows:

[0079] Step S3.1: Concatenate the static terrain features, dynamic coding features, and observation coding features along the channel dimension to generate a 192-channel fused feature map. Then, convert the format from [B, C, H, W] to [B, H, W, C] through a dimension permutation operation to adapt to the input requirements of the Transformer architecture, where B is the batch size, C is the total number of channels, and H and W are the target space height and width.

[0080] Step S3.2: Input the fused feature map into two consecutive Swing Transformer blocks, each block containing a window attention mechanism and a shift window mechanism; realize the calculation of self-attention within the local window and establish cross-window information interaction;

[0081] Step S3.3: Replace the feature dimensions after processing by the Swin Transformer block back to the [B, C, H, W] format, and then perform dimensionality reduction through two 1×1 convolutional layers. The first layer compresses 192 channels to 64 channels and activates them with ReLU. The second layer maps the 64 channels to a single-channel soil moisture prediction map while maintaining the target resolution.

[0082] Step S4: Based on the original mask obtained in step S1 and the soil moisture prediction map obtained in step S3, numerical physical constraints are applied to obtain the final downscaling result and store it, thus completing the high-resolution soil moisture generation of the target area.

[0083] The specific steps of step S4 are as follows:

[0084] Step S4.1: Based on the original mask (i.e. non-NaN region) obtained in step S1, the predicted values ​​of the regions outside the mask in the soil moisture prediction map output by the SwinTransformer block in step S3 are forcibly set to NaN to ensure that the predicted values ​​are retained only in the effective region.

[0085] Step S4.2: Extract the maximum and minimum resolution values ​​of the original satellite soil moisture data as the numerical range for numerical physics constraints. Use a truncation function to limit the resolution of the soil moisture prediction map output by the Swin Transformer block to this numerical range, ensuring that the output results conform to the physical rationality of soil moisture, and obtain the final downscaling results. Figure 4 ;

[0086] Step S4.3: Store the final downscaling results in HDF5 format, create a dataset to save the final downscaling results, create a dataset named "swc" to save the processed soil moisture data, and record the data source, resolution and numerical range information in the file attributes.

[0087] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for downscaling soil moisture based on dynamic and static factors, characterized in that, For the target area, perform the following steps S1-S4 to generate high-resolution soil moisture for the target area: Step S1: Collect DEM data, GPM satellite precipitation products, and raw satellite soil moisture data of the target area; extract the dual-scale static DEM topographic features of the target area; construct the dynamic precipitation GPM features of the target area; resample the raw satellite soil moisture data of the target area to obtain the original mask. The dynamic precipitation GPM characteristics of the target area in step S1 include current precipitation characteristics, optical flow amplitude characteristics, and precipitation change characteristics; The time window is set to 2 days, including the target day T and the day before the target day T-1; the specific steps for constructing the dynamic precipitation GPM characteristics of the target area are as follows: Step S1.2.1: Generate current precipitation characteristics: Load daily precipitation data from GPM satellites for the target area; The bilinear interpolation method was used to resample the GPM satellite daily precipitation product data of the target area to the target resolution; Fill missing data areas with zero values; Step S1.2.2: Generate optical flow amplitude characteristics: The precipitation fields on day T-1 and day T were normalized. Calculation of precipitation motion field based on Farneback optical flow algorithm; Calculate the optical flow amplitude characteristic matrix; Step S1.2.3: Generate precipitation variation characteristics: Calculate the pixel-by-pixel difference between the precipitation fields on day T and day T-1, and generate a difference matrix; Fill in the missing data areas with zero values ​​to obtain the characteristics of precipitation variation; Step S2: Load the dual-scale static data DEM topographic features of the target area as static topographic features; use an encoder network to encode the dynamic precipitation GPM features of the target area and the raw satellite soil moisture data of the target area to obtain dynamic coded features and observation coded features respectively; The specific steps of step S2 are as follows: Step S2.1: Load the HDF5 format dual-scale static data DEM terrain feature map obtained in step S1 as the static terrain feature; Step S2.2: Input the three-channel target area dynamic precipitation GPM features into the encoder network. The first convolutional layer of the encoder network uses 32 3×3 convolutional kernels to extract spatial patterns and outputs a 32-channel feature map; the second convolutional layer uses 64 3×3 convolutional kernels to perform nonlinear transformation and high-order abstraction, and finally outputs 64-channel dynamic encoded features. Step S2.3: Input the raw satellite soil moisture data of the target area of ​​a single channel into the encoder network. The first layer of the encoder network uses 32 3×3 convolutional kernels to extract local features and outputs a 32-channel feature map. The second layer uses 64 3×3 convolutional kernels to expand the channel dimension and extract the spatial distribution pattern of a medium range, and finally outputs 64-channel observation coding features. Step S3: Concatenate and recombine the static terrain features, dynamic coding features, and observation coding features to obtain a fused feature map, input it into the Swing Transformer block for processing, and output a soil moisture prediction map. Step S4: Based on the original mask obtained in step S1 and the soil moisture prediction map obtained in step S3, numerical physical constraints are applied to obtain the final downscaling result and store it, thus completing the high-resolution soil moisture generation of the target area.

2. The method for downscaling soil moisture based on dynamic and static factors according to claim 1, characterized in that, The specific steps for extracting the dual-scale static data DEM terrain features of the target area in step S1 are as follows: Step S1.1.1: Resample the DEM data of the target area to the target resolution and the original satellite soil moisture data resolution respectively to obtain the target high-resolution DEM data and the original low-resolution DEM data respectively. Convert the target high-resolution DEM data and the original low-resolution DEM data into PyTorch tensor format and add preset batch and channel dimensions to generate input tensors with shapes of [batch size, number of channels, high-resolution height, high-resolution width] and [batch size, number of channels, low-resolution height, low-resolution width] respectively, where the batch size and number of channels are 1. Step S1.1.2: Input the target high-resolution DEM data into the high-resolution branch network. The high-resolution branch network includes two 3×3 convolutional layers and outputs a 32-channel local terrain detail feature map. Upsample the original low-resolution DEM data to the high-resolution size through bilinear interpolation and input it into the low-resolution branch network. The low-resolution branch network includes two 7×7 convolutional layers and outputs a 32-channel macro-terrain trend feature map. The 32-channel feature maps output from the two branch networks are concatenated along the channel dimension to form a 64-channel fused feature map; the fused feature map is then passed through a 1×1 convolutional layer to obtain a 64-channel dual-scale static data DEM terrain feature map with dimensions [1, 64, high-resolution height, high-resolution width]. Step S1.1.3: Save the dual-scale static data DEM terrain feature map as an HDF5 format file.

3. The method for downscaling soil moisture based on dynamic and static factors according to claim 1, characterized in that, The method for obtaining the original mask in step S1 is as follows: The original satellite soil moisture data of the target area is resampled to the target resolution. Linear interpolation is performed on the resampled data. The nearest neighbor interpolation is used to fill the edges of the target area. Non-NaN regions are extracted from the resampled data as the original mask.

4. The method for downscaling soil moisture based on dynamic and static factors according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S3.1: Concatenate the static terrain features, dynamic coding features, and observation coding features along the channel dimension to generate a 192-channel fused feature map. Then, convert the format from [B, C, H, W] to [B, H, W, C] through a dimension permutation operation, where B is the batch size, C is the total number of channels, and H and W are the target space height and width. Step S3.2: Input the fused feature map into two consecutive Swing Transformer blocks, each block containing a window attention mechanism and a shift window mechanism; Step S3.3: Replace the feature dimensions after processing by the Swin Transformer block back to the [B, C, H, W] format, and then perform dimensionality reduction through two 1×1 convolutional layers. The first layer compresses 192 channels to 64 channels and activates them with ReLU. The second layer maps the 64 channels to a single-channel soil moisture prediction map while maintaining the target resolution.

5. The method for downscaling soil moisture based on dynamic and static factors according to claim 1, characterized in that, The specific steps of step S4 are as follows: Step S4.1: Based on the original mask obtained in step S1, force the predicted values ​​of the area outside the mask in the soil moisture prediction map output by the Swin Transformer block in step S3 to be NaN; Step S4.2: Extract the maximum and minimum resolution values ​​of the original satellite soil moisture data as the numerical range of numerical physics constraints. Use the truncation function to limit the resolution of the soil moisture prediction map output by the Swin Transformer block to this numerical range to obtain the final downscaling result. Step S4.3: Store the final downscaling results in HDF5 format, create a dataset to save the final downscaling results, and record the data source, resolution, and numerical range information in the file attributes.

Images