Day-by-day root zone soil humidity prediction method based on causal coupling space-time model
By constructing a causal coupled spatiotemporal Transformer model, the problem of insufficient integration of causal reasoning and spatiotemporal prediction models in existing technologies is solved, achieving high accuracy and robustness in soil moisture prediction, which is applicable to agricultural irrigation scheduling, ecosystem drought early warning, and watershed hydrological cycle simulation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-05-05
AI Technical Summary
Existing research mostly focuses on time prediction, neglecting the importance of the spatial dimension. There is a lack of systematic research that integrates causal reasoning with spatiotemporal prediction models, resulting in high uncertainty in soil moisture prediction results.
We construct a causal coupled spatiotemporal Transformer model, extract the causal relationship between soil moisture and environmental variables through a causal discovery algorithm, and combine dynamic causal weight adjustment and hierarchical multi-level feature extraction strategies to integrate prior causal knowledge and data-driven dynamic dependencies to model complex spatiotemporal dependencies.
It improves the accuracy and robustness of soil moisture prediction, can remain stable in scenarios with fluctuating environmental variables, captures long-range time memory effects and multi-scale spatial heterogeneity, and adapts to soil moisture changes under different environmental conditions.
Smart Images
Figure CN121980485A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of soil moisture monitoring, and in particular to a daily root zone soil moisture prediction method based on a causal coupled spatiotemporal model. Background Technology
[0002] Soil moisture is a key variable regulating the exchange of water and energy fluxes between the land surface and the atmosphere, and is an important component of the global hydrological cycle. In 2010, soil moisture was listed as a key climate variable for studying the global climate system due to its widespread importance. Topsoil moisture (approximately 0-5 cm) mainly affects processes such as surface evaporation and near-surface energy balance; while deep soil moisture (approximately 0-100 cm) dominates underground processes, including water uptake by plant roots and groundwater recharge, and is crucial for long-term water resource availability and ecosystem sustainability. Spatiotemporal prediction of soil moisture, that is, the process of estimating its future spatial distribution and temporal changes through predictive models, is of great significance for decision support in applications such as crop yield optimization, irrigation planning, and drought management.
[0003] Soil moisture prediction methods can be broadly categorized into process-driven models and data-driven models. Process-driven models simulate dynamic changes in soil moisture by solving physical equations describing land-atmosphere interactions (such as the Richards equation). While these models have made significant contributions to soil moisture prediction, their inherent nonlinearity and limitations in comprehensively considering factors such as topography, soil properties, and land use lead to considerable uncertainty in the prediction results. With the increasing abundance of multi-source heterogeneous observational data, data-driven machine learning models are constantly evolving, encompassing traditional statistical methods, machine learning algorithms, and deep learning models. Deep learning models are considered capable of directly extracting the relationship between input features and target variables from observational data, without needing to characterize complex physical processes in detail. Classic deep learning models used for soil moisture prediction include temporal prediction models such as Long Short-Term Memory (LSTM) networks, and spatiotemporal prediction models such as Convolutional LSTM networks. However, pure deep learning models may face problems such as capturing spurious correlations, while incorporating prior knowledge (including causal relationships) into deep learning models provides a feasible approach to solving these problems. Causal relationships reflect the directionality and meaningful interactions between variables, going beyond simple correlations and revealing the intrinsic mechanisms driving dynamic changes in the system. Embedding causal relationships allows models to more accurately capture the impact of historical variables on the target variable, thereby improving prediction accuracy.
[0004] Existing research shows that incorporating causal information into deep learning models through causal feature selection, structural optimization, and loss function constraints can improve model performance. For example, causal feature selection based on the Peter-Clark instantaneous conditionally independent causal inference method can effectively improve the prediction accuracy of deep learning models for irrigation water consumption; incorporating causal information into a long short-term memory network model for wind speed prediction by introducing a neighborhood threshold shows that the causal long short-term memory network model significantly improves prediction accuracy; and incorporating causal information into the loss function to constrain the prediction of wetland methane emissions by a long short-term memory network model has shown that embedding causal relationships into the prediction model can effectively improve model performance. However, most existing research focuses on temporal prediction, neglecting the importance of the spatial dimension, and lacks systematic research on integrating causal inference with spatiotemporal prediction models. Summary of the Invention
[0005] The purpose of this invention is to address the problem that most existing studies focus on time prediction while neglecting the importance of the spatial dimension and lack systematic research that integrates causal reasoning with spatiotemporal prediction models, and to provide a daily root zone soil moisture prediction method based on a causal coupled spatiotemporal model.
[0006] The above-mentioned objective of this application is achieved through the following technical solution: S1. Obtain data on soil moisture and environmental variables in the root zone and construct a dataset; S2. Use a causal discovery algorithm to extract the strength of the causal relationship between root zone soil moisture and environmental variables, and form prior causal knowledge; S3. Construct a causal attention module with a dynamic causal weight adjustment mechanism to integrate prior causal knowledge with data-driven dynamic dependencies. S4. Construct a spatiotemporal shift window Transformer module with a hierarchical multi-level feature extraction strategy to model complex spatiotemporal dependencies; S5. The causal attention module and the spatiotemporal shift window Transformer module are connected in series and fused to construct a causal coupled spatiotemporal Transformer model and trained using a dataset; S6. Input the new environmental variable data into the trained causal coupled spatiotemporal Transformer model to obtain the soil moisture prediction results.
[0007] Optionally, step S1 includes: Obtain daily root zone soil moisture and environmental variable data for a historical period in the study area and construct a dataset; divide the dataset into training dataset, validation dataset, and test dataset, wherein the daily soil moisture data for the subsequent 10 days corresponding to the test dataset is used as label data. The environmental variable data include: surface soil moisture, temperature, precipitation, 10-meter horizontal wind speed, 10-meter vertical wind speed, downward long-wave radiation, and downward short-wave radiation.
[0008] Optionally, step S2 includes: The causal discovery algorithm is the Peter-Clarke instantaneous conditional independence algorithm, specifically including: The parent node set for initializing the root zone soil moisture is all environmental variables and their own lagged variables from the past 1 to 7 days; Variables with a causal relationship with root zone soil moisture were screened using partial correlation tests at a significance level of 0.05. Quantify causal strength into a static causal weight matrix ,in This represents the total number of variables.
[0009] Optionally, step S3 includes: The implementation of the causal attention module includes: Construct a feature extraction network consisting of 3D convolutional layers, grouping normalization, SiLU activation function, and 3D convolutional layers, and output a dynamic dependency matrix. ; static causality weight matrix and The components are concatenated, processed through a network with the same structure, and output as a dynamic causal weight matrix. ; Through Hadamaji Applying to input features to achieve causal reinforcement:
[0010] in For input features, For enhanced features; Features are further extracted using 3D convolutional layers with 7×7×7 and 1×1×1 kernels, and residual connections are introduced to preserve the original information.
[0011] Optionally, step S4 includes: The hierarchical multi-level feature extraction strategy includes three stages: Temporal pattern capture stage: reshaping input features into A two-layer shifted window Transformer structure is used to extract time dependencies; Spatial dependency extraction stage: reshaping features into A 6-layer shifted window Transformer structure is used to extract spatial dependencies; Spatiotemporal feature integration stage: reorganizing features into A two-layer shift window Transformer structure is used to integrate spatiotemporal features; in, For batch size, For the length of time, For the height and width of the space, For the embedded dimension.
[0012] Optionally, step S5 includes: The construction of the causal coupled spatiotemporal Transformer model includes: The three-dimensional spatiotemporal sequence data is sequentially input into the causal attention module and the spatiotemporal shift window Transformer module; the three-dimensional spatiotemporal sequence data is the spatiotemporal sequence of root zone soil moisture and environmental variable data. High-dimensional features are reconstructed into a three-dimensional output with the same spatial resolution as the input by transposed 3D convolution. An autoregressive iterative approach is used, taking the previous day's forecast as the input for the next day, to generate daily soil moisture forecasts for the next 10 days.
[0013] Optionally, step S6 includes: The loss function of the causal coupled spatiotemporal Transformer model is the mean absolute error:
[0014] in For predicted values, For the true value, The number of samples; An early stopping strategy is adopted, and training is stopped when the validation set loss does not decrease for 20 consecutive rounds. Accuracy evaluation metrics for causal coupled spatiotemporal Transformer models: mean absolute error, unbiased root mean square error, Pearson correlation coefficient, and peak signal-to-noise ratio.
[0015] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform a daily root zone soil moisture prediction method based on a causal coupled spatiotemporal model.
[0016] A computer-readable storage medium storing instructions that, when executed, perform a daily root zone soil moisture prediction method based on a causal coupled spatiotemporal model.
[0017] The beneficial effects of the technical solution provided in this application are: This invention constructs a daily soil moisture spatiotemporal prediction model by building a causal attention module (including a dynamic causal weight adjustment mechanism) and a spatiotemporal shift window Transformer module. The causal attention module integrates prior causal knowledge (extracted based on the PCMCI algorithm) with data-driven dynamic dependencies, strengthening true correlation features through Hadamard product, effectively eliminating spurious correlations between temperature and soil moisture, and improving the model's robustness in scenarios with fluctuating environmental variables. The spatiotemporal shift window Transformer module employs a time-space hierarchical attention mechanism, accurately capturing the long-range temporal memory effect of root zone soil moisture (e.g., the influence of soil moisture from the previous two days on subsequent days) and multi-scale spatial heterogeneity (e.g., local moisture differences caused by topography) through shift windows and dynamic scale adjustments, overcoming the limitations of single-dimensional modeling. This invention will provide technical support for scenarios such as precision irrigation scheduling in agriculture, ecosystem drought early warning, and watershed hydrological cycle simulation. Attached Figure Description
[0018] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the daily soil moisture causal prediction method according to an embodiment of the present invention; Figure 2 This is a structural diagram of the causal attention module according to an embodiment of the present invention; Figure 3 This is a structural diagram of the spacetime shifting window Transformer module according to an embodiment of the present invention; Figure 4 This is a structural diagram of the CCTST causal spatiotemporal prediction model according to an embodiment of the present invention; Figure 5 This is a diagram showing the causal discovery results in the Qinghai-Tibet Plateau region according to an embodiment of the present invention; Figure 6 This is a comparison chart of the daily prediction accuracy of the CCTST model in the Qinghai-Tibet Plateau region according to an embodiment of the present invention with that of other models; Figure 7 This is a statistical distribution comparison of the prediction bias of the CCTST model in the Qinghai-Tibet Plateau region and other models under different environmental conditions according to an embodiment of the present invention. Figure 8 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation
[0019] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0020] The embodiments of this application provide a method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model.
[0021] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a daily root zone soil moisture prediction method based on a causal coupled spatiotemporal model in an embodiment of this application, including: S1. Obtain data on soil moisture and environmental variables in the root zone and construct a dataset; S2. Use a causal discovery algorithm to extract the strength of the causal relationship between root zone soil moisture and environmental variables, and form prior causal knowledge; S3. Construct a causal attention module with a dynamic causal weight adjustment mechanism to integrate prior causal knowledge with data-driven dynamic dependencies. S4. Construct a spatiotemporal shift window Transformer module with a hierarchical multi-level feature extraction strategy to model complex spatiotemporal dependencies; S5. The causal attention module and the spatiotemporal shift window Transformer module are connected in series and fused to construct a causal coupled spatiotemporal Transformer model and trained using a dataset; S6. Input the new environmental variable data into the trained causal coupled spatiotemporal Transformer model to obtain the soil moisture prediction results.
[0022] Step S1 includes: Obtain daily root zone soil moisture and environmental variable data for a historical period in the study area and construct a dataset; divide the dataset into training dataset, validation dataset, and test dataset, wherein the daily soil moisture data for the subsequent 10 days corresponding to the test dataset is used as label data. The environmental variable data include: surface soil moisture, temperature, precipitation, 10-meter horizontal wind speed, 10-meter vertical wind speed, downward long-wave radiation, and downward short-wave radiation.
[0023] Step S2 includes: The causal discovery algorithm is the Peter-Clarke instantaneous conditional independence algorithm, specifically including: The parent node set for initializing the root zone soil moisture is all environmental variables and their own lagged variables from the past 1 to 7 days; Variables with a causal relationship with root zone soil moisture were screened using partial correlation tests at a significance level of 0.05. Quantify causal strength into a static causal weight matrix ,in This represents the total number of variables.
[0024] As one example, the root zone soil moisture variable is initialized. The parent node set consists of all environmental variables from the past 7 days and the root zone soil moisture itself (i.e., ); iteratively remove nodes with parent nodes that are concentrated in the partial correlation test (significance level set at 0.05). Conditionally independent variables are retained, and variables with true causal associations are preserved; the causal strength obtained from the test is quantified into a static causal weight matrix. ( (Containing 1 root zone soil moisture variable + 7 environmental variables), as prior causal knowledge, among which... Indicates the first The variable for the first The causal strength of each variable.
[0025] Step S3 includes: The implementation of the causal attention module includes: Construct a feature extraction network consisting of 3D convolutional layers, grouping normalization, SiLU activation function, and 3D convolutional layers, and output a dynamic dependency matrix. ; static causality weight matrix and The components are concatenated, processed through a network with the same structure, and output as a dynamic causal weight matrix. ; Through Hadamaji Applying to input features to achieve causal reinforcement:
[0026] in For input features, For enhanced features; Features are further extracted using 3D convolutional layers with 7×7×7 and 1×1×1 kernels, and residual connections are introduced to preserve the original information.
[0027] As one embodiment, a causal attention module with a dynamic causal weight adjustment mechanism is constructed to integrate prior causal knowledge with data-driven dynamic dependencies. (Reference) Figure 2 , Figure 2 This is a schematic diagram of the causal attention module according to an embodiment of the present invention. Figure 2 The 'Causal Mask' in this context refers to the prior causal knowledge matrix extracted by the causal discovery algorithm.
[0028] Step S4 includes: The hierarchical multi-level feature extraction strategy includes three stages: Temporal pattern capture stage: reshaping input features into A two-layer shifted window Transformer structure is used to extract time dependencies; Spatial dependency extraction stage: reshaping features into A 6-layer shifted window Transformer structure is used to extract spatial dependencies; Spatiotemporal feature integration stage: reorganizing features into A two-layer shift window Transformer structure is used to integrate spatiotemporal features; in, For batch size, For the length of time, For the height and width of the space, For the embedded dimension.
[0029] As one embodiment, reference Figure 3 , Figure 3 This is a schematic diagram of the Transformer module for spatiotemporal shifting windows in an embodiment of the present invention.
[0030] Step S5 includes: The construction of the causal coupled spatiotemporal Transformer model includes: The three-dimensional spatiotemporal sequence data is sequentially input into the causal attention module and the spatiotemporal shift window Transformer module; the three-dimensional spatiotemporal sequence data is the spatiotemporal sequence of root zone soil moisture and environmental variable data. High-dimensional features are reconstructed into a three-dimensional output with the same spatial resolution as the input by transposed 3D convolution. An autoregressive iterative approach is used, taking the previous day's forecast as the input for the next day, to generate daily soil moisture forecasts for the next 10 days.
[0031] As one embodiment, the causal attention module and the spatiotemporal shift window Transformer module are fused together in series to form a causal coupled spatiotemporal Transformer model (CCTST). (Reference) Figure 4 , Figure 4 This is a structural diagram of the CCTST causal spatiotemporal prediction model according to an embodiment of the present invention. Specifically, the causal attention module and the spatiotemporal shift window Transformer module are fused in series to form a causal coupled spatiotemporal Transformer model (CCTST). Specifically, the original 3D data first passes through the causal attention module to output x_"causal", and then is input into the spatiotemporal shift window Transformer module. The high-dimensional features output by the module are reconstructed into single-step prediction features through a "transposed 3D convolutional layer". Then, through autoregressive iteration (the prediction result of the previous day is used as the input of the next day), the daily soil moisture prediction results for the next 10 days are generated.
[0032] In one embodiment, the original 3D input data (spatiotemporal sequences of soil moisture and meteorological variables) first enters the causal attention module. After dynamic causal weight adjustment, it outputs causal-enhanced features. This process strengthens true causal relationships and suppresses spurious correlations, providing a more reliable feature foundation for subsequent spatiotemporal modeling. The causal-enhanced features are input into the spatiotemporal shift window Transformer module, undergoing three levels of processing: temporal pattern capture, spatial dependency extraction, and spatiotemporal feature integration. This process gradually uncovers multi-scale 3D spatiotemporal patterns, and finally, through transposed 3D convolution, the high-dimensional features are reconstructed into a 3D output with the same spatial resolution as the input. Through autoregressive iteration, daily soil moisture prediction results for the study area for the next 10 days are generated.
[0033] Step S6 includes: The loss function of the causal coupled spatiotemporal Transformer model is the mean absolute error:
[0034] in For predicted values, For the true value, The number of samples; An early stopping strategy is adopted, and training is stopped when the validation set loss does not decrease for 20 consecutive rounds. Accuracy evaluation metrics for causal coupled spatiotemporal Transformer models: mean absolute error, unbiased root mean square error, Pearson correlation coefficient, and peak signal-to-noise ratio.
[0035] As one example, the model training uses the AdamW optimizer with an initial learning rate of 0.0005 and weight decay of . The learning rate was adjusted using the CosineAnnealingLR strategy.
[0036] In a further embodiment of the present invention, the selected study area is the Qinghai-Tibet Plateau region. In the experiment, a sliding window of length 2 was used to select root zone soil moisture and environmental variables from the past two days as input data to predict soil moisture for the next 10 days. Using the ERA5 reanalysis dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF) as the data source, a comparative experiment and robustness analysis were designed. The specific implementation process and results are as follows.
[0037] First, the PCMCI causal discovery algorithm was used to identify the causal relationship between root zone soil moisture and environmental variables. The results are as follows: Figure 5As shown, under a 1-day lag, the causal strength of each variable on root zone soil moisture was ranked as follows: lag root zone soil moisture (0.66) > precipitation (0.408) > surface soil moisture (0.133) > downward longwave radiation (0.087) > 10-meter vertical wind speed (0.052) > downward shortwave radiation (0.048) > air temperature (0.042) > 10-meter horizontal wind speed (0.031). The lag root zone soil moisture showed the highest causal strength, indicating that root zone soil moisture exhibits "state continuity" and can be considered a "memory bank" of soil moisture. Precipitation was the second largest influencing factor, further affecting root zone soil moisture after replenishing surface soil moisture through infiltration. Surface soil moisture, as the interface between the atmosphere and deep soil, regulates root zone soil moisture through processes such as evaporation and capillary rise.
[0038] Five mainstream spatiotemporal prediction models were selected as baselines for comparison with the CCTST model proposed in this invention, including ConvLSTM, E3D-LSTM, PredRNN, PredRNN++, and MIM models. All models used consistent or default parameter configurations to ensure fairness in the comparison. The mean absolute error (MAE) was used as the loss function for all models, and the MAE, unbiased root mean square error (RMSE), correlation coefficient, and peak signal-to-noise ratio (PSNR) were used to evaluate model performance. Figure 6 The accuracy changes of each model in predicting daily root zone soil moisture over the next 1-10 days are shown. Combined with the comprehensive performance indicators presented in Table 1, the following conclusions can be drawn: CCTST significantly outperforms the baseline model on all evaluation indicators, and compared with the classic ConvLSTM, the mean absolute error is reduced from 0.0146... It dropped to 0.0055 The decrease reached 62.33%, and the unbiased root mean square error decreased from 0.0135. It dropped to 0.0070 The accuracy decreased by 48.15%, the correlation coefficient improved from 0.79 to 0.92, and the peak signal-to-noise ratio improved from 33.38 dB to 40.53 dB. The prediction accuracy of all models decreased with increasing prediction days, but CCTST exhibited the slowest decay rate. For 1-day predictions, CCTST's unbiased root mean square error was 0.0062. The CCTST error is only 45.93% of that of ConvLSTM. In 10-day prediction, the unbiased root mean square error of CCTST is 0.0089 m³ / m³, which is significantly lower than that of ConvLSTM (0.0134). The CCTST model achieves 66.42% of the accuracy of traditional models such as ConvLSTM and E3D-LSTM, demonstrating its superior ability to model long-term temporal dependencies. Furthermore, the shift-window Transformer outperforms traditional models like ConvLSTM and E3D-LSTM, proving the advantage of its shift-window mechanism in capturing spatiotemporal features. Compared to the shift-window Transformer, CCTST reduces the mean absolute error by 43.30% and the unbiased root mean square error by 27.08%, demonstrating the accuracy-enhancing effect of the causal attention module.
[0039] Table 1 Comparison of the overall prediction performance of each model
[0040] As one example, Figure 7 This study presents the prediction bias distribution of various models under different environmental conditions on the Tibetan Plateau. The evaluation metric is the mean absolute error bias (MAB). The environmental conditions included are season, precipitation, topsoil moisture, root zone soil moisture, vegetation cover, soil type, and altitude. Seasons are categorized as spring, summer, autumn, and winter; soil types are categorized as coarse soil, fine soil, organic soil, and medium soil. Other attributes are categorized by their values into quartiles: <25%, 25%-50%, 50%-75%, and >75%. Statistical analysis shows that, compared to other models, the CCTST model consistently exhibits a bias closer to zero and a narrower bias range across various conditions. For example, pixels characterized by coarse soil texture show a significantly larger bias range in the baseline model than in the CCTST model. The CCTST model effectively addresses the challenges posed by extreme conditions (such as high-intensity precipitation events or drought periods) and maintains robust performance under different moisture conditions. Furthermore, the CCTST model maintains accuracy across different altitudinal gradients, further demonstrating its ability to explain the additional complexity of soil moisture dynamics caused by topographic changes. These comprehensive assessments across multiple environmental factors provide compelling evidence for the generalization ability of the CCTST model and its adaptability in capturing soil moisture dynamics under heterogeneous environmental conditions.
[0041] In summary, this invention proposes a causal attention module and a spatiotemporal shift window Transformer module, and couples these two modules to form a causal coupled spatiotemporal Transformer model (CCTST). This invention validates the daily root zone soil moisture spatiotemporal prediction performance of the CCTST model in the Qinghai-Tibet Plateau region. Prediction accuracy comparisons show that CCTST outperforms all baseline models in 1-10 day predictions, and exhibits the slowest accuracy decay rate. Bias analysis under different environmental conditions demonstrates that CCTST maintains low bias and significant robustness under various scenarios including different seasons, precipitation, soil types, vegetation, and altitudes. In conclusion, CCTST, through dynamic causal weight adjustment and hierarchical spatiotemporal feature extraction, effectively solves the problem that traditional models struggle to capture complex causal relationships and spatiotemporal dependencies, providing an efficient and reliable method for daily root zone soil moisture prediction.
[0042] This application also discloses an electronic device. (See reference...) Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.
[0043] The communication bus 502 is used to enable communication between these components.
[0044] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.
[0045] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).
[0046] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the above-described method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model.
[0047] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.
[0048] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.
Claims
1. A method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model, characterized in that, The method includes the following steps: S1. Obtain data on soil moisture and environmental variables in the root zone and construct a dataset; S2. Use a causal discovery algorithm to extract the strength of the causal relationship between root zone soil moisture and environmental variables, and form prior causal knowledge; S3. Construct a causal attention module with a dynamic causal weight adjustment mechanism to integrate prior causal knowledge with data-driven dynamic dependencies. S4. Construct a spatiotemporal shift window Transformer module with a hierarchical multi-level feature extraction strategy to model complex spatiotemporal dependencies; S5. The causal attention module and the spatiotemporal shift window Transformer module are connected in series and fused to construct a causal coupled spatiotemporal Transformer model and trained using a dataset; S6. Input the new environmental variable data into the trained causal coupled spatiotemporal Transformer model to obtain the soil moisture prediction results.
2. The method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model as described in claim 1, characterized in that, Step S1 includes: Obtain daily root zone soil moisture and environmental variable data for a historical period in the study area and construct a dataset; divide the dataset into training dataset, validation dataset, and test dataset, wherein the daily soil moisture data for the subsequent 10 days corresponding to the test dataset is used as label data. The environmental variable data include: surface soil moisture, temperature, precipitation, 10-meter horizontal wind speed, 10-meter vertical wind speed, downward long-wave radiation, and downward short-wave radiation.
3. The method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model as described in claim 1, characterized in that, Step S2 includes: The causal discovery algorithm is the Peter-Clarke instantaneous conditional independence algorithm, specifically including: The parent node set for initializing the soil moisture in the root zone is all environmental variables and their own lagged variables from the past 1 to 7 days; Variables with a causal relationship with root zone soil moisture were screened using partial correlation tests at a significance level of 0.
05. Quantify causal strength into a static causal weight matrix ,in This represents the total number of variables.
4. The method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model as described in claim 3, characterized in that, Step S3 includes: The implementation of the causal attention module includes: Construct a feature extraction network consisting of 3D convolutional layers, grouping normalization, SiLU activation function, and 3D convolutional layers, and output a dynamic dependency matrix. ; static causality weight matrix and The components are concatenated, processed through a network with the same structure, and output as a dynamic causal weight matrix. ; Through Hadamaji Applying to input features to achieve causal reinforcement: in As input features, For enhanced features; Features are further extracted using 3D convolutional layers with 7×7×7 and 1×1×1 kernels, and residual connections are introduced to preserve the original information.
5. The method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model as described in claim 1, characterized in that, Step S4 includes: The hierarchical multi-level feature extraction strategy includes three stages: Temporal pattern capture stage: reshaping input features into A two-layer shifted window Transformer structure is used to extract time dependencies; Spatial dependency extraction stage: reshaping features into A 6-layer shifted window Transformer structure is used to extract spatial dependencies; Spatiotemporal feature integration stage: reorganizing features into A two-layer shift window Transformer structure is used to integrate spatiotemporal features; in, For batch size, For the length of time, For the height and width of the space, For the embedded dimension.
6. The method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model as described in claim 1, characterized in that, Step S5 includes: The construction of the causal coupled spatiotemporal Transformer model includes: The three-dimensional spatiotemporal sequence data is sequentially input into the causal attention module and the spatiotemporal shift window Transformer module; the three-dimensional spatiotemporal sequence data is the spatiotemporal sequence of root zone soil moisture and environmental variable data. High-dimensional features are reconstructed into a three-dimensional output with the same spatial resolution as the input by transposed 3D convolution. An autoregressive iterative approach is used, taking the previous day's forecast as the input for the next day, to generate daily soil moisture forecasts for the next 10 days.
7. The method for predicting daily root zone soil moisture based on a causal coupled spatiotemporal model as described in claim 1, characterized in that, Step S6 includes: The loss function of the causal coupled spatiotemporal Transformer model is the mean absolute error: in For predicted values, For the true value, The number of samples; An early stopping strategy is adopted, and training is stopped when the validation set loss does not decrease for 20 consecutive rounds. Accuracy evaluation metrics for causal coupled spatiotemporal Transformer models: mean absolute error, unbiased root mean square error, Pearson correlation coefficient, and peak signal-to-noise ratio.
8. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform the daily root zone soil moisture prediction method based on a causal coupled spatiotemporal model as described in any one of claims 1-7.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the daily root zone soil moisture prediction method based on a causal coupled spatiotemporal model as described in any one of claims 1-7.