Soil humidity prediction method driven by space-time attention in domestic supercomputing environment

Abstract

The invention provides a time-space attention-driven soil humidity prediction method in a domestic supercomputing environment, and the method comprises the steps: constructing a time-space feature dynamic fusion soil humidity prediction network model based on attention guidance; the attention-guided spatio-temporal feature dynamic fusion soil humidity prediction network model comprises a convolutional long-short term memory network basic framework and a spatio-temporal attention mechanism module. The introduction of the ConvLSTM network effectively integrates the time sequence and space information of the soil humidity, the potential space-time dependence in the data is fully utilized, the space-time attention mechanism enables the model to adaptively pay attention to important time periods and space regions by dynamically adjusting the weight between time steps, and the accuracy of the model is improved. The limitation of fixed feature selection in a traditional method is avoided, so that the prediction precision is improved. The model can better capture space-time dynamic feature information, further identifies the relationship between the influence factor and the soil humidity, and especially shows unique advantages in complex space-time feature processing and modeling of long-time sequence data.

Description

Technical Field

[0001] The present invention relates to the field of intelligent remote sensing technology, and in particular to a spatiotemporal attention-driven soil moisture prediction method under a domestic supercomputing environment. Background Art

[0002] Soil moisture (SM) is a key state variable in climate, hydrology, and ecosystems. Its dynamics are the result of the synergistic effects of multiple factors, including meteorological conditions, surface characteristics, and biological activity. It is influenced by factors such as meteorological elements, soil texture, vegetation type, and topography. Soil moisture has a complex dependency on meteorological elements such as precipitation and soil temperature. Precipitation dominates soil water replenishment, while temperature regulates water dissipation through evapotranspiration. Meteorological elements, due to their well-defined physical mechanisms and good temporal continuity, are often used as core variables in soil moisture prediction models. They provide a foundation for constructing predictive models with clear mechanisms and spatial and temporal scalability. Appropriate soil moisture is crucial for agricultural production and farmland irrigation management. Therefore, developing accurate soil moisture prediction models can provide a scientific decision-making engine for dynamic early warning of drought and flood disasters, precise allocation of irrigation resources, and enhancing agricultural disaster resilience.

[0003] In recent years, machine learning has shown certain advantages in the field of soil moisture prediction. It reduces the dependence on data completeness through a data-driven approach and plays an important role in small and medium-sized data sets and relatively simple prediction task scenarios. Reference: Cécile G, Dharumarajan S, Jean-Baptiste F, et al. Use of Sentinel-2 Time-Series Images for Classification and Uncertainty Analysis of Inherent Biophysical Property: Case of Soil Texture Mapping[J]. Remote Sensing, 2019, 11(5): 565. Machine learning has significant limitations in capturing the temporal dependency of soil moisture. It is difficult to accurately grasp the dynamic changes of soil moisture over time, and its ability to capture time series patterns such as long-term trends and cyclical changes is relatively weak. When dealing with spatial heterogeneity problems, traditional machine learning models are difficult to fully and effectively consider the complexity of the spatial distribution of soil moisture, resulting in distortion of spatiotemporal feature decoupling, which in turn restricts prediction accuracy. Literature: Wang Jingping, Wu Xiaodan, Ma Dujuan, et al. Remote sensing inversion based on machine learning: analysis of uncertainty factors [J]. Journal of Remote Sensing, 2023, 27(03): 790-801.

[0004] Time series prediction models based on deep learning show better processing performance in soil moisture prediction tasks. Reference: Hinton GE, Osindero S, Teh Y WA Fast Learning Algorithm for DeepBeliefNets[J]. Neural Computation, 2006, 18(7): 1527-1554. In short-term soil moisture prediction tasks, deep learning models can fully consider the continuous changes in the time dimension and the distribution differences in the spatial dimension, thereby capturing more complex spatiotemporal features. Reference: Tan J, NourEldeen N, Mao K, et al. Deep Learning Convolutional Neural Network for the Retrieval of Land Surface Temperaturefrom AMSR2 Data in China[J]. Sensors, 2019, 19(13): 2987-2987. However, when processing long time series information, deep learning models generally have the problem of losing key hidden state information. To address this limitation, the attention mechanism performs weighted summation of information at different time steps, enabling the model to flexibly allocate attention resources according to the current prediction task requirements. Reference: Han J, Hong J, Chen X, et al. Integrating Convolutional Attention and Encoder-Decoder Long Short-Term Memory for Enhanced Soil Moisture Prediction[J]. Water, 2024, 16(23): 3481-3481, which can more accurately capture important patterns, dependencies, and potential laws in time series. Reference: Li X, Zhang Z, Li Q, et al. Enhancing Soil Moisture Forecasting Accuracy with REDF-LSTM: Integrating Residual En-Decoding and Feature Attention Mechanisms[J]. Water, 2024, 16(10): 1376-1400. Existing soil moisture prediction models often face problems such as differences in the temporal and spatial distribution of data, which in turn lead to feature offsets, when dealing with cross-scenario prediction tasks.Transfer learning can identify the shared knowledge between different tasks and transfer it to new tasks to achieve better performance. Reference: Kara A, Pekel E, Ozcetin E, et al. Genetic algorithm optimized a deep learning method with attention mechanism for soil moisture prediction[J]. Neural Computing and Applications, 2024, 36(4): 1761-1772. Domain adaptation technology is used to align the feature distribution of different scenes, and hierarchical fine-tuning is used to correct the model, thereby alleviating the feature offset problem and ensuring that the model still has high accuracy in new tasks.

[0005] However, existing soil moisture prediction models still have certain limitations. When dealing with the coordinated optimization of complex spatiotemporal dynamic representation and cross-scenario sample generalization, there are problems such as insufficient spatial feature mining, limited temporal dynamic modeling, and weak adaptability to cross-scenario data distribution differences. These problems make it difficult to balance the sudden response of farmland scenarios with the need for cross-regional generalization. Specifically, these limitations are reflected in two aspects: (1) In terms of spatiotemporal feature extraction, soil moisture changes have a time constant response and spatial heterogeneity. Traditional methods are difficult to achieve the dynamic coupling representation of sudden and normal meteorological characteristics, resulting in information attenuation of key features during the transmission process, thereby restricting the perception sensitivity of the model; (2) In terms of cross-scene feature adaptation, there are differences in spatial resolution between remote sensing data and measured site data. The alignment criteria of the existing model are relatively simple, making it difficult to effectively align the temporal misalignment features of soil moisture remote sensing data and measured data, thereby limiting the migration performance of the model.

[0006] Therefore, a spatiotemporal attention-driven soil moisture prediction method is provided in a domestic supercomputing environment to solve the above technical problems. Summary of the Invention

[0007] The purpose of this invention is to provide a spatiotemporal attention-driven soil moisture prediction method based on feature dependencies in a domestic supercomputer environment. This method, which integrates heterogeneous networks, can fully extract feature features and improve accuracy, time complexity, and generalization capability. To address the insufficient exploitation of spatiotemporal features in soil moisture prediction tasks, this invention proposes an attention-guided spatiotemporal feature dynamic fusion soil moisture prediction network model (AGSMP-Net). First, based on a convolutional long short-term memory network, the network leverages the fusion characteristics of convolution and recurrent structures to accurately capture the interactive relationship between soil moisture and meteorological factors in different temporal and spatial dimensions. Next, a "feature-time-space" attention mechanism module is constructed to replace the fixed weight model used in traditional convolution. This module dynamically selects key temporal and spatial features, focusing on time series information processing and capturing changes in spatial distribution, grasping long-term trends in soil moisture, and optimizing information utilization in the spatiotemporal dimensions. Compared to other soil moisture prediction models, the proposed attention-guided spatiotemporal feature dynamic fusion soil moisture prediction network model demonstrates improved accuracy and generalization capability.

[0008] The object of the present invention is achieved like this: A spatiotemporal attention-driven soil moisture prediction method in a domestic supercomputer environment is characterized by constructing a soil moisture prediction network model based on the dynamic fusion of spatiotemporal features guided by attention. The attention-guided spatiotemporal feature dynamic fusion soil moisture prediction network model includes a convolutional long short-term memory network basic framework and a spatiotemporal attention mechanism module; the spatiotemporal attention mechanism module includes a feature attention mechanism, a time attention mechanism, and a spatial attention mechanism to enhance the prediction accuracy of the model; Constructing an attention-guided spatiotemporal feature dynamic fusion soil moisture prediction network model includes the following steps: Step S1: Design a pixel-based soil moisture spatiotemporal information extraction branch ConvLSTM The soil moisture spatiotemporal information extraction branch is based on the Convolutional Long Short-Term Memory Network (Convolutional LSTM, ConvLSTM). The Convolutional Long Short-Term Memory Network is a convolutional long short-term memory neural network that combines the CNN network with the LSTM network at the bottom of the model. In order to obtain both temporal and spatial information, the ConvLSTM model changes the fully connected layer in the LSTM network to a convolutional layer. While processing multidimensional data, it can obtain the basic spatial features in the multidimensional data through convolution operations. Through convolution operations, it can effectively capture the local characteristics of soil moisture in the spatial dimension; the LSTM network unit processes information in the temporal dimension, remembers the changing trend of soil moisture over time, and solves the spatiotemporal dependency problem of soil moisture caused by meteorological conditions. Step S2: Constructing the spatiotemporal attention mechanism module The feature-time-space attention mechanism module specifically includes a feature attention mechanism, a time attention mechanism, and a space attention mechanism; the feature attention mechanism focuses on the complex nonlinear relationship between soil moisture and various meteorological elements. By calculating the correlation weights between different features, it can extract the spatiotemporal features that have a significant impact on soil moisture, thereby strengthening the role of these key features in model calculation and effectively solving the problem of spatial heterogeneity; the time attention mechanism focuses on processing information in the time series, dynamically allocating attention weights for different time steps, paying attention to the long-term trend of soil moisture changes, and keenly capturing the mutation moments in the time series, so that the model can more accurately grasp the dynamic changes of soil moisture in the time dimension; the spatial attention mechanism focuses on the differences in soil moisture in different geographical locations, highlights the characteristics of key areas, suppresses interference from irrelevant areas, effectively copes with the non-uniformity of soil moisture in spatial distribution, and improves the model's ability to learn the spatial variation laws of soil moisture, thereby providing support for accurate prediction of soil moisture while comprehensively considering spatiotemporal factors; Step S3: Validate and compare the model using the dataset A comparative validation of the soil moisture prediction network model with dynamic fusion of spatiotemporal features is conducted to verify the outstanding performance of the proposed method in complex soil moisture prediction scenarios.

[0009] The specific steps of step S1 are: Pixel-based soil moisture spatiotemporal information extraction branch ConvLSTM ConvLSTM replaces the fully connected layer in the LSTM network with a convolutional layer. While processing multidimensional data, it obtains the basic spatial features in the multidimensional data through convolution operations. The key equations of ConvLSTM are shown in Equations (1) to (5), where “*” represents the convolution operator and “∴” represents the Hadamard product: i t =σ(W xi*X t +W hi *H t-1 +W ci ∴C t-1 +b i )#(1) f t =σ(W xi *X t +W hi *H t-1 +W ci ∴C t-1 +b i )#(2) C t =f t ∴C t-1 +i t ∴tan h(W xc *X t +W hc *H t-1 +b c )#(3) o t =σ(W xo *X t +W ho *H t-1 +W co ∴C t-1 +b0)#(4) H t =o t ∴tanh(C t )#(5) Where σ is the sigmoid activation function, tanh represents the hyperbolic tangent activation function, and C t-1 is the memory unit of the previous time step, h t is the current hidden layer state, x t is the current input, f t 、i t 、C t 、o t They are forget gate, input gate, memory unit and output gate respectively, W and b are model parameters; The study conducted experiments based on the ConvLSTM network, with the following specific settings: The input data consists of influencing factors on days t, t-1, and t-2, including soil temperature, precipitation, and soil moisture. The feature map size is 61×71, so the input size of the model is (3×3×61×71), where the first "3" represents the data of three time steps and the second "3" represents the three influencing factors. The input data first enters the ConvLSTM layer containing 16 filters. The convolution kernel size of this layer is set to 3, and the hyperbolic tangent function "tanh" is selected as the activation function. In this layer, the spatial and temporal features of the time series data are effectively captured by performing convolution operations and time recursive processing on the input data. The data enters the second ConvLSTM layer, which contains 8 filters. The convolution kernel size and activation function remain unchanged. This layer further processes the data to extract deeper spatiotemporal features. To ensure that the array size before and after the convolution operation is consistent, the data is zero-padded during the processing process.

[0010] The specific steps of step S2 are: The purpose of the feature attention mechanism is to weight the time step invisible state of the data processed by the ConvLSTM network, so as to obtain temporal features and prediction features; the feature attention mechanism uses the fully connected layer Net Linea#1 Generate attention weights for the output of each time step, and then activate them through an exponential function: exp(x)=e x #(6) Where x is the function input The exponential function exp ensures that important time step features are amplified to obtain higher weights; Net Linea#1 The transfer tensor is from [bach size, T, N, d] size to [bach size, T, N, 1] size. The softmax layer is defined as follows: Where xt is the original attention score and t is the time step; The softmax layer is used for normalization. In order to ensure the limited additivity of weights, the feature attention mechanism identifies the time importance of each influencing factor. The feature attention is generated by the neural network Net Linea#1 and a softmax layer, which is defined as follows: Where xt is the original input, t is the time step, and “⊙” represents element-by-element multiplication; [FP1, FP2, FP3](FPn∈Rd) is based on the linear neural network Net Linea#2 The generated prediction features convert the tensor mf_xt from [batch size, 3, N, d] to [batch size, 1, N, d]; The spatial attention mechanism consists of two linear neural networks and a softmax layer. The output of the previous part can be spliced ​​with the h_t hidden state to obtain the feature vector, which is then passed through the linear neural network Net Linea#2, the fully connected layer is used to calculate the attention weight; the calculated weight is activated by the exponential function, and then the weight and the eigenvalue are calculated to obtain the final result, which is defined as follows: pred a =exp(Net Linea#2 (x t ))#(10) Where pred_a is the attention weight and pred_FC is the feature vector, x t is the original input, t is the time step, and i represents the sequence number. The spatial attention mechanism enhances the correlation between spatial features, and finally the concatenation vector The feature vectors at each time step are aggregated and their correlations are enhanced to obtain rich spatial features. Its linear neural network transfers a tensor of size [batch size, N, 2*d] to a tensor of size [batch size, N, 1]. For different influencing factors, different weights need to be assigned to them according to their importance. The temporal attention mechanism also consists of two linear neural networks and a softmax layer. It distinguishes the importance of each time step by assigning weights to it. It concatenates the temporal features obtained by feature attention with the input data to obtain a new temporal feature vector. After activation by an exponential function, the weights and feature values ​​are calculated to obtain the final result. It is defined as follows: temp a =exp(Net Linea#3 (x t ))#(12) In the formula, temp_a is the attention weight and temp_FC is the feature vector, x t is the original input, t is the time step, and i represents the sequence number. The temporal attention mechanism is based on the concatenation vector of temporal features and input data. The temporal feature vectors of all influencing factors are aggregated to enhance the temporal correlation of the predictor variables. Its linear neural network transfers the tensor [batch size, T, 3] to the tensor [batch size, T, 1]. The operation of the temporal attention mechanism allows us to obtain rich temporal features, and the temporal attention weights can explain the temporal correlation between predictor variables.

[0011] Beneficial effects: The present invention realizes the efficient capture of spatiotemporal characteristics of soil moisture and the powerful response to complex prediction scenarios through innovative design. Specifically, the pixel-based soil moisture spatiotemporal information extraction branch ConvLSTM effectively captures the local characteristics of soil moisture in the spatial dimension through convolution operations, and uses LSTM network units to process information in the time dimension, remembering the changing trend of soil moisture over time, and can solve the spatiotemporal dependency problem of soil moisture caused by meteorological conditions. On this basis, a spatiotemporal attention mechanism module is constructed, which includes feature attention mechanism, temporal attention mechanism and spatial attention mechanism. The feature attention mechanism digs deep into the complex nonlinear correlation between soil moisture and meteorological elements, strengthens key spatiotemporal characteristics, and breaks through the bottleneck of spatial heterogeneity; the temporal attention mechanism dynamically allocates time step weights, which not only closely follows long-term trends, but also keenly captures mutation moments, and accurately presents the dynamics of the time dimension; the spatial attention mechanism focuses on regional humidity differences, highlights key areas, shields interference, and greatly improves the learning efficiency of spatial change laws. The three work together to fully promote the accurate prediction of soil moisture. After verification and comparison of data sets, the dynamic fusion network model of spatiotemporal features has demonstrated excellent performance in complex scenarios, and has the strong ability to cope with various actual prediction conditions, providing solid support for the practical application of accurate soil moisture prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a structural diagram of the attention-guided spatiotemporal feature dynamic fusion soil moisture prediction network model (AGSMP-Net) of the present invention; Figure 2 Schematic diagram of the ERA5-Land dataset involved in the present invention; Figure 3 and Figure 4 Comparison diagrams of the visualization results of soil moisture prediction using the ERA5-Land dataset by the present invention and other methods are shown respectively; Figure 5 Graph showing the effect of the superpixel segmentation scale λ on the FSSC-GCN classification performance. (a) is the result graph of the HN-LULC dataset; (b) is the result graph of the GID dataset. DETAILED DESCRIPTION

[0013] The present invention will be further described below with reference to the accompanying drawings and examples.

[0014] like Figure 1As shown in the figure, a soil moisture prediction method driven by spatiotemporal attention in a domestic supercomputing environment is constructed, which constructs an attention-guided spatiotemporal feature dynamic fusion soil moisture prediction network model (AGSMP-Net). The model consists of two parts: a convolutional long short-term memory network framework and a spatiotemporal attention mechanism module. The ConvLSTM network, as the model's basic architecture, utilizes the fusion characteristics of convolution and recurrent structures in the network to accurately capture the interactive relationship between soil moisture and meteorological factors in different temporal and spatial dimensions. The attention mechanism module is subdivided into three parts: feature attention mechanism, temporal attention mechanism, and spatial attention mechanism. It is used to replace the fixed weight mode in traditional convolution, dynamically screen key temporal and spatial features, focus on time series information processing and capture changes in spatial distribution, grasp the long-term trend of soil moisture, and optimize information utilization in spatiotemporal dimensions.

[0015] like Figure 2 As shown, the data of Henan Province for each prediction factor. The time range covered by RA5-Land data is from 1950 to the present, updated every two to three months, providing hourly time resolution data and a spatial resolution of 0.1 degrees. ERA5-Land contains 50 climate variables, covering many aspects of surface water and energy cycles, such as temperature, precipitation, soil moisture, evapotranspiration, snow depth, etc. The article selects three climate variables: soil temperature, precipitation and soil moisture. The soil temperature is selected from the 0-7cm soil temperature, the unit is K; the soil moisture is selected from the water content per unit volume of the 0-7cm soil layer, the unit is m3 / m3; the total precipitation is the average depth of accumulated precipitation, the unit is m.

[0016] The present invention selects daily data from Henan Province from 2015 to 2024 for modeling (the number of daily samples is 3592) and preprocesses it, including steps such as data segmentation, normalization, feature and label separation. First, the soil temperature, soil moisture and precipitation data are processed. The soil temperature and humidity are averaged every 24 hours, and the precipitation is summed every 24 hours to characterize its daily scale data. Then the data is divided into a training set and a test set according to a ratio of 7:3. 70% of the data is used as a training set for training the prediction model; the remaining 30% of the data is used as a test set to verify the prediction effect of the model.

[0017] like Figure 3The figure shows the scatter density results for soil moisture prediction. The most widely used and latest models for soil moisture prediction were selected to compare and validate the proposed AGSMP-Net prediction method, demonstrating the effectiveness and outstanding performance of the proposed method in land classification based on high-resolution remote sensing images. The models selected for comparison include SVR, CNN, LSTM, Transformer, and ConvLSTM models. Experimental results show that for predictions for the next three days, the AGSMP-Net model has a coefficient of determination (R2) of 0.806 and a root mean square error (RMSE) of 0.0570. These results are significantly higher than those of SVR, CNN, LSTM, Transformer, and ConvLSTM (R2: 0.566, 0.615, 0.658, 0.717, and 0.758, respectively), and RMSE: 0.1331, 0.1193, 0.1105, 0.0816, and 0.0693, respectively).

[0018] The predicted values ​​of the six models are roughly linearly related to the true values, and for most of the predicted points, Figure 3 The red concentrated part in the image can be predicted well. The SVR model has good nonlinear fitting ability based on the support vector machine regression method, but its processing ability for large data sets is limited and it is not good at processing time series data. Figure 3 It can be seen that the results are relatively scattered, which reflects the instability of the model in the soil moisture prediction task.

[0019] The scatter plots from both the CNN and LSTM models exhibit a certain overall offset from the x=y line, meaning that the predicted results tend to be slightly larger or smaller than the true values. Because the CNN and LSTM models focus only on the spatial and temporal characteristics of soil moisture, respectively, neither can simultaneously capture predictive information from both features. CNNs primarily consist of convolutional layers, pooling layers, and fully connected layers. Through convolution operations, they achieve local connections and weight sharing, effectively capturing local features and spatial structure in the image. LSTMs build on recurrent neural networks by incorporating cell states and gating mechanisms. Through this gating mechanism, LSTMs can retain and update information in time series data, thereby capturing long-term dependencies. These two models ignore the spatiotemporal correlations of the data. Therefore, while they achieve some predictive effectiveness, they suffer from large errors and low reliability.

[0020] The basic structure of the Transformer model mainly consists of two parts: the encoder and the decoder. Each part is composed of multiple identical layers stacked together to efficiently process sequence information. Compared with LSTM, the Transformer model has stronger processing capabilities for capturing long-range dependencies. However, its predictive modeling of long-term series dependencies faces some challenges. Changes in soil moisture are affected by long-term climate factors and seasonal changes. The long-term dependencies required for its prediction task cannot be effectively captured by a single Transformer architecture, thus affecting its prediction accuracy.

[0021] ConvLSTM builds on LSTM by replacing fully connected layers with convolutional computations, replacing matrix multiplication with convolutional computations. This allows it to capture information in both temporal and spatial dimensions. However, existing models, based on a parameter-sharing mechanism, construct a globally unified convolutional kernel weight matrix, making it difficult to effectively characterize the spatially heterogeneous response relationships between soil moisture and driving factors. This static convolutional architecture, constrained by the fixed geometric constraints of the local receptive field, cannot dynamically adapt to the nonstationary coupling of soil water transport and meteorological factors within different geographic units, leading to systematic deviations in the accuracy of modeling spatiotemporally differentiated features. The AGSMP-Net model builds on this model by incorporating an attention mechanism to consider the temporal importance of temporal features and the temporal importance of individual influencing factors. The former captures the global relationships between influencing factors, while the latter distinguishes the temporal correlations among influencing factors. This further enhances the model's temporal information processing capabilities and expands the spatial information perception range, thereby improving model performance.

[0022] like Figure 4 Figure 2 shows the spatial distribution of predicted and true soil moisture values. This figure also presents a visualization of the spatial distribution of predicted and true soil moisture values ​​for the four seasons in Henan Province in 2024, using the AGSMP-Net model. Analysis shows that the AGSMP-Net model can effectively predict the spatial distribution of soil moisture in Henan Province. Regionally, due to relatively low soil moisture levels in northern and western Henan Province, the AGSMP-Net model performs well overall, with predicted values ​​close to the true values ​​and exhibiting high fitting accuracy. However, in southeastern Henan Province, where soil moisture is higher, the model's predicted values ​​are generally lower than the true values, exhibiting a certain degree of bias. Temporally, the model's predictions for spring and summer are highly accurate, with good agreement with the true values. Autumn model predictions show a significant underestimation in eastern Henan Province, while in western Henan, the predictions still reflect the true soil moisture conditions well. Winter model predictions show an overestimation in northern Henan Province, but perform well in southern Henan, with predicted values ​​closely aligned with the true values.

[0023] like Figure 5The figure shows the impact of different ratios of meteorological inputs on model prediction accuracy. To further investigate the importance of total precipitation (tp) and soil temperature (stl) in soil moisture prediction, the experiment set different input ratios for these two factors and recorded the changes in the prediction accuracy of the AGSMP-Net model under these inputs. The experimental results show that compared with the case of a 1:1 ratio of precipitation and soil temperature input, inputting either precipitation or soil temperature alone results in a decrease in model prediction accuracy. Notably, the decrease in model accuracy when inputting precipitation alone is smaller than when inputting soil temperature alone. Furthermore, reducing the input ratio of precipitation decreases model accuracy, while reducing the input ratio of soil temperature improves model accuracy.

Claims

1. A spatiotemporal attention-driven soil moisture prediction method based on a domestic supercomputer environment, characterized by: A soil moisture prediction network model based on the dynamic fusion of spatiotemporal features guided by attention is constructed. The model includes a convolutional long short-term memory network framework and a spatiotemporal attention mechanism module. The spatiotemporal attention mechanism module includes feature attention mechanism, time attention mechanism, and spatial attention mechanism to enhance the prediction accuracy of the model. Constructing an attention-guided spatiotemporal feature dynamic fusion soil moisture prediction network model includes the following steps: Step S1: Design a pixel-based soil moisture spatiotemporal information extraction branch ConvLSTM The soil moisture spatiotemporal information extraction branch is based on a convolutional long short-term memory network (Convolutional LSTM, ConvLSTM). The convolutional long short-term memory network is a convolutional long short-term memory neural network that combines a CNN network with an LSTM network at the bottom of the model. In order to simultaneously obtain temporal and spatial information, the ConvLSTM model replaces the fully connected layer in the LSTM network with a convolutional layer. While processing multidimensional data, it can obtain the basic spatial features in the multidimensional data through convolution operations. Through convolution operations, it can effectively capture the local characteristics of soil moisture in the spatial dimension. The LSTM network unit processes information in the temporal dimension, remembers the changing trend of soil moisture over time, and solves the spatiotemporal dependency problem of soil moisture caused by meteorological conditions. Step S2: Constructing the spatiotemporal attention mechanism module The feature-time-space attention mechanism module specifically includes a feature attention mechanism, a time attention mechanism, and a space attention mechanism; the feature attention mechanism focuses on the complex nonlinear relationship between soil moisture and various meteorological elements, and extracts spatiotemporal features that have a significant impact on soil moisture by calculating the correlation weights between different features, thereby strengthening the role of these key features in model calculation and effectively solving the problem of spatial heterogeneity; the time attention mechanism focuses on processing information in the time series, dynamically allocating attention weights for different time steps, paying attention to the long-term trend of soil moisture changes, and keenly capturing the mutation moments in the time series, so that the model can more accurately grasp the dynamic changes of soil moisture in the time dimension; the spatial attention mechanism focuses on the differences in soil moisture in different geographical locations, highlights the characteristics of key areas, suppresses interference from irrelevant areas, effectively copes with the non-uniformity of soil moisture in spatial distribution, and improves the model's ability to learn the spatial variation law of soil moisture, thereby providing support for accurate prediction of soil moisture while comprehensively considering spatiotemporal factors; Step S3: Validate and compare the model using the dataset A comparative validation of the soil moisture prediction network model with dynamic fusion of spatiotemporal features is conducted to verify the outstanding performance of the proposed method in complex soil moisture prediction scenarios.

2. The soil moisture prediction network model based on dynamic fusion of spatiotemporal features according to claim 1 is characterized by: The specific steps of step S1 are: Pixel-based soil moisture spatiotemporal information extraction branch ConvLSTM The soil moisture spatiotemporal information extraction branch ConvLSTM changes the fully connected layer in the LSTM network to a convolutional layer. While processing multidimensional data, it obtains the basic spatial features in the multidimensional data through convolution operations. The key equations of ConvLSTM are shown in Equations (1) to (5), where "*" represents the convolution operator and "∴" represents the Hadamard product: i t =σ(W xi *X t +W hi *H t-1 +W ci ∴C t-1 +b i )#(1) f t =σ(W xi *X t +W hi *H t-1 +W ci ∴C t-1 +b i )#(2) C t =f t ∴C t-1 +i t ∴tanh(W xc *X t +W hc *H t-1 +b c )#(3) o t =σ(W xo *X t +W ho *H t-1 +W co ∴C t-1 +b0)#(4) H t =o t ∴tanh(C t )#(5) Where σ is the sigmoid activation function, tanh represents the hyperbolic tangent activation function, and C t-1 is the memory unit of the previous time step, h t is the current hidden layer state, x t is the current input, f t 、i t 、C t 、o t They are forget gate, input gate, memory unit and output gate respectively, W and b are model parameters; The study conducted experiments based on the ConvLSTM network, with the following specific settings: The input data consists of influencing factors on days t, t-1, and t-2, including soil temperature, precipitation, and soil moisture. The feature map size is 61×71, so the model input size is (3×3×61×71), where the first "3" represents the data for three time steps, and the second "3" represents the three influencing factors. The input data first enters the ConvLSTM layer, which contains 16 filters. The convolution kernel size of this layer is set to 3, and the hyperbolic tangent function "tanh" is selected as the activation function. In this layer, the spatial and temporal features of the time series data are effectively captured by performing convolution operations and time recursion on the input data. The data enters the second ConvLSTM layer, which contains 8 filters. The convolution kernel size and activation function remain unchanged. This layer further processes the data to extract deeper spatiotemporal features. To ensure that the array size before and after the convolution operation is consistent, the data is zero-padded during processing.

3. The method for land use classification based on heterogeneous convolutional neural network remote sensing images based on feature dependency according to claim 1 is characterized by: The specific steps of step S2 are: The purpose of the feature attention mechanism is to weight the time-step invisible state of the data processed by the ConvLSTM network to obtain temporal features and prediction features; Feature attention mechanism uses the fully connected layer Net Linea#1 Generate attention weights for the output of each time step, and then activate them through an exponential function: exp(x)=e x #(6) Where x is the function input The exponential function exp ensures that important time step features are amplified to obtain higher weights; Net Linea#1 The transfer tensor is from [batch size, T, N, d] to [batch size, T, N, 1], and the softmax layer is defined as follows: Where x t is the raw attention score, t is the time step; The softmax layer is used for normalization. In order to ensure the limited additivity of weights, the feature attention mechanism identifies the time importance of each influencing factor. The feature attention is generated by the neural network Net Linea#1 and a softmax layer, which is defined as follows: Where x t is the original input, t is the time step, and "⊙" represents element-wise multiplication; [FP1, FP2, FP3](FPn∈Rd) is based on the linear neural network Net Linea#2 The generated prediction features convert the tensor mf_xt from [batch size, 3, N, d] to [batch size, 1, N, d]; The spatial attention mechanism consists of two linear neural networks and a softmax layer. The output of the previous part can be spliced ​​with the h_t hidden state to obtain the feature vector, which is then passed through the linear neural network Net Linea#2 , the fully connected layer is used to calculate the attention weight; the calculated weight is activated by the exponential function, and then the weight and the eigenvalue are calculated to obtain the final result, which is defined as follows: pred a =exp(Net Linea#2 (x t ))#(10) Where pred_a is the attention weight and pred_FC is the feature vector, x t is the original input, t is the time step, and i represents the sequence number; The spatial attention mechanism enhances the correlation between spatial features, and finally the concatenation vector The feature vectors at each time step are aggregated and their correlation is enhanced to obtain rich spatial features. Its linear neural network transfers a tensor of size [batch size, N, 2*d] to a tensor of size [batch size, N, 1]. For different influencing factors, the study needs to assign different weights to them according to their importance. The temporal attention mechanism also consists of two linear neural networks and a softmax layer. It distinguishes the importance of each time step by assigning weights to it. It concatenates the temporal features obtained by feature attention with the input data to obtain a new temporal feature vector. After activation by an exponential function, the weights and feature values ​​are calculated to obtain the final result. It is defined as follows: temp a =exp(Net Linea#3 (x t ))#(12) In the formula, temp_a is the attention weight and temp_FC is the feature vector, x t is the original input, t is the time step, and i represents the sequence number; The temporal attention mechanism is based on the concatenation vector of temporal features and input data. The temporal feature vectors of all influencing factors are aggregated to enhance the temporal correlation of the predictor variables. Its linear neural network transfers the tensor [batch size, T, 3] to the tensor [batch size, T, 1]. The operation of the temporal attention mechanism allows us to obtain rich temporal features, and the temporal attention weights can explain the temporal correlation between predictor variables.

Images