Soil moisture seamless reconstruction method and system based on graph convolution network and u-net cascade

Abstract

The application discloses a soil moisture seamless reconstruction method and system based on a graph convolution network and U-Net cascade, and the method comprises the following steps: S1, acquiring satellite-borne GNSS-R observation data, environmental auxiliary data and soil moisture benchmark data of a research area; S2, modeling the satellite-borne GNSS-R observation data as a graph node, constructing a local graph topology based on a spatial distance, and forming a dynamic graph structure; S3, using a residual graph convolution network to extract features of the dynamic graph structure, and aggregating neighbor node information to obtain a deep node embedding matrix; S4, projecting the deep node embedding matrix to a regular spatial grid according to the absolute coordinates of the graph node in the spatial grid, and generating an initial feature map; and S5, using a multi-scale U-Net network to perform spatial reasoning and gap filling on the initial feature map, and outputting a seamless soil moisture prediction field. The application aggregates sparse graph node features by using a residual graph convolution network, and performs spatial reasoning by cascading a multi-scale U-Net, so that the reconstruction from sparse observation to continuous seamless soil moisture field is realized.

Description

Technical Field

[0001] This invention relates to the field of microwave remote sensing technology, and more specifically to a method and system for seamless reconstruction of soil moisture based on graph convolutional networks and U-Net cascades. Background Technology

[0002] Soil moisture is a key variable in the Earth system, significantly impacting climate dynamics, agricultural irrigation practices, and disaster prevention and mitigation efforts. As a crucial surface state variable, soil moisture serves as a central hub connecting the terrestrial water cycle, land-atmosphere energy exchange processes, ecosystem dynamics, and extreme climate events.

[0003] Currently, methods for monitoring soil moisture mainly include traditional in-situ observation, model-based simulation, and remote sensing technology. Traditional physical in-situ monitoring methods are usually limited to point-scale observations, thus having limited ability to characterize the large-scale spatial heterogeneity of soil moisture. Microwave remote sensing currently represents the mainstream soil moisture retrieval technology, and both active and passive microwave sensors are recognized as the most effective means of monitoring soil moisture. Missions such as the Soil Moisture Active / Passive Probe Satellite (SMAP) and the Soil Moisture and Ocean Salinity Satellite (SMOS) provide large-scale spatial estimations of soil moisture, with spatial resolutions ranging from several kilometers to tens of kilometers.

[0004] Global Navigation Satellite System Reflectance (GNSS-R), as an emerging remote sensing method, inverts soil moisture by analyzing the characteristics of reflected GNSS signals, enabling continuous all-weather monitoring with short revisit times and wide spatial coverage. The Cyclone Global Navigation Satellite System (CYGNSS), as an important component of the GNSS-R constellation, has seen its enormous potential in soil moisture monitoring increasingly recognized by the academic community in recent years.

[0005] However, GNSS-R observations are highly dependent on satellite orbits and receiver field of view. Due to the irregularity of satellite ground trajectories, spatial gaps inevitably occur in the observation coverage, resulting in a pseudo-random and linearly arranged spatial distribution of observation points. In addition to these transient observation gaps, sampling density also exhibits systematic non-uniformity in geographical distribution. This spatial discontinuity and sparse, non-uniform coverage hinder the direct generation of continuous soil moisture datasets required for meteorological and hydrological simulations. Furthermore, the lack of data in key areas severely restricts the ability to detect and monitor extreme weather events or rapidly evolving hydrological processes.

[0006] To improve spatial coverage, averaging methods are commonly used to temporally aggregate multi-day GNSS-R observation data. However, this approach sacrifices instantaneous temporal resolution, and the resulting averaged data often fails to accurately reflect the true surface conditions in hydrological or agricultural applications. Traditional spatial interpolation techniques estimate values ​​for unobserved grids by mathematically interpolating fragmented observation points. Consequently, these interpolated regions rely solely on geometric relationships, neglecting the physical mechanisms (such as topography, vegetation, and soil type) that dominate soil moisture distribution. In cases of heterogeneous surface conditions and sparse observation networks, these interpolation methods frequently generate overly smoothed or biased spatial fields, limiting their ability to accurately characterize true spatial variability patterns.

[0007] In addressing the "discrete input point - continuous output surface" problem in soil moisture inversion, existing methods typically employ two independent and decoupled stages: first, geostatistical methods are used to interpolate sparse satellite trajectory points onto a dense, regular grid; then, inversion algorithms or empirical models are applied based on this. However, this step-by-step strategy of "geometric reconstruction first, then physical inversion" has inherent theoretical flaws: it severs the intrinsic connection between spatial geometric reconstruction and physical feature extraction. In the interpolation stage, the model has not yet learned the physical mapping between microwave signals and surface parameters, making it highly susceptible to introducing non-physical interpolation artifacts and over-smoothing out key local spatial heterogeneity, leading to a severe degradation in subsequent inversion accuracy.

[0008] Therefore, there is an urgent need for a method that can directly map sparse, unstructured GNSS-R observations into continuous, high-fidelity soil moisture grid products to overcome the aforementioned shortcomings of existing technologies. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for seamless reconstruction of soil moisture based on graph convolutional networks and U-Net cascades. By aggregating sparse graph node features through residual graph convolutional networks and cascading multi-scale U-Nets for spatial reasoning, the reconstruction of a continuous and seamless soil moisture field from sparse observations can be achieved.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A seamless soil moisture reconstruction method based on a graph convolutional network and a U-Net cascade includes the following steps: S1. Acquire multi-source data and soil moisture baseline data of the study area and perform preprocessing; among which, multi-source data includes: spaceborne GNSS-R observation data and environmental auxiliary data; S2. Construct a dynamic graph structure using preprocessed multi-source data, abstract the spatial units corresponding to the spaceborne GNSS-R observation data into graph nodes, and construct a local graph topology based on the spatial distance between graph nodes; S3. Use a residual graph convolutional network to extract features from the dynamic graph structure, and aggregate the physical and environmental information of neighboring nodes through a message passing mechanism to obtain a deep node embedding matrix. S4. Based on the absolute coordinates of the graph nodes in the spatial grid, project the depth node embedding matrix onto the regular spatial grid, and generate an initial feature map through explicit feature scattering operation. S5. Utilize a multi-scale U-Net network to perform multi-scale spatial reasoning and gap filling on the initial feature map, and output a spatially continuous seamless soil moisture prediction field.

[0011] Furthermore, in S1, the spaceborne GNSS-R observation data includes: surface reflectivity, signal-to-noise ratio, and bistatic radar cross-section; The environmental auxiliary data includes: digital elevation features, normalized vegetation index, and land cover type.

[0012] Furthermore, in S2, the process of constructing the dynamic graph structure is as follows: S21. Divide the study area into multiple local spatial slices; S22. Abstract the spatial units with valid spaceborne GNSS-R observation data in each local spatial slice into graph nodes; wherein, the encoding of the graph node includes: feature vectors of observation data, environmental auxiliary data, and spatial coordinates; S23. Within each local spatial slice, implement the K-nearest neighbor search algorithm based on the spatial coordinates of the effective graph nodes to filter associated nodes in the geometric neighborhood of each graph node; S24. Calculate the connection weights of directed edges using the inverse distance weighting mechanism and construct the adjacency matrix.

[0013] Furthermore, in S22, the mathematical expression for the encoding of the graph node is: ; in, For surface reflectivity, For signal-to-noise ratio and The cross-sectional area of ​​a bistatic radar characterizes the variation of the L-band surface dielectric constant; For digital elevation, For vegetation index, Land cover type is used as a static physical constraint; These are normalized floating-point coordinates in a local 128×128 coordinate system. In S24, the formula for constructing the adjacency matrix is: ; In the formula, For nodes The adjacent nodes, and They are nodes and The normalized spatial coordinate vector, where For nodes The normalized spatial coordinate vector, Let K represent the set of K nearest neighbors. A smoothing constant to ensure numerical stability.

[0014] Furthermore, in step S3, the process of obtaining the depth node embedding matrix is ​​as follows: S31. Using a multilayer perceptron as the input mapping layer, the original node features are projected onto a high-dimensional latent feature space to generate initial node representations; S32. Construct a multi-layer residual graph convolutional network and perform feature aggregation and updating by introducing an identity mapping term; S33. The high-dimensional representations of nodes after multi-layer feature aggregation are aggregated into a deep node embedding matrix.

[0015] Furthermore, in step S4, the process of projecting the depth node embedding matrix onto the regular spatial grid is as follows: S41. Employ a parameter-free and fully differentiable graph-to-mesh transformation operator; S42. Based on the absolute integer pixel index of the node, the high-dimensional graph representation of the deep node embedded in the matrix is ​​accurately located and projected into a continuous two-dimensional Euclidean space tensor through explicit feature scattering operation. The pixel positions in the non-observed area are initialized to zero vectors.

[0016] Furthermore, in S5, the process of using a multi-scale U-Net network to perform multi-scale spatial reasoning and gap filling on the initial feature map specifically includes: S51. Group normalization is used instead of batch normalization to isolate the statistical interference of spatial zero gaps on the feature distribution. S52: The encoder expands the receptive field by downsampling layer by layer, capturing large-scale macroscopic environmental context; S53. The decoder introduces skip connections, which physically concatenate the low-level fine environmental features and high-level semantic features through channel-level splicing, thereby achieving gap filling and preservation of local heterogeneity.

[0017] Furthermore, the seamless soil moisture reconstruction method also includes: optimizing the residual graph convolutional network and the multi-scale U-Net network using a dual-path supervised joint loss function based on multi-task learning; The joint loss function includes: weighted sparse regression loss, mask dense reconstruction loss, and self-supervised physical consistency loss.

[0018] Furthermore, this seamless soil moisture reconstruction method also includes: employing a slice aggregation strategy in the inference stage of the residual graph convolutional network and the multi-scale U-Net network; The slice aggregation strategy adopts a Gaussian weighted fusion method, which constructs a two-dimensional Gaussian weight matrix corresponding to the spatial dimension for the predicted slice. During global grid aggregation, all overlapping slices covering the same pixel are subjected to normalized weighted summation.

[0019] This invention also provides a system for implementing a seamless soil moisture reconstruction method based on a graph convolutional network and a U-Net cascade, comprising: The data acquisition module is used to acquire multi-source data and soil moisture baseline data of the study area and perform preprocessing; the multi-source data includes: spaceborne GNSS-R observation data and environmental auxiliary data; The graph structure construction module is used to construct a dynamic graph structure using preprocessed multi-source data, abstracting the spatial units corresponding to the spaceborne GNSS-R observation data into graph nodes, and constructing a local graph topology based on the spatial distance between graph nodes; The feature extraction module is used to extract features from the dynamic graph structure using a residual graph convolutional network, and to aggregate the physical and environmental information of neighboring nodes through a message passing mechanism to obtain a deep node embedding matrix. The spatial projection module is used to project the depth node embedding matrix onto the regular spatial grid according to the absolute coordinates of the graph nodes in the spatial grid, and generate an initial feature map through explicit feature scattering operation, wherein the grid position corresponding to the non-observation area is initialized as a zero vector; The spatial reconstruction module is used to perform multi-scale spatial inference and gap filling on the initial feature map using a multi-scale U-Net network, and output a spatially continuous seamless soil moisture prediction field.

[0020] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention constructs a dynamic graph structure by fusing multi-source environmental auxiliary data, breaking the strong dependence of traditional convolution on regular grid data. It accurately depicts the local spatial topological relationships between sparse nodes. Combining the message passing mechanism and residual connections in residual graph convolutional networks, it not only deeply aggregates the physical and environmental context information of neighboring nodes, but also effectively avoids the degradation of deep network features and preserves the unique features of the nodes themselves. This allows for the extraction of a highly expressive deep node embedding matrix. Subsequently, explicit feature scattering operations are used to adaptively map non-Euclidean graph features to Euclidean regular grids, bridging the modal gap between graph structures and image processing. Finally, relying on the powerful spatial reasoning and context restoration capabilities of multi-scale U-Net, it accurately infers and fills in the zero vector blind zone of non-observed areas, significantly improving the spatial continuity, detail integrity, and overall prediction accuracy of soil moisture reconstruction. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] The following description, in conjunction with the accompanying drawings, further illustrates the seamless soil moisture reconstruction method and system based on graph convolutional networks and U-Net cascades of the present invention. Figure 1 This is a flowchart of the seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade in Embodiment 1 of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] To better understand the purpose, structure, and function of this invention, the invention will be described in further detail below with reference to the accompanying drawings.

[0025] A seamless soil moisture reconstruction method based on a graph convolutional network and a U-Net cascade includes the following steps: S1. Acquire multi-source data of the study area and soil moisture baseline data as a supervision label; among which, multi-source data includes: spaceborne GNSS-R observation data and environmental auxiliary data; In S1, the spaceborne GNSS-R observation data includes at least surface reflectivity, signal-to-noise ratio, and bistatic radar cross-section, and the environmental auxiliary data includes at least digital elevation features, normalized vegetation index, and land cover type.

[0026] Specifically, this embodiment uses CYGNSS L1 level V3.1 scientific data product to extract three core parameters from the spaceborne GNSS-R observation data: signal-to-noise ratio (SNR), bistatic radar cross-section (BRCS), and surface reflectivity. Among them, surface reflectance It is most sensitive to changes in soil dielectric constant and was selected as the main proxy variable for characterizing soil moisture.

[0027] Environmental support data includes: topographic features extracted using a high-precision digital elevation model (DEM) released by the Space Shuttle Radar Topography Mission (SRTM); dynamic characterization of surface vegetation growth using the 16-day synthetic normalized vegetation index (NDVI) product of MOD13A2 (V061 version) from the Moderate Resolution Imaging Spectroradiometer (MODIS); and the annual land cover product of MODIS MCD12Q1 (V061) based on the International Geosphere-Biosphere Programme (IGBP) classification scheme, which finely divides the global land surface into 17 different land cover types.

[0028] The soil moisture baseline data uses the L4 global surface soil moisture product released by NASA's SMAP mission. This product integrates satellite observations and land surface hydrological models, and can provide seamless spatiotemporal continuous data with a temporal resolution of 3 hours and a spatial resolution of 9 km.

[0029] Data preprocessing stage: Removing data with an SNR lower than 0dB; For outliers less than 0, apply quantile statistical filters from 1% to 99% to eliminate extreme non-physical artifacts to the core variables; construct a global grid based on the WGS84 geographic coordinate system with a spatial resolution of 0.08°×0.08° (approximately 9km×9km) and a temporal resolution of 1 day to map all data to a unified spatiotemporal reference system; perform Min-Max normalization on all aligned feature tensors to scale the input variable values ​​to a unified range of [0,1].

[0030] S2. Based on the spaceborne GNSS-R observation data and the environmental auxiliary data, a dynamic graph structure is constructed, sparse GNSS-R observation points are modeled as graph nodes, and a local graph topology is constructed based on the spatial distance between graph nodes; In this embodiment, due to the pseudo-random nature of the CYGNSS satellite orbit, its observation data exhibits a highly sparse and unevenly distributed sampling point cloud on a spatial scale. To overcome the limitations of sparse sampling on feature representation, this invention models the soil moisture inversion problem as a node regression task on a graph structure.

[0031] The specific process of constructing a dynamic graph structure is as follows: S2.1 Divide the study area into multiple local spatial slices of 128×128 pixels each, with a sliding step size of 64 pixels. A sliding window mechanism with a 50% overlap rate is used to ensure that the model fully captures smooth neighborhood context information and effectively eliminates boundary cutting artifacts in the subsequent tile stitching stage.

[0032] S2.2 Abstract the spatial units with valid GNSS-R observation data in each local spatial slice into graph nodes. Each node Encoded as an 8-dimensional feature vector It is composed of GNSS-R observation data, environmental auxiliary data, and spatially normalized coordinates, and its mathematical expression is as follows: ; Among them, surface reflectance ( ), signal-to-noise ratio ( ) and bistatic radar cross-section ( Characterizes the variation of the L-band surface dielectric constant; Digital Elevation (DE) ), vegetation index ( ) and land cover type ( ) as a static physical constraint; These are normalized floating-point coordinates in a local 128×128 coordinate system. The feature vectors of all N valid nodes within the grid are stacked row-wise to form the input feature matrix of the graph convolution module. .

[0033] S2.3 Spatial coordinates based on valid nodes within each local data block The K-Nearest Neighbor (KNN) search algorithm is implemented. By calculating the distance between two points, the K most closely related valid nodes in the geometric neighborhood of each node are dynamically selected (K=4 in this invention) to establish a directed connection topology that crosses the sampling blind zone.

[0034] S2.4. The inverse distance weighting mechanism is used to calculate the connection weight of directed edges. The calculation formula is as follows: ; In the formula, For nodes The adjacent nodes, and They are nodes and The normalized spatial coordinate vector, where For nodes The normalized spatial coordinate vector, Let K represent the set of K nearest neighbors. A smoothing constant is used to ensure numerical stability. This physical weight setting ensures that nodes that are spatially closer have a higher contribution to message passing in graph convolution.

[0035] S3. Use a residual graph convolutional network to extract features from the dynamic graph structure, and obtain a deep node embedding matrix by aggregating the physical and environmental information of neighboring nodes. In step S3, the specific process of feature extraction using a residual map convolutional network is as follows: S31. A multilayer perceptron is used as the input mapping layer to project the original node features into a high-dimensional latent feature space. S32. Construct a multi-layer residual graph convolutional network and perform feature aggregation and updating by introducing an identity mapping term; S33. The high-dimensional representations of nodes after multi-layer feature aggregation are aggregated into a deep node embedding matrix.

[0036] S34. Through the differentiable graph to mesh bridging module, the depth node embedding matrix is ​​projected onto the regular spatial mesh according to the absolute coordinates of the node in the spatial mesh to generate an initial feature map. This embodiment is specifically as follows: GCU The first phase of the architecture aims to extract deep node embeddings from graph-structured data and fully explore the local spatial relationships between surface observation points.

[0037] First, using a multilayer perceptron ( MLP This is used as the input mapping layer to map the original 8-dimensional node features. Projecting onto a high-dimensional latent feature space generates initial node representations. This provides a high-order representation foundation for subsequent graph convolution aggregation.

[0038] To maintain the spatial heterogeneity of surface features during multi-layer information transmission and prevent the loss of local details due to over-aggregation, this invention constructs an L=4-layer residual graph convolutional network. The feature update rule for the l-th layer is defined as follows: ; In the formula, Let be the learnable weight matrix of the l-th layer. Let be the weights of the directed edges in the adjacency matrix A. (Identity mapping term) The introduction of this ensures the retention of the original characteristics of the nodes in the deep network. This is the Mish activation function. After feature aggregation through L=4 layers, the final high-dimensional representations of all nodes are integrated into a deep node embedding matrix. .

[0039] The first stage essentially acts as a learnable spatial context filter, continuously refining node representations under physical associations by aggregating neighborhood information.

[0040] In S4, the specific process of projection through the differentiable graph-to-mesh bridging module is as follows: using a parameterless and fully differentiable graph-to-mesh conversion operator, based on the absolute integer pixel index of the node, the high-dimensional graph representation embedded in the matrix of the depth node is projected into a continuous two-dimensional Euclidean space tensor through explicit feature scattering operation, and the pixel position of the non-observation area is initialized as a zero vector.

[0041] This embodiment specifically involves: embedding deep nodes generated in the first stage. Still existing in discrete graph topological domains, in order to transform them into dense tensors that can be processed by subsequent decoders, this invention designs a parameterless and fully differentiable graph-to-grid (G2G) transformation operator.

[0042] Since the graph nodes are strictly aligned with the underlying 128×128 spatial reference frame during the construction phase, the operators do not need to perform complex interpolation calculations. Instead, they use explicit feature scattering operations based on the absolute integer pixel indices provided by the nodes. The high-dimensional graph representation is precisely located and projected onto a continuous two-dimensional Euclidean space tensor: ; In the formula, The generated dense feature map is represented by F, which is the channel dimension of the embedding vector. For non-observational regions in the grid that belong to the spatial completion prediction mask, their corresponding pixel positions are uniformly initialized to zero vectors.

[0043] As a bridge connecting heterogeneous network structures, the core advantage of this operator lies in its complete differentiability. During training, the error gradient generated by the subsequent reconstruction network can be losslessly backpropagated to the front-end GCN encoder through this operator, realizing the joint dynamic optimization of graph representation and mesh completion.

[0044] S5. Use a multi-scale U-Net network to perform spatial reasoning and gap filling on the initial feature map, and output a spatially continuous seamless soil moisture prediction field.

[0045] In S5, the specific process of spatial reasoning using a multi-scale U-Net network is as follows: S51. Use group normalization instead of batch normalization for feature normalization processing; S52: The encoder expands the receptive field by downsampling layer by layer, capturing large-scale macroscopic environmental context; S53. The decoder introduces skip connections, which physically concatenate the low-level fine environmental features and high-level semantic features by splicing them together through the channel dimension.

[0046] This embodiment specifically refers to the feature map generated in the second stage. Although it possesses profound physical meaning, the sparsity of the observed trajectories results in numerous zero-value gaps in space. Therefore, in the final stage of this invention, a fully convolutional U-Net architecture is employed to perform global spatial inference, achieving high-fidelity spatial reconstruction from discrete trajectory points to a continuous physical field.

[0047] In dealing with extremely sparse In practice, the global statistics of traditional batch normalization (BN) suffer from severe bias due to the large number of zero-value pixels lowering the mean. Therefore, this invention employs group normalization (GN, with 32 groups) to strictly limit its computational domain to the channel subspace. This strategy effectively isolates the statistical interference of spatial zero-value gaps on the effective feature distribution, significantly improving the convergence and numerical stability of the network under sparse observations.

[0048] In multi-scale spatial reasoning, the encoder continuously expands its receptive field through layer-by-layer downsampling, thereby capturing large-scale macroscopic environmental context across sampling blind spots. To compensate for high-frequency details such as terrain abrupt changes and water body contours lost during downsampling, the decoder introduces skip connections. This mechanism physically concatenates the fine environmental features from the lower layers with the high-level semantic features transmitted by the decoder through channel-dimensional concatenation. The feature fusion and reconstruction process of the l-th layer can be mathematically expressed as: ; In the formula, For convolution operations, This represents the splicing operator along the channel axis. This is the convolution weight matrix at this scale. This formulaic fusion loop has a clear physical meaning: U-Net can both infer the macroscopic dryness and wetness evolution trend using upsampled features and... The forced injection of DEM and NDVI introduces local spatial heterogeneity, enabling the module to autonomously fill large-area trajectory gaps while accurately preserving water cutoffs and discontinuous boundaries in complex surface environments. The final output of this module is a localized, densely predicted tile. .

[0049] During the model training phase, this invention employs a dual-path supervised joint loss function based on multi-task learning for network optimization. This objective function not only ensures the fidelity of discrete sampling point inversion but also guarantees the hydrological rationality of the spatial field reconstructed through observation gaps via a physical constraint mechanism. Joint Loss Function Defined as: ; In the formula, , and To balance the weights of hyperparameters in relation to the magnitude and importance of each optimization objective, this invention determines the weights through empirical sensitivity analysis. , , The three types of loss supervision have different sources and complementary functions, corresponding to the three core dimensions of model optimization: the physical accuracy of the points (…). ), global spatial continuity ( ) and deep feature fidelity ( ).

[0050] Weighted Sparse Regression Loss A lightweight MLP regression head is attached to the end of the GCN encoder, relying heavily on an effective observation mask. As a spatial index, the model is directly supervised using SMAP real labels. To address the typical long-tailed distribution characteristics of soil moisture, a cost-sensitive weighted mean squared error is applied. ; Mask Dense Reconstruction Loss Focusing on the final dense prediction output of U-Net Union of the effective supervised domain Calculate the reconstruction error: ; Self-supervised physical consistency loss In the GCN stage, attribute constraint branches are introduced in parallel, and the high-dimensional embedding of GCN encoding is forced through the decoding projection head. Reverse reconstruction is performed on its original static environment inputs (such as DEM and NDVI): ; in, It is a normalized static environmental feature vector. N This represents the total number of valid nodes participating in self-supervised learning.

[0051] During the inference phase, this invention implements a large-scale inference and slice aggregation strategy to generate a seamless, global soil moisture product. In a preferred embodiment, for each predicted slice with a size of 128×128... Construct a two-dimensional Gaussian weight matrix corresponding to the spatial dimension. Its weight decreases Gaussian from the center to the edge. During global mesh aggregation, for spatial coordinates... For any grid point, the final predicted soil moisture value The calculation is performed by normalizing and weighting the sum of all overlapping slices covering this point: ; In the formula, This is a set of all slices covering the pixel. This mechanism effectively suppresses prediction noise in edge regions and eliminates non-physical spatial jumps by assigning higher fusion weights to the central region.

[0052] Through the above steps, this invention successfully transforms massive, discrete, and pseudo-random CYGNSS observation signals into a seamless, high-fidelity, and hydrophysically continuous 9km resolution daily soil moisture map, realizing a spatial scale leap from satellite signals to standardized hydrological products.

[0053] The seamless soil moisture reconstruction method based on graph convolutional networks and U-Net cascades described in this embodiment fully realizes the end-to-end direct mapping of spaceborne GNSS-R observation data from sparse discrete trajectory points to a continuous spatial grid. Through reasonable network architecture design and physical constraint mechanisms, it effectively solves the inherent defects of traditional GNSS-R soil moisture inversion techniques in terms of spatial continuity and physical consistency. The method's operation steps are logically clear and the process is standardized. The parameter settings and algorithm selections in each step are closely aligned with the actual characteristics of spaceborne GNSS-R observations, possessing good operability and engineering implementation conditions. The implementation of this method effectively enhances the application value of spaceborne GNSS-R technology in land hydrological monitoring, provides a feasible technical solution for high-precision, high spatiotemporal resolution monitoring of soil moisture in mid- and low-latitude regions globally, and lays a practical foundation for the further application and optimization of GNSS-R microwave remote sensing technology in the field of environmental monitoring.

[0054] The seamless soil moisture reconstruction method also includes: during the model training phase, a dual-path supervised joint loss function based on multi-task learning is used for network optimization. The joint loss function includes: weighted sparse regression loss, mask dense reconstruction loss, and self-supervised physical consistency loss.

[0055] The seamless soil moisture reconstruction method further includes: a slice aggregation strategy in the inference stage, wherein the slice aggregation strategy adopts a Gaussian weighted fusion method; for the prediction slice to construct a two-dimensional Gaussian weight matrix corresponding to the spatial dimension, during the global grid aggregation process, all overlapping slices covering the same pixel are normalized and weighted summed.

[0056] According to the specific embodiments provided by the present invention, the present invention also has the following technical effects: 1. Achieved end-to-end direct mapping from sparse GNSS-R observations to continuous soil moisture fields: This invention proposes a graph convolutional network and U-Net cascade architecture to directly map non-Euclidean observation space to Euclidean space grid, avoiding the cumulative error caused by the separation of spatial geometry and physical features in the traditional step-by-step strategy of "first geometric reconstruction, then physical inversion", and realizing joint dynamic optimization of feature extraction and spatial field reconstruction.

[0057] 2. Significantly improves the spatial continuity and accuracy of soil moisture retrieval: This invention aggregates dynamic dielectric features using a residual GCN encoder, achieves feature projection using a differentiable G2G bridging module, and then performs spatial inference using a multi-scale U-Net decoder, effectively filling observation gaps. Experimental verification shows that compared with the SMAP benchmark product, the correlation coefficient R reaches 0.946, and the root mean square error RMSE is 0.0367m. 3 / m 3 The error increase in areas without observation coverage is less than 5%.

[0058] 3. Ensuring the physical consistency of the inversion results: This invention introduces a self-supervised dual-path loss function, integrating the coupling relationship between surface scattering, topography, and vegetation, forcing the network to retain underlying environmental background information. By introducing environmental auxiliary data such as digital elevation, normalized vegetation index, and land cover type as physical prior constraints, the interference of vegetation attenuation and topographic distortion on microwave signals is effectively decoupled, ensuring that the reconstructed soil moisture product conforms to hydrophysical laws.

[0059] 4. Capable of monitoring extreme hydrological events: In typical climate events such as the rapid shift from drought to flood in Somalia and the monsoon rains in India, this invention successfully captured the rapid wet-dry cycle and spatiotemporal transient process of soil moisture, verifying the operational application potential of this method under extreme hydrological conditions.

[0060] 5. A feasible solution for generating seamless daily soil moisture products: This invention utilizes the high-frequency sampling advantage of the CYGNSS constellation to transform sparse daily trajectory data into spatially continuous grid products, effectively filling observation gaps and enabling seamless daily soil moisture monitoring in mid-to-low latitude regions globally at a resolution of 9km.

[0061] Example 2 This invention also provides a system for implementing a seamless soil moisture reconstruction method based on a graph convolutional network and a U-Net cascade, comprising: The data acquisition module is used to acquire multi-source data and soil moisture baseline data of the study area and perform preprocessing; the multi-source data includes: spaceborne GNSS-R observation data and environmental auxiliary data; The graph structure construction module is used to construct a dynamic graph structure using preprocessed multi-source data, abstracting the spatial units corresponding to the spaceborne GNSS-R observation data into graph nodes, and constructing a local graph topology based on the spatial distance between graph nodes; The feature extraction module is used to extract features from the dynamic graph structure using a residual graph convolutional network, and to aggregate the physical and environmental information of neighboring nodes through a message passing mechanism to obtain a deep node embedding matrix. The spatial projection module is used to project the depth node embedding matrix onto the regular spatial grid according to the absolute coordinates of the graph nodes in the spatial grid, and generate an initial feature map through explicit feature scattering operation, wherein the grid position corresponding to the non-observation area is initialized as a zero vector; The spatial reconstruction module is used to perform multi-scale spatial inference and gap filling on the initial feature map using a multi-scale U-Net network, and output a spatially continuous seamless soil moisture prediction field.

[0062] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for seamless reconstruction of soil moisture based on a graph convolutional network and a U-Net cascade.

[0063] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A seamless soil moisture reconstruction method based on a cascaded graph convolutional network and U-Net, characterized in that, Includes the following steps: S1. Acquire multi-source data and soil moisture baseline data of the study area and perform preprocessing; among which, multi-source data includes: spaceborne GNSS-R observation data and environmental auxiliary data; S2. Construct a dynamic graph structure using preprocessed multi-source data, abstract the spatial units corresponding to the spaceborne GNSS-R observation data into graph nodes, and construct a local graph topology based on the spatial distance between graph nodes; S3. Use a residual graph convolutional network to extract features from the dynamic graph structure, and aggregate the physical and environmental information of neighboring nodes through a message passing mechanism to obtain a deep node embedding matrix. S4. Based on the absolute coordinates of the graph nodes in the spatial grid, project the depth node embedding matrix onto the regular spatial grid, and generate an initial feature map through explicit feature scattering operation. S5. Utilize a multi-scale U-Net network to perform multi-scale spatial reasoning and gap filling on the initial feature map, and output a spatially continuous seamless soil moisture prediction field.

2. The seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade as described in claim 1, characterized in that, In S1, the spaceborne GNSS-R observation data includes: surface reflectivity, signal-to-noise ratio, and bistatic radar cross-section; The environmental auxiliary data includes: digital elevation features, normalized vegetation index, and land cover type.

3. The seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade as described in claim 1, characterized in that, In S2, the construction process of the dynamic graph structure is as follows: S21. Divide the study area into multiple local spatial slices; S22. Abstract the spatial units with valid spaceborne GNSS-R observation data in each local spatial slice into graph nodes; wherein, the encoding of the graph node includes: feature vectors of observation data, environmental auxiliary data, and spatial coordinates; S23. Within each local spatial slice, implement the K-nearest neighbor search algorithm based on the spatial coordinates of the effective graph nodes to filter associated nodes in the geometric neighborhood of each graph node; S24. Calculate the connection weights of directed edges using the inverse distance weighting mechanism and construct the adjacency matrix.

4. The seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade as described in claim 3, characterized in that, In S22, the mathematical expression for the encoding of the graph node is: ； in, For surface reflectivity, For signal-to-noise ratio and The cross-sectional area of ​​a bistatic radar characterizes the variation of the L-band surface dielectric constant; For digital elevation, For vegetation index, Land cover type is used as a static physical constraint; These are normalized floating-point coordinates in a local 128×128 coordinate system. In S24, the formula for constructing the adjacency matrix is: ； In the formula, For nodes The adjacent nodes, and They are nodes and The normalized spatial coordinate vector, where For nodes The normalized spatial coordinate vector, Let K represent the set of K nearest neighbors. A smoothing constant to ensure numerical stability.

5. The seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade as described in claim 1, characterized in that, In S3, the process of obtaining the depth node embedding matrix is ​​as follows: S31. Using a multilayer perceptron as the input mapping layer, the original node features are projected onto a high-dimensional latent feature space to generate initial node representations; S32. Construct a multi-layer residual graph convolutional network and perform feature aggregation and updating by introducing an identity mapping term; S33. The high-dimensional representations of nodes after multi-layer feature aggregation are aggregated into a deep node embedding matrix.

6. The seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade as described in claim 1, characterized in that, In step S4, the process of projecting the depth node embedding matrix onto the regular spatial grid is as follows: S41. Employ a parameter-free and fully differentiable graph-to-mesh transformation operator; S42. Based on the absolute integer pixel index of the node, the high-dimensional graph representation of the deep node embedded in the matrix is ​​accurately located and projected into a continuous two-dimensional Euclidean space tensor through explicit feature scattering operation. The pixel positions in the non-observed area are initialized to zero vectors.

7. The seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade as described in claim 1, characterized in that, In step S5, the process of using a multi-scale U-Net network to perform multi-scale spatial reasoning and gap filling on the initial feature map is as follows: S51. Group normalization is used instead of batch normalization to isolate the statistical interference of spatial zero gaps on the feature distribution. S52: The encoder expands the receptive field by downsampling layer by layer, capturing large-scale macroscopic environmental context; S53. The decoder introduces skip connections, which physically concatenate the low-level fine environmental features and high-level semantic features through channel-level splicing, thereby achieving gap filling and preservation of local heterogeneity.

8. The seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade as described in claim 1, characterized in that, The seamless soil moisture reconstruction method also includes: optimizing the residual graph convolutional network and the multi-scale U-Net network using a dual-path supervised joint loss function based on multi-task learning; The joint loss function includes: weighted sparse regression loss, mask dense reconstruction loss, and self-supervised physical consistency loss.

9. The seamless soil moisture reconstruction method based on graph convolutional network and U-Net cascade as described in claim 1, characterized in that, This seamless soil moisture reconstruction method also includes: a slice aggregation strategy is adopted in the inference stage of residual graph convolutional networks and multi-scale U-Net networks; The slice aggregation strategy adopts a Gaussian weighted fusion method, which constructs a two-dimensional Gaussian weight matrix corresponding to the spatial dimension for the predicted slice. During global grid aggregation, all overlapping slices covering the same pixel are subjected to normalized weighted summation.

10. A seamless soil moisture reconstruction system based on a cascaded graph convolutional network and U-Net, used to implement the seamless soil moisture reconstruction method based on a cascaded graph convolutional network and U-Net as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-source data and soil moisture baseline data of the study area and perform preprocessing; the multi-source data includes: spaceborne GNSS-R observation data and environmental auxiliary data; The graph structure construction module is used to construct a dynamic graph structure using preprocessed multi-source data, abstracting the spatial units corresponding to the spaceborne GNSS-R observation data into graph nodes, and constructing a local graph topology based on the spatial distance between graph nodes; The feature extraction module is used to extract features from the dynamic graph structure using a residual graph convolutional network, and to aggregate the physical and environmental information of neighboring nodes through a message passing mechanism to obtain a deep node embedding matrix. The spatial projection module is used to project the depth node embedding matrix onto the regular spatial grid according to the absolute coordinates of the graph nodes in the spatial grid, and generate an initial feature map through explicit feature scattering operation, wherein the grid position corresponding to the non-observation area is initialized as a zero vector; The spatial reconstruction module is used to perform multi-scale spatial inference and gap filling on the initial feature map using a multi-scale U-Net network, and output a spatially continuous seamless soil moisture prediction field.

Images