Sudden drought identification method and system based on space-time double-branch fusion model

Abstract

The invention discloses a sudden drought identification method and system based on a space-time double-branch fusion model. The method comprises the steps that meteorological data are acquired and preprocessed; calculating a composite sudden drought index according to the obtained data; performing data dimension reduction on the obtained data through singular value decomposition (SVD); a deep learning model is adopted to construct a time branch model, a graph attention network GAT is adopted to construct a space branch model, and the time branch model and the space branch model are dynamically fused through a cross attention mechanism to construct a space-time double-branch fusion model; according to the method, data dimensions are compressed and model complexity is reduced by fusing multi-source variables and combining singular value decomposition (SVD), meanwhile, geographic neighborhood weights are dynamically learned by adopting Transforme and based on a graph attention network (GAT), dynamic fusion of spatial-temporal characteristics is finally realized through a cross attention mechanism, a sudden drought recognition result is generated, and the method has the advantages of being high in robustness, high in accuracy and high in reliability. The limitation of a traditional method on nonlinear feature capture, space-time modeling splitting and generalization ability is broken through.

Description

Technical Field

[0001] The invention belongs to the technical field of meteorological disaster identification methods and discloses a sudden drought identification method and system based on a spatiotemporal double-branch fusion model. Background Art

[0002] Traditional drought monitoring methods primarily target slow-moving climate drought events, using thresholds on single indicators such as accumulated precipitation and soil moisture to determine drought status. However, the frequent occurrence of sudden droughts in recent years, characterized by their sudden onset, rapid evolution, and multifactorial coupling, presents significant challenges in their application. Existing mainstream methods often rely on single physical indicators or combinations of multiple indicators with fixed weights, making it difficult to dynamically capture the nonlinear synergistic effects of multiple factors such as temperature, precipitation, and evapotranspiration, resulting in insufficient characterization of the triggering mechanisms of sudden droughts. Traditional statistical models and early machine learning methods, limited by their structure, are unable to effectively model the rapid temporal evolution and spatial geographic propagation of sudden droughts. Due to limitations in sequence computation, traditional statistical models have limited ability to capture long-term dependencies, while spatial analysis methods based on fixed neighborhood rules struggle to adaptively learn complex geographic relationships. Furthermore, existing methods are often designed for specific climate regions, and their parameter systems rely on local statistical characteristics of historical data, making them inadequate for adapting to the significant differences in the mechanisms of sudden droughts across my country's diverse climate zones. In particular, in the arid-humid transition zone, traditional threshold-based methods experience a significant increase in misjudgment rates due to the complex coupling of environmental factors. Summary of the Invention

[0003] The purpose of this invention is to address the existing sudden drought identification methods that rely on a single indicator or a combination of multiple factors with fixed weights, making it difficult to dynamically capture the nonlinear synergistic effects of multiple factors such as temperature and precipitation, and unable to effectively model rapid temporal decay and complex spatial propagation patterns. When fusing multi-source data, key features are easily lost or the model complexity is excessively high. By constructing a spatiotemporal dual-branch model to dynamically fuse multi-source data, and combining it with a cross-attention mechanism to adaptively assign spatiotemporal feature weights, this method overcomes the limitations of traditional time series models and spatial analysis methods, achieving high-precision identification of sudden drought events and cross-regional generalization capabilities, and addressing the problems of insufficient nonlinear feature capture, fragmented spatiotemporal modeling, and imbalance between computational efficiency and accuracy in existing technologies.

[0004] To solve the above technical problems, the present invention provides a method for identifying sudden drought based on a spatiotemporal dual-branch fusion model, comprising the following steps:

[0005] Step 1: Obtain daily meteorological data of the study area and preprocess them;

[0006] Step 1.1: Perform spatial and temporal unification of daily meteorological data, unifying data from different sources into daily grid data.

[0007] Step 1.2: Perform a data processing process on the data processed in Step 1.1, including but not limited to missing value imputation and outlier correction;

[0008] Step 1.3: Collect historical sudden drought event data and match the spatio-temporal coordinates with the data processed in Step 1.2;

[0009] The daily-scale meteorological data includes temperature, precipitation, root zone soil moisture, vegetation index, and evapotranspiration data;

[0010] Step 2: Calculate the composite sudden drought index for the data obtained in Step 1;

[0011] Step 2.1: Calculate the standardized temperature index STI using the daily-scale temperature data. The calculation formula for STI is:

[0012] STI = (T i - μ T ) / σ T ;

[0013]

[0014] In the formula, T i represents the temperature value on the i-th day, and N is the number of samples;

[0015] Step 2.2: Calculate the standardized precipitation index SPI using the daily-scale precipitation data. The calculation formula for SPI is:

[0016] Let the precipitation at the initial scale be x, then the probability density function of its gamma distribution is:

[0017]

[0018] In the formula, α3 is the shape parameter; β3 is the scale parameter; Γ is the gamma function;

[0019] Let the length of the precipitation series be n, and the number of terms equal to zero be m. Let Then the cumulative probability calculation formula for a certain time scale is:

[0020] H(x) = q + (1 - q)G(x);

[0021]

[0022] Transform the cumulative probability distribution H(x) into a standard normal distribution to obtain the corresponding SPI value:

[0023] When 0 < H(x) ≤ 0.5;

[0024] Let Then

[0025] When 0.5 < H(x) < 1;

[0026] Let Then

[0027] In the formula, c0 = 2.515517; c1 = 0.802853; c2 = 0.010328; d1 = 1.432788; d2 = 0.189269; d3 = 0.001308.

[0028] Step 2.3: Calculate the standardized root zone soil moisture SMCI using the daily-scale root zone soil moisture data; the calculation formula for SMCI is:

[0029] SMCI = (SMrz i - μ s ) / σ s ;

[0030]

[0031] In the formula, SMrz i is the root zone soil moisture value on the i-th day, μ s is the mean value of the root zone soil moisture, σ s is the standard deviation of the root zone soil moisture, and N is the number of samples.

[0032] Step 2.4: Based on the radial basis function RBF, convert the vegetation index NDVI into the kernel vegetation index KNDVI, and the calculation formula for KNDVI is:

[0033]

[0034] In the formula, NIR is the reflectance in the near-infrared band, Red is the reflectance in the red band, and σ1 is the bandwidth parameter.

[0035] Step 2.5: Calculate the standardized evaporation stress ratio RSES based on the evapotranspiration data, and the calculation formula for RSES is:

[0036]

[0037] In the formula, ESR represents the evaporation stress ratio on the daily scale, R sES is the standardized evaporation stress ratio, ET represents the daily-scale evapotranspiration, PET is the daily-scale potential evapotranspiration, and σ2 is the standard deviation of ESR.

[0038] Step 2.6: Perform normalization processing on the standardized temperature index STI, the standardized precipitation index SPI, the standardized root zone soil moisture SMCI, the kernel vegetation index KNDVI, and the standardized evaporation stress ratio RSES.

[0039]

[0040] Where, X i represents the original data, X′ i represents the normalized data, μ is the mean of the original data, and σ3 is the standard deviation of the original data.

[0041] Step 3: Perform data dimension reduction on the data obtained in step 2 by using singular value decomposition (SVD);

[0042] Step 3.1: Perform singular value decomposition on the normalized data matrix F and obtain the decomposition formula:

[0043] F=UΣV T ;

[0044] Where U is the left singular matrix; Σ is the singular value matrix; V is the right singular matrix;

[0045] Step 3.2: Calculate F T The eigenvalue λ of F i , and sort them in descending order;

[0046] Calculating λ i The corresponding eigenvector u i , forming the matrix U, and standardizing the matrix U to obtain:

[0047] U=XA1X T ;

[0048] Matrix X=(x1,x2…x 5m ) is 5m×5m in size, and the column vector (x1, x2…x 5m ) is F T The eigenvector of F is the left singular vector of matrix F. The size of matrix A1 is 5m×5m;

[0049] Calculate FF T The eigenvector v i , forming a matrix V, and standardizing the matrix V:

[0050] V=YA2Y T ;

[0051] Matrix Y = (y1,y2…y n ) is n×n in size, and the column vector (y1, y2…y n ) is F T The eigenvector of F is the right singular vector of the matrix F. The size of the matrix A2 is n×n;

[0052] Step 3.3: Get the sorted eigenvalue λ according to step 3.2 i , extract the corresponding first k1 singular values ​​σ′i ;

[0053] Solve the singular value matrix Σ, the matrix Σ expression is:

[0054] Σ=diag(σ1,σ2,···,σ i );

[0055]

[0056] Where, σ i is a singular value;

[0057] Step 3.4: Based on the first k1 singular values ​​obtained in step 3.3, reorganize the left singular matrix U, the singular value matrix Σ, and the right singular matrix V to obtain the reduced-dimensional data matrix;

[0058] The singular values ​​are sorted from large to small, the number of principal components is 4, that is, k1=4, the cumulative variance contribution rate P≥85%, and the matrix is ​​reorganized to obtain the data after dimensionality reduction.

[0059] Step 4: Use the deep learning model to build a temporal branch model, use the graph attention network (GAT) to build a spatial branch model, and dynamically fuse the temporal branch model and the spatial branch model through the cross-attention mechanism to build a spatiotemporal dual-branch fusion model.

[0060] Step 4.1: Use a deep learning model to build a time branching model. The input of the time branching model is the data obtained in step 3, and the output is the global features of the time dependency relationship.

[0061] The deep learning model is a transformer model, which uses a transformer deep learning module. The transformer deep learning module adopts an encoding and decoding architecture, and the input data enters the encoding layer through position encoding and input embedding;

[0062] The position coding uses sine and cosine functions to generate a position coding matrix PE. The calculation formula of the position coding matrix PE is:

[0063]

[0064] Where pos is the position index in the sequence, 2i and 2i+1 represent even and odd bits respectively, and d model Represents the total number of dimensions;

[0065] The feature embedding directly inputs the n-dimensional features of each time step into the linear layer through the fully connected layer and maps them to the embedding dimension d model , the output calculation formula of input data through feature embedding is:

[0066] X embed=Linear(X input );

[0067] Where Xinput is the original input data, Linear(·) is the linear layer;

[0068] The final transformer input data is:

[0069] X ′ =X embed +PE;

[0070] In the encoding layer, the global dependency is calculated through the multi-head self-attention mechanism. ′ , the calculation formula for Query, Key, and Value is:

[0071]

[0072]

[0073] Where, is the training parameter, Q tr Query, K tr is Key, V tr is Value;

[0074] The attention weight calculation formula is:

[0075]

[0076] Where h is the number of attention heads, is the scaling factor, d k is the dimension of the Key, d v is the dimension of Value, and SoftMax is the normalization function;

[0077] Using a multi-attention head learning model, the dependency is captured by parallelizing multiple independent attention heads. The calculation formula is:

[0078] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O ;

[0079]

[0080] Where W O is the output projection matrix, Concat(·) is the concatenation of the outputs of h heads along the feature dimension, is the projection matrix corresponding to the multi-head attention Query, is the projection matrix corresponding to the multi-head attention Key, is the projection matrix corresponding to the multi-head attention Value;

[0081] For each time step output, a two-layer feedforward network is used:

[0082] FFN(x)=max(0,xW1+b1)W2+b2;

[0083] Where x is the input vector at a certain time step, W1 is the weight matrix of the first layer, b1 is the bias term of the first layer; W2 is the weight matrix of the second layer, b2 is the bias term of the second layer, and max(0, ) is the ReLU activation function;

[0084] After each sub-layer, layer normalization and residual connection are performed:

[0085]

[0086] output=LayerNorm(x+Sublayer(x));

[0087] Where x represents the input vector; μ is the mean of the input vector x; σ is the standard deviation of the input vector x; γ is the scaling parameter, β is the translation parameter, ∈ is the minimum value, x is the input vector of the sublayer, and Sublayer is the output of the sublayer;

[0088] Construct masked multi-head self-attention;

[0089] Apply an upper triangular mask to the self-attention matrix. The triangular mask calculation formula is:

[0090]

[0091] Add mask when calculating attention weights:

[0092]

[0093] Encoder-decoder multi-head attention mechanism;

[0094] The cross attention between the decoder Query and the encoder Key-Value is calculated as follows:

[0095]

[0096] Where Q dec is the current layer of the decoder, K enc is the key matrix of the final output of the encoder, V enc is the value matrix of the final output of the encoder;

[0097] Furthermore, the transformer model is trained using Adam and the learning rate is 3×10 -4 , first-order moment attenuation rate β1 = 0.9, second-order moment attenuation rate β1 = 0.999, minimum value ∈ = 10 -8 , weight decay weight_decay uses 0.01.

[0098] Furthermore, the learning rate is used for preheating, and the formula for calculating the number of preheating steps is:

[0099] Warmup Steps=max(100,Total Steps×0.03);

[0100] Where Total Steps is the total number of training steps.

[0101] Furthermore, the encoding layer and the decoding layer are stacked with 6 layers, d model =128, hidden layer output size d ff =512, the number of multi-head attention heads is 4, d K =d V =32, dropout=0.1, Batch Size=64, Epoch=150, and the validation loss is terminated early if it does not decrease for 10 to 20 consecutive Epochs.

[0102] Step 4.2: Use the graph attention network (GAT) to build a spatial branch model. The input of the spatial branch model is the data obtained in step 3, and the output is the spatial correlation feature.

[0103] The graph attention network GAT builds a GAT deep learning module;

[0104] The input data features are:

[0105]

[0106] Where N is the number of nodes and d is the number of feature dimensions.

[0107] Perform a linear transformation on the input data features:

[0108]

[0109] Where W (l) is the weight matrix of the lth layer, i is the number of the current node, is the input feature vector of the i-th node in the l-th layer;

[0110] The adjacency matrix is ​​constructed using the eight-neighborhood connection method. The node distance threshold is ≤ 17 kilometers. The expression of the adjacency matrix is:

[0111]

[0112] Where N(i) is the set of neighborhood relations of node i, and j is the number of the neighbor node;

[0113] The LeakyReLU activation function is used to perform nonlinear transformation on spatial information. The calculation formula for the attention coefficient of node i and its adjacent node j in the graph attention layer is as follows:

[0114]

[0115] Where, Indicates concatenating the feature vectors of nodes i and j, α1 T is the training parameter vector, LeakyReLU is the activation function;

[0116] For e ij Perform softmax normalization:

[0117]

[0118] The calculation formula of GAT adjacency feature aggregation is:

[0119]

[0120] Where σ(·) is the ReLU activation function;

[0121] A multi-head attention mechanism is used to repeat the linear transformation and attention calculation k times. Each attention head calculates the output independently. For the middle layer, the calculation formula is as follows:

[0122]

[0123] Where W1 (k) is the weight matrix of the kth attention head;

[0124] For the last layer, the average output is taken:

[0125]

[0126] Where W1 (k) is the weight matrix of the kth attention head;

[0127] Furthermore, the Adam optimizer was used in the training process, with a learning rate of 0.0001, a weight decay of 0.0001, a number of GAT layers Lspace = 2, a number of attention heads per layer hhead = 4, and a hidden layer dimension of

[0128] dhidden=64,dropout=0.5.

[0129] Step 4.3: The spatiotemporal dual-branch fusion model dynamically fuses the output features of the temporal branch model and the spatial branch model through a cross-attention mechanism to generate sudden drought recognition results;

[0130] The cross-attention mechanism dynamically integrates the output features of the temporal branch model and the spatial branch model. The cross-attention mechanism includes a multi-layer stack of spatiotemporal interaction modules, which in turn include:

[0131] 1) Time-to-space (T2S) cross-attention submodule, which is used to map temporal dimension features to spatial dimension and perform feature interaction;

[0132] 2) Space-to-time (S2T) cross-attention submodule, which is used to map spatial dimension features to temporal dimension and perform feature interaction;

[0133] The calculation formula of time-to-space T2S cross attention is:

[0134]

[0135] K time =H time W K1 ;

[0136] V time =H time W V1 ;

[0137] Where H time is the temporal feature output of transformer, H space is the temporal feature output of GAT, W Q1 、W K1 、W V1 is the training parameter, t is the recognition time step;

[0138] The time-to-space cross attention weight is calculated as follows:

[0139]

[0140] The calculation formula of space-to-time cross attention Query (Q), Key (K), and Value (V) is as follows:

[0141] Q time =H time W Q2 ;

[0142]

[0143] Where, H time Represents the temporal feature output of transformer, H spaceRepresents the temporal characteristic output of GAT, W Q2 、W K2 、W V2 are the training parameters;

[0144] The calculation method of cross attention weight from space to time is:

[0145]

[0146] The calculation formula for time feature update is:

[0147] Z time ′=LayerNorm(Z time +Attention S2T );

[0148] Where Z time is the intermediate state matrix of time characteristics

[0149] The calculation formula for spatial feature update is:

[0150]

[0151] Where, is the intermediate state matrix of spatial features

[0152] The calculation formula for residual connection and normalization is:

[0153]

[0154] The output Z of layer l time′(l) 、 As the input of the l+1 layer;

[0155] In the output layer, identify the occurrence sequence of sudden drought at the time node:

[0156]

[0157] Where, MLP is the temporal decoder;

[0158] Identify the occurrence distribution grid of sudden drought at spatial nodes:

[0159]

[0160] In the formula, GNN is the spatial decoder.

[0161] Furthermore, the Adam optimizer is used in the training process, and the initial learning rate η=3×10 -4 , the number of cross attention heads is 4, dropout = 0.3, and weight decay is 0.0001.

[0162] Step 5: Set the classification accuracy and comprehensive classification performance indicators to iteratively optimize the model hyperparameters until the model reaches convergence.

[0163] Furthermore, a sudden drought identification system based on a sudden drought identification method of a spatiotemporal dual-branch fusion model includes:

[0164] Data preprocessing module, which obtains daily-scale meteorological data and performs preprocessing;

[0165] The spatiotemporal dual-branch model module uses a deep learning model to build a temporal branch model and a graph attention network (GAT) to build a spatial branch model.

[0166] The dynamic fusion module dynamically fuses the time branch model and the space branch model through the cross-attention mechanism to build a spatiotemporal dual-branch fusion model.

[0167] Compared with the prior art, the present invention has the following beneficial effects:

[0168] (1) The present invention dynamically captures the nonlinear synergistic mutation effect of multiple factors by fusing multiple source variables such as temperature, precipitation, evapotranspiration, root zone soil moisture and vegetation index, and combines singular value decomposition (SVD) to compress the data dimension and reduce the model complexity. At the same time, the Transformer is adopted to capture the global characteristics of time dependency using the self-attention mechanism, and the geographic neighborhood weights are dynamically learned based on the graph attention network (GAT), and spatial correlation features are adaptively obtained in complex terrain. Finally, the dynamic fusion of spatiotemporal features is achieved through the cross-attention mechanism to generate sudden drought recognition results, breaking through the limitations of traditional methods in nonlinear feature capture, spatiotemporal modeling separation and generalization ability.

[0169] (2) The present invention breaks through the limitations of a single indicator or fixed weight combination by integrating multiple source variables such as temperature, precipitation, evapotranspiration, root zone soil moisture, and vegetation index, dynamically captures the nonlinear synergistic mutation effects of multiple factors, and significantly improves the sensitivity and accuracy of sudden drought identification.

[0170] (3) The present invention recalculates the original data, directly correlates it with the dynamic changes of vegetation, significantly improves the correlation of the actual impact of sudden drought events, and avoids the misjudgment of delayed ecological response; the introduction of the kernel normalized difference vegetation index (kNDVI) breaks through the linear assumption limitation of traditional NDVI, effectively handles the nonlinear relationship between vegetation growth and spectral reflectance, provides more reliable ecological response data support for sudden drought identification, and solves the technical defects of the existing methods such as distortion of vegetation drought representation and weak correlation with ecological impact.

[0171] (4) The present invention adopts a data dimensionality reduction method based on singular value decomposition (SVD). SVD accurately screens out the first several singular values ​​and their corresponding singular vectors that can effectively characterize the characteristics of the original data through the principle of matrix decomposition, and constructs an optimal low-dimensional feature space. It not only achieves efficient compression of complex data structures, but also significantly reduces the loss of key information in the dimensionality reduction process. Compared with the limited requirements of PCA on the covariance matrix, the SVD scheme of the present invention has significantly expanded the scope of application of the algorithm, and can directly process non-square matrices, sparse matrices and non-centralized data, showing better flexibility and robustness when dealing with multi-source heterogeneous data. This improved mathematical processing method enables the data set after dimensionality reduction to retain the core characteristics of the original data while significantly improving the efficiency and classification accuracy of subsequent machine learning model training, providing more reliable technical support for high-dimensional data analysis.

[0172] (5) The present invention introduces the Transformer model to achieve a breakthrough improvement in the characteristics of long-sequence data in the meteorological field. The self-attention mechanism is used to dynamically capture the multidimensional correlation patterns of soil moisture, precipitation and temperature parameters over a long time span, effectively overcoming the gradient vanishing and memory decay problems caused by sequence order calculations in traditional recursive neural networks. Its parallel computing architecture significantly improves the efficiency of model training, while enhancing the ability to deeply analyze the laws of meteorological evolution through global dependency modeling. The present invention uses a multi-head attention mechanism to hierarchically mine the spatiotemporal coupling characteristics of meteorological elements, significantly reducing hardware resource consumption while maintaining prediction accuracy, and providing highly robust algorithmic support for real-time analysis and disaster warning of massive meteorological data.

[0173] (6) The present invention introduces the graph attention network (GAT) algorithm framework to achieve a breakthrough in modeling capabilities for complex environmental prediction tasks such as sudden drought identification. The core technical advantage of GAT lies in the adaptive analysis of the nonlinear correlation strength between meteorological element nodes through a learnable dynamic attention weight mechanism, breaking through the limitations of traditional methods that rely on fixed adjacency matrices or manually defined node relationships. This architecture uses multi-head parallel attention layers to mine the potential spatiotemporal coupling characteristics in meteorological data from multiple dimensions, significantly enhancing the model's collaborative perception of local meteorological mutations and global evolution trends. Through a multi-level attention weight allocation strategy, the system can accurately identify abnormal correlation patterns between key meteorological nodes, thereby improving the accuracy of sudden drought event detection. In addition, the model autonomously optimizes the calculation path during training, significantly reducing hardware resource consumption while maintaining high prediction accuracy, and providing an efficient and robust technical solution for real-time dynamic analysis of complex meteorological systems.

[0174] (7) The present invention dynamically integrates spatiotemporal dual-branch features through a cross-attention mechanism, breaking through the technical bottleneck of insufficient regional generalization capabilities of traditional meteorological models. Through a dynamic weight allocation mechanism, deep coupling of spatiotemporal dual-branch features is achieved, and the multi-scale triggering patterns of sudden drought events in different climate zones are autonomously analyzed. This innovative design not only overcomes the defects of traditional methods that rely on regional customized modeling, but also significantly reduces the computational complexity of cross-domain reasoning through an online feature reconstruction mechanism, providing a more universal intelligent analysis paradigm for monitoring extreme meteorological events in the context of global climate change. BRIEF DESCRIPTION OF THE DRAWINGS

[0175] The present invention will be further described below with reference to the accompanying drawings and examples.

[0176] Figure 1 This is a flowchart of the steps of a sudden drought identification method based on a spatiotemporal dual-branch fusion model according to an embodiment of the present invention.

[0177] Figure 2 It is a flowchart of the steps of reconstructing the principal components of data using the singular value decomposition method according to an embodiment of the present invention.

[0178] Figure 3 2 is a schematic diagram of the Transformer structure of the time modeling branch of an embodiment of the present invention.

[0179] Figure 4 Schematic diagram of GAT calculation weight coefficients and adjacency matrix of the spatial modeling branch of an embodiment of the present invention.

[0180] Figure 5 This is a comparison chart of the identification effect of the 2022 severe drought in the Yangtze River Basin based on the spatiotemporal dual-branch fusion model in an embodiment of the present invention and the single variable identification effect.

[0181] Figure 6 This is a comparison chart of the average duration of sudden drought in the Yangtze River Basin from 1980 to 2023 based on the spatiotemporal dual-branch fusion model of an embodiment of the present invention and other studies. DETAILED DESCRIPTION

[0182] In order to make the purpose and specific implementation steps of the present invention clearer, it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. The following is a detailed introduction to the present invention in conjunction with the accompanying drawings and specific embodiments.

[0183] In 2022, a sudden drought occurred in the Yangtze River Basin. This example selected this event as the verification object because it occurred in a humid climate region and there are sufficient observation records of sudden drought events. It is suitable for testing the sudden drought recognition ability of this technical solution in humid areas.

[0184] like Figure 1 and Figure 2 and Figure 3 and Figure 4 and Figure 5 and Figure 6 As shown, a sudden drought identification method based on a spatiotemporal dual-branch fusion model includes the following steps:

[0185] Step 1: Obtain daily meteorological data of the study area and preprocess them;

[0186] Step 1.1: Perform spatial and temporal unification of daily meteorological data, unifying data from different sources into daily grid data.

[0187] Step 1.2: Perform data processing on the data processed in step 1.1, including but not limited to missing value interpolation and outlier correction;

[0188] Step 1.3: Collect historical sudden drought event data and match the spatiotemporal coordinates with the data processed in step 1.2;

[0189] Select a study area and obtain daily temperature, precipitation, root zone soil moisture, vegetation index, actual and potential evapotranspiration data for the study area. This data needs to be processed, including but not limited to data screening, missing value filling, and outlier processing. The data should be unified to the daily time scale and unified to grid data at the same spatial scale. Historical sudden drought data should be collected and aligned with the study area data.

[0190] Step 2: Calculate the composite sudden drought index based on the data obtained in step 1;

[0191] Step 2.1: Calculate the Standardized Temperature Index (STI) using daily temperature data. The calculation formula for STI is:

[0192] STI=(T i -μ T ) / σ T ;

[0193]

[0194] Where, T i represents the temperature value of the i-th day, and N is the number of samples;

[0195] Step 2.2: Calculate the Standardized Precipitation Index (SPI) using daily precipitation data. The calculation formula for SPI is:

[0196] Assuming the initial scale precipitation is x, the probability density function of its gamma distribution is:

[0197]

[0198] Where, α3 is the shape parameter; β3 is the scale parameter; Γ is the gamma function;

[0199] Let the length of the precipitation series be n, and the number of terms equal to zero be m. Let Then the cumulative probability calculation formula for a certain time scale is:

[0200] H(x) = q + (1 - q)G(x);

[0201]

[0202] By transforming the cumulative probability distribution H(x) into a standard normal distribution, the corresponding SPI value can be obtained:

[0203] When 0 < H(x) ≤ 0.5;

[0204] Let Then

[0205] When 0.5 < H(x) < 1;

[0206] Let Then

[0207] Where, c0 = 2.515517; c1 = 0.802853; c2 = 0.010328; d1 = 1.432788; d2 = 0.189269; d3 = 0.001308.

[0208] Step 2.3: Calculate the standardized root zone soil moisture SMCI using the daily scale root zone soil moisture data; the calculation formula for SMCI is:

[0209] SMCI = (SMrz i - μ s ) / σ s ;

[0210]

[0211] Where, SMrz i is the root zone soil moisture value on the i-th day, μ s is the mean of the root zone soil moisture, σ s is the standard deviation of the root zone soil moisture, and N is the number of samples.

[0212] Step 2.4: Based on the radial basis function RBF, transform the vegetation index NDVI into the kernel vegetation index KNDVI. The calculation formula for KNDVI is:

[0213]

[0214] Where NIR is the reflectivity of the near-infrared band, Red is the reflectivity of the red light band, and σ1 is the bandwidth parameter.

[0215] Step 2.5: Calculate the standardized evaporative stress ratio RSES based on the evapotranspiration data. The calculation formula of RSES is:

[0216]

[0217] Where ESR represents the daily evaporation stress ratio, R SES is the standardized evaporative stress ratio, ET represents daily evapotranspiration, PET is daily potential evapotranspiration, and σ2 is the standard deviation of ESR.

[0218] Step 2.6: Normalize the standardized temperature index (STI), standardized precipitation index (SPI), standardized root zone soil moisture (SMCI), kernel vegetation index (KNDVI), and standardized evaporative stress ratio (RSES).

[0219]

[0220] Where, X i represents the original data, X′ i represents the normalized data, μ is the mean of the original data, and σ3 is the standard deviation of the original data.

[0221] Step 3: Perform data dimension reduction on the data obtained in step 2 by using singular value decomposition (SVD);

[0222] Step 3.1: Perform singular value decomposition on the normalized data matrix F and obtain the decomposition formula:

[0223] F=UΣV T ;

[0224] Where U is the left singular matrix; Σ is the singular value matrix; V is the right singular matrix;

[0225] Step 3.2: Calculate F T The eigenvalue λ of F i , and sort them in descending order;

[0226] Calculating λ i The corresponding eigenvector u i , forming the matrix U, and standardizing the matrix U to obtain:

[0227] U=XA1X T ;

[0228] Matrix X = (x1, x2…x 5m ) is 5m×5m in size, and the column vector (x1, x2…x 5m ) is F TThe eigenvector of F is the left singular vector of matrix F. The size of matrix A1 is 5m×5m;

[0229] Calculate FF T The eigenvector v i , forming a matrix V, and standardizing the matrix V:

[0230] V=YA2Y T ;

[0231] Matrix Y = (y1,y2…y n ) is n×n in size, and the column vector (y1, y2…y n ) is F T The eigenvector of F is the right singular vector of matrix F. The size of matrix A2 is n×n

[0232] Step 3.3: Get the sorted eigenvalue λ according to step 3.2 i , extract the corresponding first k1 singular values ​​σ′ i ;

[0233] Solve the singular value matrix Σ, the matrix Σ expression is:

[0234] Σ=diag(σ1,σ2,···,σ i );

[0235]

[0236] Where, σ i is a singular value;

[0237] Step 3.4: Based on the first k1 singular values ​​obtained in step 3.3, reorganize the left singular matrix U, the singular value matrix Σ, and the right singular matrix V to obtain the reduced-dimensional data matrix.

[0238] The singular values ​​are sorted from large to small, the number of principal components is 4, that is, k1=4, the cumulative variance contribution rate P≥85%, and the matrix is ​​reorganized to obtain the data after dimensionality reduction.

[0239] Step 4: Use the deep learning model to build a temporal branch model, use the graph attention network (GAT) to build a spatial branch model, and dynamically fuse the temporal branch model and the spatial branch model through the cross-attention mechanism to build a spatiotemporal dual-branch fusion model.

[0240] Step 4.1: Use a deep learning model to build a time branching model. The input of the time branching model is the data obtained in step 3, and the output is the global features of the time dependency relationship.

[0241] The deep learning model is a transformer model, which uses a transformer deep learning module. The transformer deep learning module adopts an encoding and decoding architecture. The input data enters the encoding layer through position encoding and input embedding.

[0242] Position coding uses sine and cosine functions to generate the position coding matrix PE. The calculation formula of the position coding matrix PE is:

[0243]

[0244] Where pos is the position index in the sequence, 2i and 2i+1 represent even and odd bits respectively, and d model Represents the total number of dimensions;

[0245] Feature embedding, the n-dimensional features of each time step are directly input into the linear layer through the fully connected layer and mapped to the embedding dimension d model , the output calculation formula of input data through feature embedding is:

[0246] X embed =Linear(X input );

[0247] Where Xinput is the original input data, Linear(·) is the linear layer;

[0248] The final transformer input data is:

[0249] X ′ =X embed +PE;

[0250] In the encoding layer, the global dependency is calculated through the multi-head self-attention mechanism. ′ , the calculation formula for Query, Key, and Value is:

[0251]

[0252] Where, is the training parameter, Q tr Query, K tr is Key, V tr is Value;

[0253] The attention weight calculation formula is:

[0254]

[0255] Where h is the number of attention heads, is the scaling factor, d k is the dimension of the Key, dv is the dimension of Value, and SoftMax is the normalization function;

[0256] Using a multi-attention head learning model, the dependency is captured by parallelizing multiple independent attention heads. The calculation formula is:

[0257] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O ;

[0258]

[0259] Where W O is the output projection matrix, Concat(·) is the concatenation of the outputs of h heads along the feature dimension, is the projection matrix corresponding to the multi-head attention Query, is the projection matrix corresponding to the multi-head attention Key, is the projection matrix corresponding to the multi-head attention Value;

[0260] For each time step output, a two-layer feedforward network is used:

[0261] FFN(x)=max(0,xW1+b1)W2+b2;

[0262] Where x is the input vector at a certain time step, W1 is the weight matrix of the first layer, b1 is the bias term of the first layer; W2 is the weight matrix of the second layer, b2 is the bias term of the second layer, and max(0, ) is the ReLU activation function;

[0263] After each sub-layer, layer normalization and residual connection are performed:

[0264]

[0265] output=LayerNorm(x+Sublayer(x));

[0266] Where x represents the input vector; μ is the mean of the input vector x; σ is the standard deviation of the input vector x; γ is the scaling parameter, β is the translation parameter, ∈ is the minimum value, x is the input vector of the sublayer, and Sublayer is the output of the sublayer;

[0267] Construct masked multi-head self-attention;

[0268] Apply an upper triangular mask to the self-attention matrix. The triangular mask calculation formula is:

[0269]

[0270] Add mask when calculating attention weights:

[0271]

[0272] Encoder-decoder multi-head attention mechanism;

[0273] The cross attention between the decoder Query and the encoder Key-Value is calculated as follows:

[0274]

[0275] Where Q dec is the current layer of the decoder, K enc is the key matrix of the final output of the encoder, V enc is the value matrix of the final output of the encoder;

[0276] The transformer model is trained using Adam and the learning rate is 3×10 -4 , first-order moment attenuation rate β1 = 0.9, second-order moment attenuation rate β1 = 0.999, minimum value ∈ = 10 -8 , weight decay weight_decay uses 0.01.

[0277] Using learning rate preheating, the preheating step calculation formula is:

[0278] Warmup Steps=max(100,Total Steps×0.03);

[0279] Where Total Steps is the total number of training steps.

[0280] The encoding layer and decoding layer are stacked with 6 layers, d model =128, hidden layer output size d ff =512, the number of multi-head attention heads is 4, d K =d V =32, dropout=0.1, BatchSize=64, Epoch=150, the validation loss does not decrease for 10 to 20 consecutive Epochs and is terminated early.

[0281] Step 4.2: Use the graph attention network (GAT) to build a spatial branch model. The input of the spatial branch model is the data obtained in step 3, and the output is the spatial correlation feature.

[0282] Graph Attention Network (GAT), building a GAT deep learning module;

[0283] The input data characteristics are:

[0284]

[0285] Where N is the number of nodes and d is the number of feature dimensions.

[0286] Perform a linear transformation on the input data features:

[0287]

[0288] Where W (l) is the weight matrix of the lth layer, i is the number of the current node, is the input feature vector of the i-th node in the l-th layer;

[0289] The adjacency matrix is ​​constructed using the eight-neighborhood connection method. The node distance threshold is ≤ 17 kilometers. The expression of the adjacency matrix is:

[0290]

[0291] Where N(i) is the set of neighborhood relations of node i, and j is the number of the neighbor node;

[0292] The LeakyReLU activation function is used to perform nonlinear transformation on spatial information. The calculation formula for the attention coefficient of node i and its adjacent node j in the graph attention layer is as follows:

[0293]

[0294] Where, Indicates concatenating the feature vectors of nodes i and j, α1 T is the training parameter vector, LeakyReLU is the activation function;

[0295] For e ij Perform softmax normalization:

[0296]

[0297] The calculation formula of GAT adjacency feature aggregation is:

[0298]

[0299] Where σ(·) is the ReLU activation function;

[0300] A multi-head attention mechanism is used to repeat the linear transformation and attention calculation k times. Each attention head calculates the output independently. For the middle layer, the calculation formula is as follows:

[0301]

[0302] Where W1 (k) is the weight matrix of the kth attention head;

[0303] For the last layer, the average output is taken:

[0304]

[0305] Where W1 (k) is the weight matrix of the kth attention head;

[0306] The Adam optimizer was used during training, with a learning rate of 0.0001, weight decay of 0.0001, the number of GAT layers Lspace = 2, the number of attention heads per layer hhead = 4, the hidden layer dimension dhidden = 64, and dropout = 0.5.

[0307] Step 4.3: The spatiotemporal dual-branch fusion model dynamically fuses the output features of the temporal branch model and the spatial branch model through a cross-attention mechanism to generate sudden drought recognition results;

[0308] The cross-attention mechanism dynamically integrates the output features of the temporal branch model and the spatial branch model. The cross-attention mechanism includes a multi-layer stack of spatiotemporal interaction modules, which include:

[0309] 1) Time-to-space (T2S) cross-attention submodule, which is used to map temporal dimension features to spatial dimension and perform feature interaction;

[0310] 2) Space-to-time (S2T) cross-attention submodule, which is used to map spatial dimension features to temporal dimension and perform feature interaction;

[0311] The calculation formula of time-to-space T2S cross attention is:

[0312]

[0313] K time =H time W K1 ;

[0314] V time =H time W V1 ;

[0315] Where H time is the temporal feature output of transformer, H space is the temporal feature output of GAT, W Q1 、W K1 、W V1 is the training parameter, t is the recognition time step;

[0316] The time-to-space cross attention weight is calculated as follows:

[0317]

[0318] The calculation formula of space-to-time cross attention Query (Q), Key (K), and Value (V) is as follows:

[0319] Q time =H time W Q2 ;

[0320]

[0321] Where, H time Represents the temporal feature output of transformer, H space Represents the temporal characteristic output of GAT, W Q2 、W K2 、W V2 are the training parameters;

[0322] The calculation method of cross attention weight from space to time is:

[0323]

[0324] The calculation formula for time feature update is:

[0325] Z time′ =LayerNorm(Z time +Attention S2T );

[0326] Where Z time is the intermediate state matrix of time characteristics

[0327] The calculation formula for spatial feature update is:

[0328]

[0329] Where, is the intermediate state matrix of spatial features

[0330] The calculation formula for residual connection and normalization is:

[0331]

[0332] The output Z of layer l time′(l) 、 As the input of the l+1 layer;

[0333] In the output layer, identify the occurrence sequence of sudden drought at the time node:

[0334]

[0335] Where, MLP is the temporal decoder;

[0336] Identify the occurrence distribution grid of sudden drought at spatial nodes:

[0337]

[0338] In the formula, GNN is the spatial decoder.

[0339] The Adam optimizer is used in the training process, and the initial learning rate η=3×10 -4 , the number of cross attention heads is 4, dropout = 0.3, and weight decay is 0.0001.

[0340] Step 5: Set the model's prediction accuracy, F1 score, ROC-AUC, and recall rate and iteratively optimize the model hyperparameters until the model converges.

[0341] The model's prediction accuracy, F1 score, ROC-AUC, and recall rate are used as the final evaluation indicators.

[0342] The sudden drought identification system based on the sudden drought identification method of the spatiotemporal dual-branch fusion model includes:

[0343] Data preprocessing module, which obtains daily-scale meteorological data and performs preprocessing;

[0344] The spatiotemporal dual-branch model module uses a deep learning model to build a temporal branch model and a graph attention network (GAT) to build a spatial branch model.

[0345] The dynamic fusion module dynamically fuses the time branch model and the space branch model through the cross-attention mechanism to build a spatiotemporal dual-branch fusion model.

[0346] The system captured a sudden change in precipitation in early June, providing a 10-day early warning compared to traditional methods. It also identified the drought's propagation path from the Ganjiang River basin around Poyang Lake to the main stream. Attention weighting revealed that high temperatures in July contributed 65% to evapotranspiration, while insufficient precipitation in August dominated, accounting for 80%. The recall rate was 92%, compared to 78% for traditional methods, and the underreporting rate was reduced by 14%. The ROC-AUC was 0.94, compared to 0.85 for traditional methods, demonstrating significant improvement in spatial consistency. Therefore, the present invention is able to accurately identify sudden drought events.

[0347] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the claims and description of the present invention.

Claims

1. A sudden drought identification method based on a spatiotemporal dual-branch fusion model, characterized in that: The following steps are involved: Step 1: Obtain meteorological data and preprocess it; Step 2: Calculate the composite sudden drought index based on the data obtained in step 1; Step 3: Perform data dimension reduction on the data obtained in step 2 by using singular value decomposition (SVD); Step 4: Use the deep learning model to build a temporal branch model, use the graph attention network (GAT) to build a spatial branch model, and dynamically fuse the temporal branch model and the spatial branch model through the cross-attention mechanism to build a spatiotemporal dual-branch fusion model. Step 4.1: Use a deep learning model to build a time branching model. The input of the time branching model is the data obtained in step 3, and the output is the global features of the time dependency relationship. Step 4.2: Use the graph attention network (GAT) to build a spatial branch model. The input of the spatial branch model is the data obtained in step 3, and the output is the spatial correlation feature. Step 4.3: The spatiotemporal dual-branch fusion model dynamically fuses the output features of the temporal branch model and the spatial branch model through a cross-attention mechanism to generate sudden drought recognition results.

2. The method for sudden drought identification based on a spatiotemporal dual-branch fusion model according to claim 1, characterized in that: The meteorological data include temperature, precipitation, root zone soil moisture, vegetation index and evapotranspiration data, and the time scale of the meteorological data is daily.

3. The method for sudden drought identification based on a spatiotemporal dual-branch fusion model according to claim 1, characterized in that: The step 1 includes the following sub-steps: Step 1.1: Select the study area and obtain the meteorological data of the study area; Step 1.2: Perform temporal and spatial scale unification on the meteorological data, unifying the meteorological data into daily grid data; Step 1.3: Perform data processing on the data processed in step 1.2, including missing value interpolation and outlier correction; Step 1.4: Collect historical sudden drought event data and match the spatiotemporal coordinates with the data processed in step 1.

3.

4. The method for sudden drought identification based on a spatiotemporal dual-branch fusion model according to claim 1, characterized in that: The step 2 includes the following sub-steps: Step 2.1: Calculate the standardized temperature index (STI) based on the temperature data; Step 2.2: Calculate the Standardized Precipitation Index (SPI) based on precipitation data; Step 2.3: Calculate the standardized root zone soil moisture (SMCI) based on the root zone soil moisture data; Step 2.4: The vegetation index NDVI is converted into the kernel vegetation index KNDVI through the radial basis function RBF; Step 2.5: Calculate the standardized evaporative stress ratio (RSES) based on the evapotranspiration data; Step 2.6: Normalize the standardized temperature index (STI), standardized precipitation index (SPI), standardized root zone soil moisture (SMCI), kernel vegetation index (KNDVI), and standardized evaporative stress ratio (RSES).

5. The method for sudden drought identification based on a spatiotemporal dual-branch fusion model according to claim 4, characterized in that: In step 2.4, the vegetation index NDVI is converted into the kernel vegetation index KNDVI through the radial basis function RBF. The calculation formula of KNDVI is: Where NIR is the reflectivity of the near-infrared band, Red is the reflectivity of the red light band, and σ1 is the bandwidth parameter.

6. The method for sudden drought identification based on a spatiotemporal dual-branch fusion model according to claim 1, characterized in that: Described step 3 comprises the following sub-steps: Step 3.1: Perform singular value decomposition on the normalized data matrix F and obtain the decomposition formula: F=UΣV T ; Where U is the left singular matrix; Σ is the singular value matrix; V is the right singular matrix; Step 3.2: Calculate F T The eigenvalue λ of F i , and sort them in descending order; Step 3.3: Get the sorted eigenvalue λ according to step 3.2 i , extract the corresponding first k1 singular values ​​σ′ i ; Step 3.4: Based on the first k singular values ​​obtained in step 3.3, reorganize the left singular matrix U, the singular value matrix Σ, and the right singular matrix V to obtain the reduced-dimensional data matrix.

7. The method for sudden drought identification based on a spatiotemporal dual-branch fusion model according to claim 1, characterized in that: In step 4.1, the deep learning model is a transformer model, using a transformer deep learning module, which uses an encoding and decoding architecture. The input data enters the encoding layer through position encoding and input embedding; The position coding uses sine and cosine functions to generate a position coding matrix PE. The calculation formula of the position coding matrix PE is: Where pos is the position index in the sequence, 2i and 2i+1 represent even and odd bits respectively, and d model Represents the total number of dimensions; The feature embedding directly inputs the n-dimensional features of each time step into the linear layer through the fully connected layer and maps them to the embedding dimension d model , the output calculation formula of input data through feature embedding is: X embed =Linear(X input ); Where Xinput is the original input data, Linear(·) is the linear layer; The final transformer input data is: X′=X embed +PE; In the encoding layer, the global dependency is calculated through the multi-head self-attention mechanism. For the input X′, the calculation formula of Query, Key, and Value is: Where, is the training parameter, Q tr Query, K tr is Key, V tr is Value; The attention weight calculation formula is: Where h is the number of attention heads, is the scaling factor, d k is the dimension of the Key, d v is the dimension of Value, and SoftMax is the normalization function; Using a multi-attention head learning model, the dependency is captured by parallelizing multiple independent attention heads. The calculation formula is: MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O ; Where W O is the output projection matrix, Concat(·) is the concatenation of the outputs of h heads along the feature dimension, is the projection matrix corresponding to the multi-head attention Query, is the projection matrix corresponding to the multi-head attention Key, is the projection matrix corresponding to the multi-head attention Value; For each time step output, a two-layer feedforward network is used: FFN(x)=max(0,xW1+b1)W2+b2; Where x is the input vector at a certain time step, W1 is the weight matrix of the first layer, b1 is the bias term of the first layer; W2 is the weight matrix of the second layer, b2 is the bias term of the second layer, and max(0, ) is the ReLU activation function; After each sub-layer, layer normalization and residual connection are performed: output=LayerNorm(x+Sublayer(x)); Where x represents the input vector; μ is the mean of the input vector x; σ is the standard deviation of the input vector x; γ is the scaling parameter, β is the translation parameter, ∈ is the minimum value, x is the input vector of the sublayer, and Sublayer is the output of the sublayer; Construct masked multi-head self-attention; Apply an upper triangular mask to the self-attention matrix. The triangular mask calculation formula is: Add mask when calculating attention weights: Encoder-decoder multi-head attention mechanism; The cross attention between the decoder Query and the encoder Key-Value is calculated as follows: Where Q dec is the current layer of the decoder, K enc is the key matrix of the final output of the encoder, V enc is the value matrix of the final output of the encoder; Using learning rate preheating, the preheating step calculation formula is: Warmup Steps=max(100,TotalSteps×0.03); Where Total Steps is the total number of training steps.

8. The method for sudden drought identification based on a spatiotemporal dual-branch fusion model according to claim 1, characterized in that: In step 4.2, the graph attention network GAT is used to construct a GAT deep learning module; The input data features are: Where N is the number of nodes and d is the number of feature dimensions; Perform a linear transformation on the input data features: Where W (l) is the weight matrix of the lth layer, i is the number of the current node, is the input feature vector of the i-th node in the l-th layer; The adjacency matrix is ​​constructed by eight-neighborhood connection. The expression of the adjacency matrix is: Where N(i) is the set of neighborhood relations of node i, and j is the number of the neighbor node; The LeakyReLU activation function is used to perform nonlinear transformation on spatial information. The calculation formula for the attention coefficient of node i and its adjacent node j in the graph attention layer is as follows: Where, To concatenate the feature vectors of nodes i and j, α1 T is the training parameter vector, LeakyReLU is the activation function; For e ij Perform softmax normalization: The calculation formula of GAT adjacency feature aggregation is: Where σ(·) is the ReLU activation function; A multi-head attention mechanism is used to repeat the linear transformation and attention calculation k times. Each attention head calculates the output independently. For the middle layer, the calculation formula is as follows: Where W1 (k) is the weight matrix of the kth attention head; For the last layer, the average output is taken: Where W1 (k) is the weight matrix of the kth attention head.

9. The method for sudden drought identification based on a spatiotemporal dual-branch fusion model according to claim 1, characterized in that: In step 4.3, the cross-attention mechanism dynamically fuses the output features of the temporal branch model and the spatial branch model. The cross-attention mechanism includes a multi-layer stacked spatiotemporal interaction module, which includes: 1) Time-to-space (T2S) cross-attention submodule, which is used to map temporal dimension features to spatial dimension and perform feature interaction; 2) Space-to-time (S2T) cross-attention submodule, which is used to map spatial dimension features to temporal dimension and perform feature interaction; The calculation formula of time-to-space T2S cross attention is: K time =H time W K1 ; V time =H time W V1 ; Where H time is the temporal feature output of transformer, H space is the temporal feature output of GAT, W Q1 、W K1 、W V1 is the training parameter, t is the recognition time step; The time-to-space cross attention weight is calculated as follows: The calculation formula of space-to-time cross attention Query (Q), Key (K), and Value (V) is as follows: Q time =H time W Q2 ; Where, H time Represents the temporal feature output of transformer, H space Represents the temporal characteristic output of GAT, W Q2 、W K2 、W V2 are the training parameters; The calculation method of cross attention weight from space to time is: The calculation formula for time feature update is: Z time′ =LayerNorm(Z time +Attention S2T ); Where Z time is the intermediate state matrix of time characteristics; The calculation formula for spatial feature update is: Where, is the intermediate state matrix of spatial features; The calculation formula for residual connection and normalization is: The output Z of layer l time′(l) 、 As the input of the l+1 layer; In the output layer, identify the occurrence sequence of sudden drought at the time node: Where, MLP is the temporal decoder; Identify the occurrence distribution grid of sudden drought at spatial nodes: In the formula, GNN is the spatial decoder.

10. The sudden drought identification system according to any one of claims 1 to 9, characterized in that: Includes the following modules: Data preprocessing module, which obtains meteorological data and performs preprocessing; The spatiotemporal dual-branch model module uses a deep learning model to build a temporal branch model and a graph attention network (GAT) to build a spatial branch model. The dynamic fusion module dynamically fuses the time branch model and the space branch model through the cross-attention mechanism to build a spatiotemporal dual-branch fusion model.

Images