Soil moisture dynamic inversion method based on graph attention and multi-scale time feature fusion

Abstract

The invention discloses a soil moisture dynamic inversion method based on graph attention and multi-scale time feature fusion, and belongs to the technical field of intelligent remote sensing hydrology. S2, spatial feature modeling is carried out; s3, time feature modeling is carried out; step S4, spatial-temporal feature fusion; and S5, predicting a sliding window and constructing a training sample. According to the method, multivariable time sequence features are extracted by adopting the bidirectional long-short-term memory network, and a multi-head self-attention mechanism and a multi-scale time attention mechanism are combined, so that fine modeling of key moments and different time scale modes in a time sequence is realized, the expression diversity of a time sequence dependency relationship is improved, and the accuracy of time sequence modeling is improved. Integrating spatial features, context attention features and scale perception features through a feature fusion module; a sliding window prediction mechanism is further introduced, an input-output sample sequence is dynamically generated through a fixed-length window, the adaptability of the model to time sequence dynamic change is enhanced, and the prediction stability and the data utilization efficiency are also improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence remote sensing hydrology, and particularly relates to a soil moisture dynamic inversion method based on graph attention and multi-scale time feature fusion. BACKGROUND

[0002] Surface soil moisture is a crucial hydrological variable in the land-atmosphere interaction, and has a significant impact on drought monitoring, agricultural management, and ecosystem dynamics. As a key variable driven by multiple factors such as atmosphere, hydrology, and topography, soil moisture has significant spatial heterogeneity and temporal non-stationarity. In recent years, remote sensing technology has become an important data source for soil moisture monitoring due to its advantages in large-scale and continuous acquisition of surface information. At the same time, deep learning models have shown great ability in modeling nonlinear relationships and extracting high-dimensional features. The integration of the two provides a new path to improve the spatial resolution and prediction accuracy of soil moisture estimation. With the help of remote sensing and deep learning methods, high-precision prediction of soil moisture has become a hot research direction in the field of remote sensing hydrology and artificial intelligence integration.

[0003] Currently, soil moisture inversion and prediction technology based on remote sensing data has become a hot research direction in the field of remote sensing hydrology and agricultural remote sensing. Existing methods can be roughly divided into three categories: spatial feature modeling methods, time series modeling methods, and spatio-temporal joint modeling methods. Spatial feature modeling methods are represented by convolutional neural networks, which can extract local spatial features but have difficulty in capturing non-Euclidean spatial structures under complex terrain conditions. Time series modeling methods such as long short-term memory networks are good at modeling temporal dependencies but have memory degradation problems in long time series prediction. Spatio-temporal joint modeling methods can handle both spatial and temporal features, but still have deficiencies in multi-scale feature fusion and dynamic adaptation mechanisms.

[0004] Specifically, the existing technology has the following main technical problems: In the spatial dimension, traditional methods rely on regular grid representation of geographical space, which cannot truly reflect the irregularity and connectivity of surface structure; In the time dimension, existing models have difficulty in capturing both long-term trends and short-term fluctuations in soil moisture changes; In the spatio-temporal joint modeling aspect, there is a lack of adaptive ability to multi-scale dynamics, which cannot dynamically adjust the time granularity and spatial range of attention according to different conditions. These problems collectively limit the accuracy and stability of soil moisture prediction models. SUMMARY

[0005] Problems to be solved

[0006] In view of the problems raised in the existing background technology, the application provides a soil moisture dynamic inversion method based on graph attention and multi-scale time feature fusion.

[0007] Technical scheme

[0008] To solve the above problems, the application adopts the technical scheme as follows.

[0009] A soil moisture dynamic inversion method based on graph attention and multi-scale time feature fusion, the steps are as follows:

[0010] Step S1, data preprocessing: obtain monthly scale data of soil moisture, land surface temperature, precipitation, NDVI, terrain factor and soil property, uniformly resample to 1km resolution, and extract the mean value based on the hexagonal grid unit to convert into spatial point vector data; all input variables are subjected to Z-Score standardization processing, in addition, a spatial cross-validation strategy based on geographic location is introduced, the samples are divided into training set, validation set and test set according to geographic coordinates, to prevent evaluation deviation caused by adjacent point information leakage;

[0011] Step S2, spatial feature modeling: a non-Euclidean space graph structure is constructed with a hexagonal grid, the grid center is taken as a graph node, and the adjacent unit constitutes a graph edge; an improved graph attention network is used to adaptively weight and aggregate the information of the node and its neighborhood unit; the spatial coding module includes an input transformation layer, a dynamic adjacency matrix construction, a multi-layer GAT feature aggregation and a spatial feature extraction layer;

[0012] Step S3, time feature modeling: a multi-dimensional modeling framework of time sequence features is constructed by combining bidirectional long short-term memory network, multi-head self-attention mechanism and multi-scale time attention mechanism; wherein, the bidirectional LSTM captures the bidirectional dependence of long sequence, the multi-head attention captures the multi-subspace time sequence relationship in parallel, and the multi-scale attention identifies short-term fluctuations, medium-term trends and long-term evolution patterns through scale-specific query vectors and adaptive down-sampling mechanism, and fuses the features of each scale;

[0013] Step S4, spatio-temporal feature fusion: the spatial coding features in step S2 and the time context and multi-scale features in step S3 are jointly modeled, and a unified spatio-temporal representation vector is obtained by fusion, which is used for soil moisture dynamic estimation.

[0014] Step S5, sliding window prediction and training sample construction: a symmetric sliding window mechanism is used to construct input-output pairs, training samples are generated by monthly sliding, and the output results of multiple windows corresponding to the prediction time are weighted integrated to improve robustness and stability.

[0015] Preferably, the improved GAT spatial encoder in step S2 includes the following steps:

[0016] S201, input transformation layer: the two-dimensional geographic coordinates C of the node are mapped to high-dimensional spatial feature representation (including longitude and latitude information) to enhance the non-linear expression ability; N×2 ​

[0017] S202, dynamic adjacency matrix construction: based on the Euclidean distance and the Gaussian kernel function to adaptively adjust the spatial relationship weight;

[0018] S203, multi-layer GAT feature aggregation: through two-layer GAT structure to gradually aggregate neighborhood information, the first layer captures local spatial relationship, the second layer integrates more wide range of spatial context information, and introduces residual connection mechanism;

[0019] S204, feature extraction layer: convert the spatial features aggregated by multi-layer GAT into representation vectors suitable for time modeling.

[0020] Further, the specific steps of step S3 time feature modeling include:

[0021] S301, long sequence modeling of bidirectional LSTM: through bidirectional information flow design to process forward and reverse time sequence information at the same time, enhance the modeling ability of long-term dependence;

[0022] S302, multi-head self-attention processing: parallel computing multiple attention heads to capture the diversity of time sequence features from different representation subspaces;

[0023] S303, multi-scale time attention processing: through learnable query vector and adaptive downsampling strategy, respectively capture short-term fluctuations, medium-term trends and long-term evolution patterns;

[0024] S304, multi-scale feature fusion: through scale fusion network to integrate features of different time scales, and introduce residual connection mechanism.

[0025] Further, the spatio-temporal feature fusion of step S4 adopts feature splicing and full connection transformation structure to jointly map the spatial features and scale-aware time features to the final inversion space, and the specific steps are:

[0026] S401, feature splicing and mapping: splice the spatial features and time features in the feature dimension, and input to the full connection mapping layer;

[0027] S402, regularization and residual enhancement: the fusion features are regularized by LayerNorm and Dropout, and residual connection structure is introduced to improve network stability;

[0028] S403, prediction layer output and sliding window integration: input the fusion features to the full connection prediction layer, output the soil moisture prediction value corresponding to the window; since the target time may be in multiple sliding windows, the window fusion strategy is adopted to integrate the prediction values.

[0029] Preferably, in step S5, the sliding window is constructed using a symmetrical sliding strategy, that is, taking k steps forward and k steps backward to form an input window centered on the current prediction time; and performing a weighted average fusion of multiple output results from different sliding windows at the same prediction time to improve the stability of the prediction results.

[0030] Beneficial effects

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] (1) This invention uses a hexagonal grid to construct a non-Euclidean spatial graph structure, which can realistically reflect the irregularity and connectivity of the surface structure, breaking through the limitations of regular grids in expressing spatial topological relationships and better reflecting the actual spatial distribution characteristics of soil moisture. A dynamic adjacency matrix is ​​constructed based on Euclidean distance and Gaussian kernel function, which can adaptively adjust spatial relationship weights according to data distribution, effectively capturing the influence of factors such as complex terrain, vegetation, and soil type on spatial dependence, and solving the problem that static adjacency matrices cannot adapt to complex geographical relationships. By gradually aggregating neighborhood information through a two-layer graph attention network, combined with a residual connection mechanism, the limitation of the receptive field of a single-layer graph network is overcome, improving the modeling ability for long-distance spatial interactions and enhancing the model's adaptability to complex spatial heterogeneity.

[0033] (2) In this invention, the bidirectional long short-term memory network, through its bidirectional information flow design, simultaneously processes forward and reverse time series information, significantly enhancing its ability to capture long-term dependencies. This alleviates the gradient vanishing and error accumulation problems inherent in traditional recurrent networks in long-term data processing, and improves the stability of multi-step prediction. The multi-head self-attention mechanism, through parallel computation of multiple attention heads, captures the diversity of temporal features from different representation subspaces, enabling simultaneous attention to multiple important time points in the sequence. This solves the problems of information loss or uneven weight distribution in long sequences in traditional recurrent networks. By introducing a multi-scale temporal attention mechanism, learnable query vectors, and adaptive downsampling strategies, short-term fluctuations, medium-term trends, and long-term evolution patterns are captured respectively, achieving multi-scale decoupled modeling of soil moisture temporal features. This solves the feature confusion problem caused by single-scale modeling in traditional models and enhances the ability to model complex temporal dynamics.

[0034] (3) This invention constructs input-output sample pairs through a symmetrical sliding window mechanism and significantly expands the effective sample capacity by monthly sliding windows; it performs an arithmetic average of the prediction results of multiple windows at a future time, which reduces the error accumulation effect in time series prediction and improves the stability of medium- and long-term prediction.

[0035] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments. Generally, the components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0037] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0038] Example 1:

[0039] like Figure 1 As shown, a method for dynamic inversion of soil moisture based on graph attention and multi-scale temporal feature fusion includes the following steps: Step S1, Data Preprocessing: Monthly scale data of soil moisture, surface temperature, precipitation, NDVI, topographic factors, and soil properties are acquired, uniformly resampled to 1km resolution, the mean is extracted based on hexagonal grid cells and converted into spatial point vector data, and Z-score standardization is performed on all input variables. In addition, a spatial cross-validation strategy based on geographic location is adopted, dividing all samples into training set, validation set, and test set according to geographic coordinates to avoid evaluation bias caused by leakage of neighboring point information; Step S2, Spatial Feature Modeling: A non-Euclidean spatial graph structure is constructed using a hexagonal grid, with the centroid of each grid as a graph node, and adjacent hexagons forming graph edges. An improved graph attention network is used to adaptively weight and aggregate the information of nodes and their adjacent cells; Step S3, Temporal Feature Modeling: A multi-dimensional model of temporal features is constructed by combining a bidirectional long short-term memory network, a multi-head self-attention mechanism, and a multi-scale temporal attention mechanism. The framework captures dynamic evolution features at different time scales; Step S4, Spatiotemporal Feature Fusion Modeling: The spatial features output by the GAT encoder and the multi-scale temporal representation vector output by the temporal modeling module are concatenated or weighted and fused through the feature fusion network to jointly learn the spatiotemporal interaction features, and a residual connection mechanism is introduced to enhance information retention and gradient propagation; Step S5, Sliding Window Prediction and Training Set Construction: Input-output sample pairs are constructed using a symmetric sliding window mechanism, training samples are generated by monthly sliding windows, and the arithmetic average of the prediction results of multiple windows at a future time is performed to obtain the final prediction value.

[0040] In practical implementation, with a window length of L, each sample group uses historical data from L consecutive months to predict soil moisture values ​​for the following L months. By using a monthly sliding window to generate training samples cyclically, the effective sample capacity is significantly expanded. Furthermore, for situations where a future moment may be predicted by multiple sliding windows, an averaging strategy is adopted, performing an arithmetic average of the predicted values ​​obtained from multiple windows at that moment, further improving the model's ability to express continuous temporal changes and its output stability.

[0041] This application represents geospatial information as a graph structure using a hexagonal grid, with dynamic construction of connections between nodes, which can realistically reflect the irregularity and connectivity of the surface structure. It adopts a dynamic adjacency matrix based on Euclidean distance and Gaussian kernel function, which can adaptively adjust the spatial relationship weights according to the data distribution and capture the influence of factors such as complex terrain, vegetation and soil type on spatial dependence. Through two layers of GAT, neighborhood information is gradually aggregated. The first layer captures local spatial relationships, and the second layer integrates a wider range of spatial context information, which significantly improves the model's ability to model long-distance spatial interactions.

[0042] The model accurately reflects spatial structure: the design of hexagonal grid and dynamic adjacency matrix enables the model to adapt to complex surface structures, breaking through the limitations of regular grids; and enhances regional adaptability: the multi-layer GAT structure can capture long-distance spatial dependencies, significantly improving the model's adaptability to complex spatial heterogeneity.

[0043] The improved GAT spatial encoder in step S2 includes the following steps: S201, Input Transformation Layer: Mapping two-dimensional geographic coordinates to a high-dimensional feature space to enhance expressive power; S202, Dynamic Adjacency Matrix Construction: Adaptively adjusting spatial relationship weights based on Euclidean distance and Gaussian kernel function; S203, Multi-layer GAT Feature Aggregation: Gradually aggregating neighborhood information through a two-layer GAT structure. The first layer captures local spatial relationships, and the second layer integrates a wider range of spatial context information and introduces a residual connection mechanism; S204, Feature Extraction Layer: Converting the spatial features aggregated by the multi-layer GAT into a representation vector suitable for time modeling.

[0044] This application proposes a dynamic adaptation mechanism through joint modeling of spatiotemporal features. Using a dynamic adjacency matrix and a multi-layered GAT structure, the model can adaptively adjust spatial relationship weights to capture the spatial features of complex surface structures. Through a multi-scale temporal attention mechanism, the model can dynamically adjust the temporal granularity of attention, capturing short-term, medium-term, and long-term temporal features respectively. A feature fusion network is designed to jointly model spatial and temporal features, enhancing the model's expressive power.

[0045] This model can dynamically adjust its focus based on the spatiotemporal characteristics of the input data, significantly improving prediction accuracy in complex environments. The spatiotemporal joint modeling design allows the model to focus on the most important spatiotemporal features under different conditions.

[0046] The specific steps of time feature modeling in step S3 include: S301, long sequence modeling with bidirectional LSTM: by designing bidirectional information flow, forward and reverse time series information are processed simultaneously to enhance the ability to model long-term dependencies; S302, multi-head self-attention processing: multiple attention heads are computed in parallel to capture the diversity of time series features from different representation subspaces; S303, multi-scale time attention processing: short-term fluctuations, medium-term trends and long-term evolution patterns are captured respectively through learnable query vectors and adaptive downsampling strategies; S304, multi-scale feature fusion: features at different time scales are integrated through a scale fusion network and a residual connection mechanism is introduced.

[0047] The formula for the multi-scale temporal attention mechanism is:

[0048] ① Learnable query vectors: Dedicated query vectors are designed for each time scale, enabling the model to focus on key features at different scales. Short-scale features capture instantaneous changes, medium-scale features capture short-term trends, and long-scale features capture long-term patterns.

[0049]

[0050] Where s∈1,2,...,S represents the time scale, and S=3 is the number of scales set in this study.

[0051] ② Multi-scale downsampling: The input sequence is subjected to average pooling operations of different degrees to obtain representations with different temporal resolutions. This ensures that the first scale maintains the original resolution, the second scale is downsampled to half the original resolution, and the third scale is downsampled to one-quarter of the original resolution.

[0052] X s =AvgPool1D(X,kernel_size=2 s-1 )

[0053] ③ Scale-specific attention: Apply independent multi-head attention mechanisms to the feature sequences at each scale:

[0054] A s =MultiHeadAttention(X) s )

[0055] ④ Scale fusion: The attention outputs at different scales are concatenated and nonlinearly transformed to obtain the fused feature representation.

[0056] F multi=MLP(Concat[A1,A2,...,A2]) S ])

[0057] MLP consists of a linear layer, layer normalization, ReLU activation function, and Dropout class.

[0058] In this application, the bidirectional LSTM, through its bidirectional information flow design, processes both forward and backward time series information simultaneously, significantly enhancing the model's ability to capture long-term dependencies. The multi-head self-attention mechanism, by computing multiple attention heads in parallel, captures the diversity of temporal features from different representation subspaces, enabling it to simultaneously focus on multiple important time points in the sequence. By introducing a multi-scale temporal attention mechanism and employing learnable query vectors and adaptive downsampling strategies, it captures short-term fluctuations, medium-term trends, and long-term evolution patterns, respectively.

[0059] The model's significant effects include: enhanced long-term dependency modeling capabilities, with the bidirectional LSTM design significantly mitigating the gradient vanishing problem and improving the model's stability in multi-step predictions; and accurate capture of temporal dynamics, with the dual attention mechanism simultaneously focusing on global temporal dependencies and features at different temporal granularities, significantly improving the model's ability to model complex temporal dynamics.

[0060] Step S4, Feature Fusion and Result Output: After obtaining spatial and temporal features, a feature fusion module is used to jointly model the two, further improving the model's expressive power and prediction accuracy. The feature fusion module includes the following steps: S401, Feature Concatenation and Mapping: Spatial and temporal features are concatenated along the feature dimension and nonlinearly mapped through a fully connected layer to obtain a fused feature representation; S402, Regularization and Residual Connection: To prevent overfitting and enhance model stability, Dropout and LayerNorm operations are introduced, and a residual connection mechanism is added to maintain the continuity and stability of the feature flow; S403, Prediction Output Layer: The fused features are input to the final prediction layer, outputting the predicted soil moisture value at the target time. A fully connected neural network is used for regression modeling, with the mean squared error (MSE) as the loss function.

[0061] This application adopts a three-stage structure of space-time-fusion, which enables the model to first extract high-quality spatial structure information and multi-granular temporal dynamic features, and then integrate the information through a unified fusion mechanism, which significantly improves the accuracy and generalization ability of soil moisture inversion.

[0062] To enhance the model's generalization ability and prediction stability, a sliding window mechanism is used to construct the training and test sets (step S5), and multi-window fusion prediction output is implemented, specifically including the following steps:

[0063] S501. Training Sample Generation: Traverse the time series using a fixed-length symmetrical sliding window to generate input-output sample pairs. Each sample contains the time series input and the corresponding target month prediction value.

[0064] S502, Multi-window Prediction Fusion: During the testing phase, for the target time point to be predicted, it may fall within the output range of multiple sliding windows. An arithmetic averaging strategy is used to integrate multiple prediction results to improve stability.

[0065] In summary, this invention constructs a non-Euclidean spatial representation by introducing a graph attention mechanism, achieves hierarchical modeling of temporal features by combining multi-scale temporal attention, and effectively characterizes the dynamic changes of soil moisture under complex terrain and multiple spatiotemporal driving factors by leveraging a spatiotemporal feature fusion structure and a sliding window prediction mechanism. Compared with traditional methods, this invention not only improves the model's ability to perceive spatial heterogeneity and temporal dependence but also enhances the fitting effect on nonlinear evolution processes, exhibiting higher prediction accuracy and adaptability. It is suitable for multi-regional and multi-scale soil moisture retrieval tasks and has broad application prospects in remote sensing hydrology.

[0066] The above embodiments only illustrate preferred embodiments of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications, improvements, and substitutions without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for inverting soil moisture dynamics based on graph attention and multi-scale temporal feature fusion, characterized in that, The steps are as follows: Step S1: Data Preprocessing: Monthly data on soil moisture, surface temperature, precipitation, NDVI, topographic factors, and soil properties are acquired, uniformly resampled to 1km resolution, and the mean is extracted based on hexagonal grid cells to convert them into spatial point vector data; all input variables are Z-score standardized; in addition, a spatial cross-validation strategy based on geographic location is introduced, dividing the samples into training set, validation set, and test set according to geographic coordinates to prevent evaluation bias caused by leakage of neighboring point information; Step S2, Spatial Feature Modeling: Construct a non-Euclidean spatial graph structure using a hexagonal grid, with the grid centroid as the graph node and adjacent units forming the graph edge; Adaptive weighted aggregation of information from nodes and their neighboring units is performed using an improved graph attention network; The spatial coding module includes an input transformation layer, a dynamic adjacency matrix construction layer, a multi-layer GAT feature aggregation layer, and a spatial feature extraction layer. Step S3, Temporal Feature Modeling: Combining bidirectional long short-term memory networks, multi-head self-attention mechanisms, and multi-scale temporal attention mechanisms, a multi-dimensional modeling framework for temporal features is constructed. Among them, bidirectional LSTM captures the bidirectional dependencies of long sequences, multi-head attention captures the temporal relationships of multiple subspaces in parallel, and multi-scale attention identifies short-term fluctuations, medium-term trends, and long-term evolution patterns through scale-specific query vectors and adaptive downsampling mechanisms, and integrates features at each scale. Step S4, Spatiotemporal Feature Fusion: The spatial coding features in Step S2 are jointly modeled with the temporal context and multi-scale features in Step S3, and fused to obtain a unified spatiotemporal representation vector for dynamic estimation of soil moisture. Step S5, Sliding window prediction and training sample construction: The input-output pair is constructed using a symmetric sliding window mechanism. Training samples are generated by monthly sliding, and the output results of multiple windows corresponding to the prediction time are weighted and integrated to improve robustness and stability.

2. The method for soil moisture dynamic inversion based on graph attention and multi-scale temporal feature fusion according to claim 1, characterized in that: The improved GAT spatial encoder in step S2 includes the following sub-steps: S201, Input Transformation Layer: Transform the node's two-dimensional geographic coordinates C∈R N×2 (Including latitude and longitude information) is mapped to a high-dimensional spatial feature representation to enhance nonlinear expressive power; S202, Dynamic Adjacency Matrix Construction: Adaptive Adjustment of Spatial Relationship Weights Based on Euclidean Distance and Gaussian Kernel Function; S203, Multi-layer GAT feature aggregation: Neighborhood information is aggregated step by step through a two-layer GAT structure. The first layer captures local spatial relationships, and the second layer integrates a wider range of spatial context information and introduces a residual connection mechanism. S204, Feature Extraction Layer: Converts the spatial features aggregated from multiple GAT layers into representation vectors suitable for time modeling.

3. The method for soil moisture dynamic inversion based on graph attention and multi-scale temporal feature fusion according to claim 2, characterized in that: The specific steps of time feature modeling in step S3 include: S301, Long-sequence Modeling with Bidirectional LSTM: By designing a bidirectional information flow, it simultaneously processes forward and reverse time series information, enhancing the ability to model long-term dependencies. S302, Multi-head self-attention processing: Parallel computation of multiple attention heads to capture the diversity of temporal features from different representation subspaces; S303, Multi-scale temporal attention processing: Captures short-term fluctuations, medium-term trends, and long-term evolution patterns through learnable query vectors and adaptive downsampling strategies; S304, Multi-scale Feature Fusion: Integrates features from different time scales through a scale fusion network and introduces a residual connection mechanism.

4. The method for soil moisture dynamic inversion based on graph attention and multi-scale temporal feature fusion according to claim 3, characterized in that: Step S4 describes a spatiotemporal feature fusion method that employs feature splicing and fully connected transformation to jointly map spatial features and scale-aware temporal features to the final inversion space. The specific steps are as follows: S401, Feature Concatenation and Mapping: Spatial and temporal features are concatenated along the feature dimension and input into a fully connected mapping layer; S402, Regularization and Residual Enhancement: The fused features are regularized using LayerNorm and Dropout, and a residual connection structure is introduced to improve network stability; S403. Prediction layer output and sliding window integration: The fused features are input to the fully connected prediction layer, and the corresponding window's predicted soil moisture value is output. Since the target time may be within multiple sliding windows, a window fusion strategy is used to integrate the predicted values.

5. The method for soil moisture dynamic inversion based on graph attention network and multi-scale temporal feature fusion according to claim 1, characterized in that: In step S5, the sliding window is constructed using a symmetrical sliding strategy, that is, taking k steps forward and k steps backward from the current prediction time as the center to form the input window; and performing weighted average fusion on multiple output results from different sliding windows at the same prediction time to improve the stability of the prediction results.

Images