An intelligent soil moisture condition prediction method of AttnConvLSTM fused with CLDAS

Abstract

The application discloses a kind of AttnConvLSTM soil moisture intelligent prediction methods of fusing CLDAS, comprising: collecting the soil moisture data of CLDAS and the meteorological data of corresponding area observed by weather observation station;The relationship between meteorological data and soil moisture data is analyzed using Pearson correlation coefficient and Granger causality test, and the meteorological data with high correlation is selected;AttnConvLSTM soil moisture prediction model is constructed, high-correlation meteorological data is used as input, corresponding soil moisture data is used as label value, combined with the spatial attention of precipitation anomaly, AttnConvLSTM soil moisture prediction model is trained until mean square error loss function converges;Real-time acquisition of weather observation station forecast meteorological data and precipitation anomaly, input into trained AttnConvLSTM soil moisture prediction model, and the soil moisture of corresponding area is predicted.The application realizes high-precision, robust prediction of soil moisture.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of soil moisture prediction, in particular to an AttnConvLSTM soil moisture intelligent prediction method fusing CLDAS. BACKGROUND

[0002] Soil moisture, i.e. soil humidity condition, is of great significance to agricultural production, ecological environment monitoring, water resource management and many other fields. Accurate soil moisture prediction can help farmers arrange irrigation reasonably, improve water resource utilization efficiency, prevent drought and flood disasters from damaging crops, and thus ensure food security. In terms of ecological environment, soil moisture affects ecological processes such as vegetation growth and land surface evaporation, and accurate prediction is helpful for the protection and restoration of ecological systems.

[0003] Currently, traditional soil moisture prediction methods mainly include physical model-based and statistical model-based methods. Physical model-based methods, such as land surface models and hydrological models, rely on complex land-atmosphere feedback control equations and auxiliary elements such as precipitation, soil temperature and evapotranspiration observed in situ. However, these models have limitations in representing key processes and human activities, and are computationally expensive, making it difficult to meet the demand for real-time and high-precision prediction. Statistical models establish prediction models by statistically analyzing historical data, but often fail to capture the complex spatiotemporal variation characteristics of soil moisture and are less adaptable to nonlinear and non-stationary data.

[0004] With the development of deep learning technology, data-driven models have shown great predictive power in many fields. Convolutional Neural Network (CNN) can effectively extract spatial features of data, and Long Short-Term Memory (LSTM) is good at handling long-term dependencies in time series data. Convolutional Long Short-Term Memory (ConvLSTM) combines the advantages of CNN and LSTM and has great potential in capturing both temporal relationships and spatial features of detection targets, and has been applied in some hydrological and climate variable predictions.

[0005] However, existing ConvLSTM-based soil moisture prediction still has some problems: on the one hand, when considering the historical dynamic changes of soil moisture, the model does not fully exploit the complex spatiotemporal characteristics, resulting in limited prediction accuracy; on the other hand, many current studies only consider a single parameter of soil moisture and do not fully combine other factors related to soil moisture change trends, such as soil temperature, evapotranspiration and rainfall, which also affect the prediction performance of the model. SUMMARY

[0006] In view of the problems in the prior art, the application provides an AttnConvLSTM soil moisture intelligent prediction method fusing CLDAS, so as to solve the problems of low soil moisture prediction accuracy and insufficient consideration of the influence of multiple meteorological elements in the prior art, and realize high-precision and robust prediction of soil moisture.

[0007] To achieve the above technical purposes, the application adopts the following technical solutions:

[0008] An AttnConvLSTM soil moisture intelligent prediction method fusing CLDAS comprises the following steps:

[0009] Step S1: Collecting soil moisture data of CLDAS and meteorological data of a meteorological observation station observing the corresponding region;

[0010] Step S2: Analyzing the relationship between the meteorological data and the soil moisture data by using the Pearson correlation coefficient and the Granger causality test, and screening out meteorological data with high correlation;

[0011] Step S3: Constructing an AttnConvLSTM soil moisture prediction model based on Seq2Seq;

[0012] Step S4: Taking the meteorological data with high correlation as the input of the AttnConvLSTM soil moisture prediction model, taking the corresponding soil moisture data as the label value, combining the spatial attention of the precipitation anomaly, training the AttnConvLSTM soil moisture prediction model until the mean square error loss function converges, and completing the training of the AttnConvLSTM soil moisture prediction model;

[0013] Step S5: Real-time acquisition of the meteorological data predicted by the meteorological observation station and the precipitation anomaly, input into the trained AttnConvLSTM soil moisture prediction model, and prediction of the soil moisture of the corresponding region.

[0014] Further, the meteorological data with high correlation screened out in step S2 is meteorological data with an absolute value of the Pearson correlation coefficient greater than 0.5 and a p value of the Granger causality test less than 0.05, and the meteorological data with high correlation includes precipitation, air temperature, relative humidity and radiation.

[0015] Further, the AttnConvLSTM soil moisture prediction model comprises an encoder and a decoder, the encoder is composed of a plurality of AttnConvLSTM units stacked, and the high-correlation meteorological data and the precipitation anomaly are encoded into a context vector containing spatial and temporal features; the decoder adopts the same AttnConvLSTM unit as the encoder, and gradually generates a soil moisture map by combining the context vector containing spatial and temporal features and the precipitation anomaly.

[0016] Furthermore, each AttnConvLSTM unit includes a ConvLSTM and an attention module, wherein the attention module is based on precipitation anomalies and... T The hidden state at time -1 and the computational spatial attention weight matrix; the ConvLSTM is based on highly correlated meteorological data and T At time -1, the hidden state is calculated by considering the input gate, forget gate, output gate, and candidate cell state. The candidate cell state is spatially weighted using a spatial attention weight matrix. The cell state is then updated using the input and forget gates, and finally, the output gate is used to obtain the hidden state. T The hidden state at any given moment.

[0017] Furthermore, the calculation process of the spatial attention weight matrix is ​​as follows:

[0018] Precipitation anomaly P and T Hidden state at time -1 H T-1 By performing concatenation along the channel dimension, we obtain T Multi-channel feature maps at different times Combined T =Concat( H T-1 , P ), where Concat() represents the concatenation operation;

[0019] Combining multi-channel feature maps using 1×1 convolutional layers T After nonlinear transformation and compression, the result is mapped to a preliminary attention score map (Score) for a single channel. T =σ(W p Combined T +b p ), where W p This indicates that the convolutional layer can learn the parameters of the convolutional kernel, b p σ represents the bias of the convolutional kernel that the convolutional layer can learn, and σ represents the Sigmoid function;

[0020] Score of the initial attention graph T Softmax normalization is performed to obtain the spatial attention matrix. =Softmax(Score T ).

[0021] Furthermore, the aforementioned T The process of acquiring the hidden state at a given time is as follows:

[0022] ConvLSTM is based on highly correlated meteorological data and THidden state at time -1 T Input gate, forget gate, output gate and candidate cell state at time t:

[0023]

[0024]

[0025]

[0026]

[0027] wherein, represents T Input gate at time t, σ represents a Sigmoid function, represents T Highly relevant meteorological data input at time t, represents T Hidden state at time -1, represents a weight matrix of the input gate, represents a hidden layer weight matrix of the input gate, represents a bias of the input gate; represents T Forget gate at time t, represents a weight matrix of the forget gate, represents a hidden layer weight matrix of the forget gate, represents a bias of the forget gate; represents T Output gate at time t, represents a weight matrix of the output gate, represents a hidden layer weight matrix of the output gate, represents a bias of the output gate; represents T Candidate cell state at time t, represents a weight matrix of the cell state, represents a hidden layer weight matrix of the cell state, represents a bias of the cell state, represents a hyperbolic tangent function, represents a convolution operation;

[0028] Utilizing a spatial attention weight matrix to spatially weight modulate the candidate cell state combining the input gate and the forget gate to update the cell state :

[0029]

[0030] updating the cell state binding output gate obtaining T hidden state at time step t wherein, denotes element-wise multiplication.

[0031] Further, for the AttnConvLSTM unit in the encoder, if T -1 = 0, then T the hidden state at time step -1 is 0; for the AttnConvLSTM unit in the decoder, if T -1 = 0, then T the hidden state at time step -1 is the context vector containing the spatio-temporal features encoded by the encoder, including: T the cell state and the hidden state of the encoder output at time step t.

[0032] Further, an independent 1x1 convolutional layer is added to the hidden state of each time step at the end of the decoder, and the number of filters of the convolutional layer is set to 1.

[0033] Further, the calculation process of the mean square error loss function is as follows:

[0034]

[0035] wherein, denotes the batch size, denotes the length of the corresponding region of soil moisture, denotes the width of the corresponding region of soil moisture, denotes the index of denotes the index of b denotes the index of H denotes the index of c denotes the label value of the soil moisture in the region range of denotes the predicted value of the soil moisture in the region range of b c b c

[0036] ​​​​​​​​Further, in the initial round of AttnConvLSTM soil moisture condition prediction model training, the label value is directly used as the input of the decoder in the AttnConvLSTM soil moisture condition prediction model to train the AttnConvLSTM soil moisture condition prediction model; in the subsequent training process of the AttnConvLSTM soil moisture condition prediction model, the use probability of the label value is gradually reduced, and the context vector containing the spatio-temporal features coded by the encoder in the AttnConvLSTM soil moisture condition prediction model is increased as the input of the decoder to train the AttnConvLSTM soil moisture condition prediction model;

[0037] The use probability of the label value The calculation process is as follows:

[0038]

[0039] Wherein, Indicates a fixed parameter, 0 ; Indicates the training round.

[0040] Compared with the prior art, the present application has the following beneficial effects:

[0041] (1) The AttnConvLSTM soil moisture condition intelligent prediction method fusing CLDAS of the present application uses the Pearson correlation coefficient and Granger causality test to analyze the relationship between meteorological data and soil moisture condition data, and selects meteorological data with high correlation, so that the AttnConvLSTM soil moisture condition prediction model can more comprehensively capture the meteorological features related to soil moisture condition, and improve the prediction accuracy and reliability of soil moisture condition;

[0042] (2) The AttnConvLSTM soil moisture condition intelligent prediction method fusing CLDAS of the present application introduces the spatial attention of precipitation anomaly into the encoder and decoder of the AttnConvLSTM soil moisture condition prediction model, wherein the encoder encodes the meteorological data with high correlation and the precipitation anomaly into a context vector containing spatio-temporal features, and uses the spatial attention mechanism of the precipitation anomaly to perform feature enhancement processing on the data of each time step, thereby dynamically enhancing the feature expression of the precipitation anomaly area, and providing a solid information foundation for the subsequent soil moisture condition prediction work of the decoder; the decoder gradually generates a soil moisture condition map by combining the context vector containing spatio-temporal features with the precipitation anomaly, and realizes accurate focusing on the key change area of soil moisture condition by deeply integrating the spatial attention mechanism of the precipitation anomaly, which can effectively identify the precipitation anomaly area and modulate the cell state update, and ensure the rationality of the prediction accuracy of soil moisture condition in spatial distribution. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1A flow chart of the AttnConvLSTM soil moisture intelligent prediction method of the application fusing CLDAS;

[0044] Figure 2 A schematic diagram of the AttnConvLSTM soil moisture prediction model in the application;

[0045] Figure 3 A schematic diagram of the AttnConvLSTM unit in the application;

[0046] Figure 4 A comparison schematic diagram of the AttnConvLSTM soil moisture prediction result fusing CLDAS in the application and the actual soil moisture. DETAILED DESCRIPTION

[0047] The technical solutions of the application will be further explained and described below with reference to the accompanying drawings.

[0048] As Figure 1 A flow chart of the AttnConvLSTM soil moisture intelligent prediction method of the application fusing CLDAS, the soil moisture intelligent prediction method comprising the following steps:

[0049] Step S1: Collecting soil moisture data of CLDAS and meteorological observation data of a corresponding region of a meteorological observation station; Specifically, the China Meteorological Administration Land Data Assimilation System (CLDAS) selects a suitable land surface model to simulate spatially and temporally continuous soil moisture and other land surface variables at different depths on spatial grid points, finally uses an assimilation algorithm such as EnKF to integrate soil moisture and satellite microwave brightness temperature and other information into the simulation results of the previous step, and uses the assimilation results as the initial field of the next step of the model input, continuously optimizes the model calculation process, so as to obtain soil moisture data that more truly reflects the actual situation; The meteorological data collected by the meteorological observation station includes: precipitation, air temperature, relative humidity, radiation, sunshine duration, wind speed, cloud amount and water vapor pressure data.

[0050] The collected soil moisture data and meteorological data are subjected to outlier processing and missing value processing, and the soil moisture data and meteorological data are integrated according to time and space dimensions to ensure that the soil moisture data and meteorological data correspond. Since the soil moisture data is in GeoTiff format with spatial resolution of 6.25KM, and the meteorological data is in one-dimensional CSV format, the dimensions of the two data do not match, and the soil moisture data needs to be interpolated to the same spatial resolution by IDW interpolation method to fill in the missing spatial dimension.

[0051] Step S2: Pearson correlation coefficient and Granger causality test are used to analyze the relationship between meteorological data and soil moisture data, and meteorological data with high correlation is screened out, so that the AttnConvLSTM soil moisture prediction model can more comprehensively capture the meteorological features related to soil moisture, and improve the prediction accuracy and reliability of soil moisture; specifically:

[0052] The Pearson correlation coefficient is a statistical measure of the degree of linear correlation between two variables. A two-dimensional array is formed by combining soil moisture data and each meteorological data, and the Pearson correlation coefficient between them is calculated. The Pearson correlation coefficient is between-1 and 1, and the closer the absolute value is to 1, the stronger the correlation.

[0053] Granger causality test is used to determine whether a variable can predict the future changes of another variable, thereby determining the causal relationship between variables. Granger causality test is used to verify whether meteorological data is the Granger cause of soil moisture data changes one by one. The original hypothesis is that a certain meteorological data is not the Granger cause of soil moisture. The F-statistic and p-value are calculated to determine whether to reject the original hypothesis. If the p-value is less than 0.05, the original hypothesis is rejected, indicating that the meteorological data is the Granger cause of soil moisture, i.e. the change of the meteorological data can predict the change of soil moisture to a certain extent.

[0054] In one technical solution of the present application, meteorological data with an absolute value of Pearson correlation coefficient greater than 0.5 and a p-value of Granger causality test less than 0.05 is considered as high-correlation meteorological data, including precipitation, temperature, relative humidity and radiation. Among them, precipitation and relative humidity are positively correlated with soil moisture data, and temperature and radiation are negatively correlated with soil moisture data.

[0055] Step S3: An AttnConvLSTM soil moisture prediction model based on Seq2Seq is constructed, and a convolution structure is used to replace the fully connected layer in the traditional LSTM, so as to capture spatial features while preserving time dependence, realize intelligent prediction of soil moisture, and enhance feature abstraction capability. For example Figure 2 The AttnConvLSTM soil moisture prediction model includes an encoder and a decoder. The encoder is composed of multiple layers of AttnConvLSTM units stacked together, and the high-correlation meteorological data and precipitation anomaly are encoded into context vectors containing spatial and temporal features. This process implicitly learns spatial correlation through convolution operation, while explicitly modeling long-term temporal dependence through LSTM gating mechanism. A spatial attention mechanism based on precipitation anomaly is introduced to each time step TThe data undergoes feature enhancement processing to dynamically strengthen the feature representation of precipitation anomaly areas, providing a solid information foundation for subsequent soil moisture prediction by the decoder. The decoder uses the same AttnConvLSTM unit as the encoder. This symmetrical design ensures the consistency of feature processing capabilities, guaranteeing lossless information transmission and efficient decoding. The context vector containing spatiotemporal features is combined with precipitation anomalies to gradually generate soil moisture maps. Through a spatial attention mechanism that deeply integrates precipitation anomalies, the key areas of soil moisture change are accurately focused, effectively identifying precipitation anomaly areas and modulating cell state updates, ensuring the spatial distribution of soil moisture prediction accuracy.

[0056] like Figure 3 Each AttnConvLSTM unit includes a ConvLSTM and an attention module, which integrates convolutional layers and LSTM layers to form a unique spatiotemporal feature processing architecture. Combined with the spatial attention mechanism of precipitation anomalies, which has a significant impact on soil moisture changes, the accuracy of soil moisture prediction is improved.

[0057] Attention module based on precipitation anomaly and T The hidden state at time -1 and the computational spatial attention weight matrix are used. Precipitation anomaly is the difference between actual precipitation and the long-term historical average precipitation for the same period. It can effectively eliminate the influence of seasonal cycles, highlight whether the current precipitation is excessive or insufficient, and more directly reflect the state of water stress or surplus. By introducing the spatial attention weight matrix of precipitation anomaly, it is possible to dynamically focus on areas of precipitation anomaly; specifically:

[0058] Precipitation anomaly P and T Hidden state at time -1 H T-1 By performing concatenation along the channel dimension, we obtain T Multi-channel feature maps at different times Combined T =Concat( H T-1 , P ), where Concat() represents the concatenation operation;

[0059] Combining multi-channel feature maps using 1×1 convolutional layers T Perform nonlinear transformation and compression for efficient fusion. T Hidden state at time -1 H T-1 and precipitation anomaly P Information from different spatial locations is used to automatically learn the interactions between them, and this is mapped to a single-channel initial attention score map. T =σ(W p Combined T +b p ), where W p This indicates that the convolutional layer can learn the parameters of the convolutional kernel, b p This indicates that the convolutional layer can learn the bias of the convolutional kernel, and σ represents the Sigmoid function, ensuring that the initial attention score is between 0 and 1;

[0060] Score of the initial attention map T Softmax normalization is performed to obtain the spatial attention matrix. =Softmax(Score T ).

[0061] ConvLSTM is based on highly correlated meteorological data and T Hidden state calculation at time -1: Input gate, forget gate, output gate, and candidate cell state:

[0062]

[0063]

[0064]

[0065]

[0066] in, express T The input gate at time t, where σ represents the Sigmoid function. express T Highly correlated meteorological data input at any given time needs to be Z-score standardized. express T The hidden state at time -1 This represents the weight matrix of the input gate. This represents the hidden layer weight matrix of the input gate. Indicates the bias of the input gate; express T The Gate of Oblivion in Time The weight matrix represents the forget gate. This represents the hidden layer weight matrix of the forget gate. Indicates the bias of the forget gate; express T Output gate at any given time, This represents the weight matrix of the output gate. This represents the hidden layer weight matrix of the output gate. Indicates the bias of the output gate; express TThe state of candidate cells at a given time. The weight matrix representing the cell state. The hidden layer weight matrix represents the unit state. Indicates the bias of the cell state. Represents the hyperbolic tangent function. Indicates the convolution operation;

[0067] A gated modulation strategy was adopted, using a spatial attention weight matrix. For candidate cell states Spatial weighted modulation is performed, through T Hidden state at time -1 H T-1 and precipitation anomaly P This is used to calculate the spatial attention weight matrix, achieving a deep fusion of data-driven implicit attention and physical prior explicit attention, forming a dual-focusing mechanism, combined with the input gate. And the Gate of Oblivion Update cell state :

[0068]

[0069] Updated cell state Combined with output gate get T Hidden state at any moment ,in, This indicates element-wise multiplication.

[0070] In one technical solution of this invention, the encoder is composed of multiple stacked AttnConvLSTM units, forming a deep spatiotemporal feature extraction network. As the network depth increases, the number of filters increases layer by layer, enabling shallower networks to learn basic local spatiotemporal features, while deeper networks can capture more abstract and global complex patterns. The convolutional kernel size of each AttnConvLSTM unit is uniformly set to 3×3. To avoid gradient vanishing in deep networks, a fully connected layer is inserted between every two AttnConvLSTM units, with the number of neurons dynamically adjusted according to the feature dimension of the preceding and following layers, ensuring that feature information can be transmitted forward losslessly and efficiently. Simultaneously, after each AttnConvLSTM unit, a max-pooling operation in the temporal dimension is introduced, effectively reducing the data dimensionality and subsequent computational load while retaining the most salient features, thus achieving feature extraction.

[0071] In one technical solution of the present invention, for the AttnConvLSTM unit in the encoder, if T -1=0, then TThe hidden state at the time of -1 is 0; for the AttnConvLSTM unit in the decoder, if T -1=0, then T The hidden state at the time of -1 is the context vector encoded by the encoder and containing the spatial-temporal features, including: T The cell state and the hidden state of the encoder output at the time.

[0072] In one technical solution of the present application, the output of the AttnConvLSTM unit of each layer of the decoder is a hidden state feature map, which needs to be converted into a final prediction value through an output layer. Therefore, an independent 1x1 convolutional layer is added to the hidden state of each time step at the end of the decoder, the number of filters of the convolutional layer is set to 1, to realize the fusion and compression of the channel dimension, and map the multi-channel hidden state feature map to a unique target channel--soil moisture value, so as to realize the pixel-level accurate prediction. For each time step, the decoder outputs a two-dimensional soil moisture prediction map, and after multiple cycles, a complete soil moisture prediction sequence is finally obtained.

[0073] Step S4: Taking the meteorological data with high correlation as the input of the AttnConvLSTM soil moisture prediction model and the corresponding soil moisture data as the label value, combining the spatial attention of the precipitation anomaly, training the AttnConvLSTM soil moisture prediction model until the mean square error loss function converges, and completing the training of the AttnConvLSTM soil moisture prediction model;

[0074] The calculation process of the mean square error loss function in the present application is as follows:

[0075]

[0076] Among them, represents the batch size, represents the length of the corresponding region of the soil moisture, represents the width of the corresponding region of the soil moisture, represents the index of represents the index of b represents the index of H represents the index of c represents the index of represents the index of represents the label value of the soil moisture in the region range of batch b x c represents the prediction value of the soil moisture in the region range of batch x b c ​​

[0077] In one technical solution of the present application, a teacher forcing mechanism is used to accelerate the convergence process of the AttnConvLSTM soil moisture prediction model training and improve the stability of the training: in the initial round of AttnConvLSTM soil moisture prediction model training, the label value is directly used as the input of the decoder in the AttnConvLSTM soil moisture prediction model, and the AttnConvLSTM soil moisture prediction model is trained. This strategy helps the AttnConvLSTM soil moisture prediction model quickly learn the stable mapping relationship between the input sequence and the target sequence, accelerates the initial convergence, and lays a good parameter foundation for subsequent learning; in the subsequent training process of the AttnConvLSTM soil moisture prediction model, the probability of using the label value is gradually reduced, and the context vector containing the spatio-temporal features coded by the encoder in the AttnConvLSTM soil moisture prediction model is used as the input of the decoder to train the AttnConvLSTM soil moisture prediction model. By gradually reducing the dependence on the teacher forcing mechanism during training, it forces the AttnConvLSTM soil moisture prediction model to learn how to process and correct its own previous errors, thereby better simulating the real scene during reasoning, thereby guiding the AttnConvLSTM soil moisture prediction model to smoothly transition from complete dependence on the label value to a reasoning mode that must rely on its own prediction; when the training is about to end, the probability of using the label value is reduced to a very small value, at this time, the decoder uses its own prediction value for the next step calculation in most cases, and its training environment is very close to the final reasoning environment, effectively bridging the conditional distribution gap between training and reasoning, and enhancing the robustness and generalization ability of the AttnConvLSTM soil moisture prediction model.

[0078] The calculation process of the probability of using the label value is as follows:

[0079]

[0080] wherein, is a fixed parameter, 0 ; represents the training round.

[0081] Step S5: Real-time acquisition of meteorological observation station forecast meteorological data and precipitation anomaly, input into the trained AttnConvLSTM soil moisture prediction model, and prediction of the soil moisture of the corresponding area.

[0082] As shown in Table 1, the soil moisture prediction results of the AttnConvLSTM soil moisture intelligent prediction method of the application and the soil moisture prediction results of the ConvLSTM soil moisture intelligent prediction method are compared, it can be seen that the root mean square error of the soil moisture prediction results of the application is significantly lower than the soil moisture prediction results based on ConvLSTM, and the accuracy of the soil moisture prediction of the application in the area with more rainfall and the area with less rainfall is improved compared with the soil moisture prediction results based on ConvLSTM.

[0083] Table 1: Comparison of soil moisture prediction results of AttnConvLSTM soil moisture intelligent prediction method of the application and ConvLSTM soil moisture intelligent prediction method

[0084]

[0085] As Figure 4 The comparison diagram of the AttnConvLSTM soil moisture prediction results of the application and the actual soil moisture is given, it can be seen that the error between the soil moisture prediction value and the actual value in most areas is between-10% and 15%, which shows that the prediction of soil moisture of the application has high precision and robustness.

[0086] In one of the technical solutions of the application, a computer readable storage medium is provided, which stores a computer program, and the computer program enables a computer to execute the AttnConvLSTM soil moisture intelligent prediction method of the application.

[0087] In one of the technical solutions of the application, an electronic device is provided, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the AttnConvLSTM soil moisture intelligent prediction method of the application is realized.

[0088] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium, which can contain or store programs for use by or in conjunction with an instruction execution system, device or apparatus. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or apparatus, or any suitable combination of the above. More specific examples of computer storage media can include one or more wires, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0089] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0090] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall be considered as the protection scope of the present application.

Claims

1. An intelligent prediction method for soil moisture content based on CLDAS-AttnConvLSTM, characterized in that, The method comprises the following steps: Step S1: collecting soil moisture data of CLDAS and meteorological data of a corresponding region observed by a meteorological observation station; Step S2: analyzing the relationship between the meteorological data and the soil moisture data by using a Pearson correlation coefficient and a Granger causality test, and screening meteorological data with high correlation; Step S3: constructing an AttnConvLSTM soil moisture prediction model based on Seq2Seq; Step S4: taking the meteorological data with high correlation as the input of the AttnConvLSTM soil moisture prediction model, taking the corresponding soil moisture data as the label value, combining the spatial attention of the precipitation anomaly, training the AttnConvLSTM soil moisture prediction model until the mean square error loss function converges, and completing the training of the AttnConvLSTM soil moisture prediction model; Step S5: acquiring the predicted meteorological data and the precipitation anomaly of the meteorological observation station in real time, inputting the trained AttnConvLSTM soil moisture prediction model, and predicting the soil moisture of the corresponding region; The AttnConvLSTM soil moisture prediction model comprises an encoder and a decoder, the encoder is composed of a plurality of AttnConvLSTM units stacked in multiple layers, and the high-correlation meteorological data and the precipitation anomaly are encoded into a context vector containing spatial and temporal features; the decoder adopts the same AttnConvLSTM unit as the encoder, and gradually generates a soil moisture map by combining the context vector containing spatial and temporal features and the precipitation anomaly; The AttnConvLSTM unit each comprises a ConvLSTM and an attention module, the attention module is used for calculating a spatial attention weight matrix according to a precipitation anomaly and a hidden state at a time point of -1. T The ConvLSTM is used for calculating an input gate, a forget gate, an output gate and a candidate cell state according to high-correlation meteorological data and a hidden state at a time point of -1. T The spatial attention weight matrix is used for spatially weighting and modulating the candidate cell state, the cell state is updated in combination with the input gate and the forget gate, and a hidden state at a time point of 0 is obtained in combination with the output gate. T The hidden state at the time point of 0. The calculation process of the spatial attention weight matrix is as follows: Precipitation anomaly P And T Hidden state at time -1 H T-1 Concatenate in the channel dimension to obtain T Multi-channel feature map Combined at time t T = Concat( H T-1 , P ), wherein Concat() represents a concatenation operation; Combining multi-channel feature maps using 1×1 convolutional layers T After nonlinear transformation and compression, the result is mapped to a preliminary attention score map (Score) for a single channel. T =σ(W p Combined T +b p ), where W p This indicates that the convolutional layer can learn the parameters of the convolutional kernel, b p σ represents the bias of the convolutional kernel that the convolutional layer can learn, and σ represents the Sigmoid function; Score T Softmax normalization is performed to obtain a spatial attention matrix =Softmax(Score T ). 2.The AttnConvLSTM soil moisture intelligent prediction method of claim 1, wherein, The high-correlation meteorological data screened in step S2 is meteorological data with an absolute value of the Pearson correlation coefficient greater than 0.5 and a p value of the Granger causality test less than 0.05, and the high-correlation meteorological data include precipitation, air temperature, relative humidity and radiation. 3.The AttnConvLSTM soil moisture intelligent prediction method of claim 1, wherein, The T The acquisition process of the hidden state at the moment is as follows: ConvLSTM according to the high correlation of meteorological data and T hidden state at time -1 T input gate, forget gate, output gate, and candidate cell state at time wherein, denotes T denotes the input gate at time t, denotes the Sigmoid function, denotes T denotes the highly correlated weather data input at time t, denotes T denotes the hidden state at time -1, denotes the weight matrix of the input gate, denotes the hidden layer weight matrix of the input gate, denotes the bias of the input gate; denotes T denotes the forget gate at time t, denotes the weight matrix of the forget gate, denotes the hidden layer weight matrix of the forget gate, denotes the bias of the forget gate; denotes T denotes the output gate at time t, denotes the weight matrix of the output gate, denotes the hidden layer weight matrix of the output gate, denotes the bias of the output gate; denotes T denotes the candidate cell state at time t, denotes the weight matrix of the cell state, denotes the hidden layer weight matrix of the cell state, denotes the bias of the cell state, denotes the hyperbolic tangent function, denotes the convolution operation; Utilizing spatial attention weight matrices On candidate cell states Performing spatially weighted modulation in conjunction with input gates And forget gates Updating cell states : updating the cell state combining the output gate obtaining T hidden state at a time step wherein, denotes element-wise multiplication. 4.The method of claim 1, wherein, For the AttnConvLSTM unit in the encoder, if T -1 = 0, then T the hidden state at time -1 is 0; for the AttnConvLSTM unit in the decoder, if T -1 = 0, then T the hidden state at time -1 is the context vector containing the spatio-temporal features encoded by the encoder, including: T the cell state and hidden state of the encoder output at time t. 5.The AttnConvLSTM soil moisture intelligent prediction method of claim 1, wherein, An independent 1×1 convolution layer is added to the end of the decoder for the hidden state of each time step, and the number of filters of the convolution layer is set to 1. 6.The method of Claim 1, wherein, The mean square error loss function The calculation process is: wherein, denotes a batch size, denotes a length of the respective area of soil moisture, denotes a width of the respective area of soil moisture, denotes an index of b an index of H an index of c an index of an index of denotes a batch b x c a label value of soil moisture in the area range of denotes a batch b x c a predicted value of soil moisture in the area range of 7. The AttnConvLSTM soil moisture intelligent prediction method of claim 1, wherein, In the initial round of training of the AttnConvLSTM soil moisture prediction model, the label value is directly used as the input of the decoder in the AttnConvLSTM soil moisture prediction model to train the AttnConvLSTM soil moisture prediction model; in the subsequent training process of the AttnConvLSTM soil moisture prediction model, the use probability of the label value is gradually reduced, and the context vector containing spatial and temporal features encoded by the encoder in the AttnConvLSTM soil moisture prediction model is gradually increased as the input of the decoder to train the AttnConvLSTM soil moisture prediction model; The probability of use of the tag value The calculation process is as follows: wherein, represents a fixed parameter, 0 ; represents a training epoch.

Images