Agricultural drought prediction method and system based on spatio-temporal feature learning

Abstract

This invention provides an agricultural drought prediction method and system integrating spatiotemporal feature learning, comprising: data collection and preprocessing; constructing a drought prediction model, including a spatial feature extraction module, a spatiotemporal feature learning module, and a sequence prediction module; constructing a drought event identification and prediction model; the drought event identification and prediction model analyzes the standardized soil moisture index (SMI) sequence for the next M consecutive days predicted by the sequence prediction module, extracts drought events, analyzes the spatial centroid movement characteristics of severe drought events, and predicts the future spatiotemporal changes of drought. This invention comprehensively considers the temporal and spatial attributes of drought, utilizes high-precision raster meteorological data, and inputs multiple hydrometeorological elements into a pre-trained convolutional neural network coupled with a deep learning model to predict future drought indicators. While maintaining model complexity within an appropriate range, it efficiently provides early warning of drought, offering technical guidance for drought prevention and water resource management.

Description

Technical Field

[0001] This invention belongs to the field of drought prediction technology, specifically an agricultural drought prediction method and system that integrates spatiotemporal feature learning. Background Technology

[0002] To address the challenges of drought prediction, numerous scholars have proposed various methods. Early drought prediction primarily relied on dynamic forecasting based on hydrological model simulations, which involved complex calculations and introduced significant uncertainties from the hydrological models themselves. In recent years, with the continuous development of machine learning, researchers have discovered its potential in drought prediction, making deep learning-based drought prediction a research hotspot. However, existing drought prediction methods typically focus only on the temporal variations of regional drought, neglecting the spatial development process, thus making it difficult to comprehensively and accurately predict agricultural drought. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an agricultural drought prediction method and system that integrates spatiotemporal feature learning, which can effectively solve the above-mentioned problems.

[0004] The technical solution adopted in this invention is as follows:

[0005] This invention provides an agricultural drought prediction method that integrates spatiotemporal feature learning, comprising the following steps:

[0006] Step S1, Data Collection and Preprocessing:

[0007] Step S1.1, the prediction region is a rasterized region of H*W; where H and W represent the number of grid cells in the height and width directions of the prediction region, respectively;

[0008] Step S1.2: Obtain the time series of meteorological and hydrological elements for the prediction area over the most recent N days, and perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, represented as: set X = (X1, X2, ..., X...). N ); where X1, X2, ..., X N , representing the standardized indices of meteorological and hydrological elements for the forecast area arranged chronologically on historical day 1, historical day 2, ..., historical day N; for X k , k = 1, 2, ..., N, where X is the standardized index of meteorological and hydrological elements in the prediction area on historical day k; k The spatial dimension contains H*W raster data, and each raster data has C=5 dimensional features, namely: the standardized maximum temperature index ST. max Standardized minimum temperature index ST min Standardized precipitation index (SPI), standardized runoff index (SRI), and standardized soil moisture index (SMI); C represents the dimension of the raster data.

[0009] Step S2: Construct a drought prediction model; the drought prediction model includes a spatial feature extraction module, a spatiotemporal feature learning module, and a sequence prediction module;

[0010] Step S2.1, set X = (X1, X2, ..., X...) N The data is input to the spatial feature extraction module, which then sequentially processes the standardized index X of meteorological and hydrological elements for each historical day k. k Spatial feature extraction is performed to extract the spatial feature vector SF corresponding to each historical day k. k This forms the spatial feature matrix SF = (SF1, SF2, ..., SF2). N );

[0011] Step S2.2, convert the spatial feature matrix SF = (SF1, SF2, ..., SF2) into a single matrix. N The input is fed into the spatiotemporal feature learning module, which sequentially processes the spatial feature vector SF of each historical day k. k Spatiotemporal feature learning is performed to obtain the spatiotemporal feature vector (TSF) corresponding to each historical day k. k This forms the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF). N );

[0012] Step S2.3, convert the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF) N The input is given to the sequence prediction module, which predicts the standardized soil moisture index (SMI) sequence for the next M consecutive days, expressed as: SMI = (SMI... N+1 SMI N+2 ,...,SMI N+M ); where M < N, SMI N+m The standardized soil moisture index SMI is the predicted standard soil moisture index for the m-th future day, where m = 1, 2, ..., M, and has H*W grids corresponding to the standardized soil moisture index SMI in the spatial dimension.

[0013] Step S3: Construct a drought event identification and prediction model; the drought event identification and prediction model analyzes the standardized soil moisture index (SMI) sequence for the next M consecutive days predicted by the sequence prediction module, extracts drought events, analyzes the spatial centroid movement characteristics of severe drought events, and predicts the future spatiotemporal changes of drought.

[0014] Preferably, the procedure before step S1.2 further includes:

[0015] Collect ERA5-Land data products and hydrological station data for the prediction area, and perform spatial downscaling to obtain the time series of meteorological and hydrological elements for the prediction area over the most recent N days.

[0016] Preferably, in step S1.2, standardization is performed using formula (1):

[0017]

[0018] in, This represents the standardized index of the c-th meteorological and hydrological element in the h-th row and w-th column of the raster in the prediction area on historical day k; where h = 1, 2, ..., H; w = 1, 2, ..., W; c = 1, 2, ..., C;

[0019] μ represents the average value of the standardized index of the c-th meteorological and hydrological element over N historical days, and σ represents the standard deviation of the standardized index of the c-th meteorological and hydrological element over N historical days.

[0020] represent Values ​​before standardization.

[0021] Preferably, step S2.1 specifically includes:

[0022] Step S2.1.1, the spatial feature extraction module includes a Conv2D convolutional layer, a BatchNormalization normalization layer, and a PReLu activation layer;

[0023] Step S2.1.2, Standardized index X of meteorological and hydrological elements for historical day k k A data matrix of size H*W*C;

[0024] Standardized index of meteorological and hydrological elements of historical day k X k The input is fed into a Conv2D convolutional layer configured with 3×3 kernels. Local features are captured on the input data matrix by sliding the convolution kernel. Spatial feature vectors with spatial correlation are extracted using formula (2).

[0025]

[0026] Among them, K i,j,c The weights represent the convolution kernel;

[0027] Standardized index X of meteorological and hydrological elements representing historical day k k In the diagram, the value of the standardized index of the c-th meteorological and hydrological element in the raster of the (h+i)-th row and (w+j)-th column; c = 1, 2, ..., C;

[0028] This represents the spatial feature vector of the standardized index of the c-th meteorological and hydrological element in the h-th row and w-th column of the raster in the prediction area on historical day k.

[0029] Step S2.1.3: Superimpose the spatial feature vectors of the standardized indices of the C meteorological and hydrological elements in the h-th row and w-th column of the raster for historical day k in the prediction area to obtain the spatial feature vector of the h-th row and w-th column of the raster for historical day k in the prediction area.

[0030] Step S2.1.4, the BatchNormalization normalization layer, for each h-th row and w-th column of the prediction region of historical day k, the spatial feature vectors... After normalization, the vector is activated by a PReLU activation layer to obtain the spatial feature vector SF of historical day k. k .

[0031] Preferably, step S2.2 specifically includes:

[0032] Step S2.2.1, the spatiotemporal feature learning module includes C layers of ConvLSTM2D convolutional layers; for the c-th ConvLSTM2D convolutional layer, c = 1, 2, ..., C, the convolution operation is performed using formula (3):

[0033] output(TSF k,c ) = f(conv(Input(SF) k,c ),W conv,c ,b c (3)

[0034] Where: Input(SF) k,c ) represents the spatial feature vector of the standardized index of the c-th meteorological and hydrological element in the historical day k of the input prediction region; f represents the activation function; conv represents the convolution operation; W conv,c b represents the convolutional kernel of layer c; c It is the bias term of the c-th layer;

[0035] output(TSF k,c ) is the spatiotemporal feature vector of the standardized index of the c-th meteorological and hydrological element in the historical day k of the prediction region;

[0036] Step S2.2.2 yields the spatiotemporal feature vectors of the standardized meteorological and hydrological element indices for C prediction regions on historical day k, namely: TSF k,1 ,TSF k,2 ,...,TSF k,C ; For TSF k,1 ,TSF k,2 ,...,TSFk,C By superimposing the vectors, we obtain the spatiotemporal feature vector TSF corresponding to historical day k. k .

[0037] Preferably, step S2.3 specifically includes:

[0038] Step S2.3.1, the sequence prediction module includes M ConvLSTM2D convolutional layers and M Conv2D convolutional layers;

[0039] Step S2.3.2, convert the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF) N The input is fed into the first ConvLSTM2D convolutional layer ConvLSTM2D1, and the spatiotemporal feature vector TSF of day N+1 is obtained by using formula (4). N+1 :

[0040] output(TSF N+1 )=ConvLSTM2D1(Input(TSF1,TSF2,...,TSF N ),hide1,cell1) (4)

[0041] Where: hide1 and cell1 represent the hidden state and cell state of the first ConvLSTM2D convolutional layer ConvLSTM2D1, respectively;

[0042] The spatiotemporal feature vector TSF of day N+1 N+1 The input is fed into the first Conv2D convolutional layer, and the output is the standardized soil moisture index (SMI) for day N+1. N+1 ;

[0043] Step S2.3.3: Convert the spatiotemporal feature vector matrix TSF2,...,TSF N ,TSF N+1 The input is fed into the second ConvLSTM2D convolutional layer, ConvLSTM2D2, and Equation (5) is used to obtain the spatiotemporal feature vector TSF for the (N+2)th day. N+2 :

[0044] output(TSF N+2 )=ConvLSTM2D2(Input(TSF2,...,TSF N ,TSF N+1 ),hide2,cell2)(5)

[0045] Where: hide2 and cell2 represent the hidden state and cell state of the second ConvLSTM2D convolutional layer, respectively;

[0046] The spatiotemporal feature vector TSF of day N+2 N+2 The input is fed into the second Conv2D convolutional layer, and the output is the standardized soil moisture index (SMI) for day N+2. N+2 ;

[0047] Step S2.3.4, continue shifting one position to the right in this cycle to obtain the standardized soil moisture index (SMI) sequence for the next M consecutive days: SMI = (SMI... N+1 SMI N+2 ,...,SMI N+M ).

[0048] Preferably, step S3 specifically includes:

[0049] Step S3.1: For each grid cell in the prediction area, obtain the standardized soil moisture index (SMI) sequence for the historical N days and the next M consecutive days. Therefore, N+M standardized soil moisture indices (SMI) are obtained, represented as: Sequence SMI N+M =(SMI1,SMI2,...,SMI N SMI N+1 SMI N+2 ,...,SMI N+M );

[0050] Step S3.2, for sequence SMI N+M Perform trend analysis on the time-varying characteristics to determine whether the trend is significant. If so, then the series is SMI (Search Engine Management). N+M Assign the value to the corresponding grid cell; otherwise, assign the value 0 to the grid cell.

[0051] Step S3.3: Perform daily analysis of the distribution of the standardized soil moisture index (SMI) of the predicted area for the most recent M days and the next M consecutive days to determine the centroid of the severely drought area each day, thereby obtaining the movement characteristics of the centroid of the severely drought area over the most recent M days and the next M consecutive days.

[0052] Preferably, step S3.2 specifically includes:

[0053] Step S3.2.1: Using formula (6), the sequence SMI is calculated. N+M =(SMI1,SMI2,...,SMI N SMI N+1 SMI N+2 ,...,SMI N+M The cumulative difference d:

[0054]

[0055] Where: q = 1, 2, ..., N+M-1; p = q+1, q+2, ..., N+M; d represents the sequence SMI. N+M The cumulative difference between the various SMI data points, where d > 0 indicates that the SMI data sequence tends to increase, and vice versa; sign() represents the sign function;

[0056] Step S3.2.2, using formula (7), the concentration Var(d) of the cumulative difference d is obtained:

[0057]

[0058] in:

[0059] R represents SMI1, SMI2, ..., SMI N SMI N+1 SMI N+2 ,...,SMI N+M In the given information, there are R distinct SMIs that repeat at least twice; for the r-th repeating SMI, its number of repetitions is Num. r ;

[0060] Step S3.2.3, using formula (8), obtain the statistic St:

[0061]

[0062] Step S3.2.4: If the statistic St exceeds the set threshold, it indicates that the sequence SMI has occurred. N+M It has a significant trend; otherwise, it represents the SMI sequence. N+M It does not show a significant trend.

[0063] Preferably, in step S3.3, the method for determining the centroid of the severely drought-stricken area each day is as follows:

[0064] Step S3.3.1: Take each day in the most recent historical M days and the next consecutive M days as the identification day. On each identification day, read the standardized soil moisture index SMI of each grid in the prediction area on the identification day.

[0065] Step S3.3.2: Identify whether a drought event occurred on the identification day for each grid cell:

[0066] The standardized soil moisture index (SMI) of each grid cell on the identification day is compared with 0. If the SMI is less than 0, it has negative variation, indicating that the soil moisture is low and a drought event has occurred on the identification day. The SMI value is then used as the drought intensity of the grid cell on the identification day. Conversely, if the SMI is greater than or equal to 0, it has positive variation, indicating that the soil moisture is high and no drought event has occurred. The grid cell does not have drought intensity on the identification day.

[0067] Step S3.3.3: Determine the severely drought-stricken areas in the prediction region on each identification day:

[0068] The prediction area is traversed in a 5×5 window. For each window area, the weighted average drought intensity of each grid cell in the window area on the identification day is calculated to obtain the average drought intensity corresponding to the window area.

[0069] The window area with the highest average drought intensity is identified as the severely drought-stricken area on the identification date in the prediction region;

[0070] Step S3.3.4: Using formulas (9) and (10), the coordinates of the centroid center of the predicted region on the identification date are obtained:

[0071]

[0072] Where: C lon and C lat , representing the longitude and latitude of the center of mass, respectively; Lon z and Lat z , representing the longitude and latitude of the z-th grid in the severely drought-stricken area, respectively.

[0073] This invention also provides a system for implementing the aforementioned agricultural drought prediction method that integrates spatiotemporal feature learning, comprising:

[0074] The data collection and preprocessing module is used to acquire the time series of meteorological and hydrological elements in the prediction area over the most recent N days, and to perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, represented as: set X = (X1, X2, ..., X...). N ); where the prediction region is a rasterized region of H*W; where H and W represent the number of grid cells in the height and width directions of the prediction region, respectively; X1, X2, ..., X N , representing the standardized indices of meteorological and hydrological elements for the forecast area arranged chronologically on historical day 1, historical day 2, ..., historical day N; for X k , k = 1, 2, ..., N, where X is the standardized index of meteorological and hydrological elements in the prediction area on historical day k; k The spatial dimension contains H*W raster data, and each raster data has C=5 dimensional features, namely: the standardized maximum temperature index ST. max Standardized minimum temperature index ST min Standardized precipitation index (SPI), standardized runoff index (SRI), and standardized soil moisture index (SMI); C represents the dimension of the raster data.

[0075] The drought prediction model includes a spatial feature extraction module, a spatiotemporal feature learning module, and a sequence prediction module;

[0076] The spatial feature extraction module is used to extract the set X = (X1, X2, ..., X...). N Using this as input data, the standardized index X of meteorological and hydrological elements for each historical day k is sequentially processed. k Spatial feature extraction is performed to extract the spatial feature vector SF corresponding to each historical day k. k This forms the spatial feature matrix SF = (SF1, SF2, ..., SF2). N );

[0077] The spatiotemporal feature learning module is used to input the spatial feature matrix SF = (SF1, SF2, ..., SF3). N ), sequentially processing the spatial eigenvectors SF of each historical day k k Spatiotemporal feature learning is performed to obtain the spatiotemporal feature vector (TSF) corresponding to each historical day k. k This forms the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF). N );

[0078] The sequence prediction module is used to input the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF) N The standardized soil moisture index (SMI) sequence for the next M consecutive days is predicted and expressed as: SMI = (SMI N+1 SMI N+2 ,...,SMI N+M ); where M < N, SMI N+m The standardized soil moisture index SMI is the predicted standard soil moisture index for the m-th future day, where m = 1, 2, ..., M, and has H*W grids corresponding to the standardized soil moisture index SMI in the spatial dimension.

[0079] The drought event identification and prediction model is used to analyze the standardized soil moisture index (SMI) sequence for the next M consecutive days predicted by the sequence prediction module, extract drought events, analyze the spatial centroid movement characteristics of severe drought events, and predict the future spatiotemporal changes of drought.

[0080] The agricultural drought prediction method and system that integrates spatiotemporal feature learning provided by this invention has the following advantages:

[0081] This invention provides an agricultural drought prediction method and system that integrates spatiotemporal feature learning. It comprehensively considers the temporal and spatial attributes of drought, utilizes high-precision raster meteorological data, and inputs multiple hydrological and meteorological elements into a pre-trained convolutional neural network coupled with a deep learning model to predict future drought indicators. It estimates the spatiotemporal variation characteristics of future drought disasters in the region, fully considers the multidimensional attributes of drought and the important meteorological and hydrological elements affecting drought, and efficiently provides early warning of drought while maintaining the model complexity within an appropriate range, providing technical guidance for drought prevention and water resource management. Attached Figure Description

[0082] Figure 1 An overall diagram of an agricultural drought prediction method that integrates spatiotemporal feature learning provided by this invention;

[0083] Figure 2 This is a structural diagram of the drought prediction model provided by the present invention;

[0084] Figure 3 A graph of the standardized soil moisture index (SMI) provided in this embodiment of the invention;

[0085] Figure 4 The image shows the prediction results of the drought event identification and prediction model provided in this embodiment of the invention for drought events in the next 7 days. Detailed Implementation

[0086] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0087] To address the problem that current drought prediction methods only consider the temporal occurrence of drought and ignore its spatial variation, this paper proposes an agricultural drought prediction method and system that integrates spatiotemporal feature learning, fully considering the temporal attributes and spatial characteristics of drought. Furthermore, to address the issue that current drought prediction methods often use single variables, this paper employs standardized indices of multiple meteorological and hydrological elements affecting drought as model inputs, fully considering the influence of multiple variables and reducing redundant information in the original data, thus achieving efficient drought prediction.

[0088] Therefore, this invention provides an agricultural drought prediction method and system that integrates spatiotemporal feature learning. It comprehensively considers the temporal and spatial attributes of drought, utilizes high-precision raster meteorological data, and inputs multiple hydrological and meteorological elements into a pre-trained convolutional neural network coupled with a deep learning model to predict future drought indicators. It estimates the spatiotemporal variation characteristics of future drought disasters in the region, fully considering the multidimensional attributes of drought and the important meteorological and hydrological elements affecting drought. While maintaining model complexity within an appropriate range, it efficiently provides drought early warning, offering technical guidance for drought prevention and water resource management. This invention has a clear concept, is easy to operate, and is highly practical.

[0089] See Figure 1 This invention provides an agricultural drought prediction method that integrates spatiotemporal feature learning, comprising the following steps:

[0090] Step S1, Data Collection and Preprocessing:

[0091] Step S1.1, the prediction region is a rasterized region of H*W; where H and W represent the number of grid cells in the height and width directions of the prediction region, respectively;

[0092] Step S1.2: Collect ERA5-Land data products and hydrological station data for the prediction area, and perform spatial downscaling to obtain the time series of meteorological and hydrological elements for the prediction area over the most recent N days.

[0093] ERA5-Land, provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), is a high-resolution global surface reanalysis data product offering higher spatial resolution and making it more suitable for regional-scale applications and research. High spatial resolution meteorological raster data sequences can be obtained by using remote sensing meteorological data products such as ERA5-Land, or by interpolating regional hydrological station data. Specifically, spatial interpolation methods, such as bilinear interpolation and kriging interpolation, can be used to rasterize and downscale hydrological station data.

[0094] Obtain the time series of meteorological and hydrological elements for the forecast area over the most recent N days, and perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, represented as: set X = (X1, X2, ..., X...). N ); where X1, X2, ..., X N , representing the standardized indices of meteorological and hydrological elements for the forecast area arranged chronologically on historical day 1, historical day 2, ..., historical day N; for X k , k = 1, 2, ..., N, where X is the standardized index of meteorological and hydrological elements in the prediction area on historical day k; kThe spatial dimension contains H*W raster data, and each raster data has C=5 dimensional features, namely: the standardized maximum temperature index ST. max Standardized minimum temperature index ST min Standardized Precipitation Index (SPI), Standardized Runoff Index (SRI), and Standardized Soil Moisture Index (SMI); C represents the dimension of the raster data; such as Figure 3 The image shown is a distribution map of the standardized soil moisture index (SMI) for a region on May 2, 2016, provided in the embodiment.

[0095] In this step, standardization is specifically performed using formula (1):

[0096]

[0097] in, This represents the standardized index of the c-th meteorological and hydrological element in the h-th row and w-th column of the raster in the prediction area on historical day k; where h = 1, 2, ..., H; w = 1, 2, ..., W; c = 1, 2, ..., C;

[0098] μ represents the average value of the standardized index of the c-th meteorological and hydrological element over N historical days, and σ represents the standard deviation of the standardized index of the c-th meteorological and hydrological element over N historical days.

[0099] represent Values ​​before standardization.

[0100] Step S2, construct a drought prediction model; such as Figure 2 As shown, the drought prediction model includes a spatial feature extraction module, a spatiotemporal feature learning module, and a sequence prediction module; the drought prediction model can be a convolutional LSTM coupled model.

[0101] Step S2.1, set X = (X1, X2, ..., X...) N The data is input to the spatial feature extraction module, which then sequentially processes the standardized index X of meteorological and hydrological elements for each historical day k. k Spatial feature extraction is performed to extract the spatial feature vector SF corresponding to each historical day k. k This forms the spatial feature matrix SF = (SF1, SF2, ..., SF2). N );

[0102] Step S2.2, convert the spatial feature matrix SF = (SF1, SF2, ..., SF2) into a single matrix. N The input is fed into the spatiotemporal feature learning module, which sequentially processes the spatial feature vector SF of each historical day k. k Spatiotemporal feature learning is performed to obtain the spatiotemporal feature vector (TSF) corresponding to each historical day k. k This forms the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF). N );

[0103] Step S2.3, convert the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF) N The input is given to the sequence prediction module, which predicts the standardized soil moisture index (SMI) sequence for the next M consecutive days, expressed as: SMI = (SMI... N+1 SMI N+2 ,...,SMI N+M ); where M < N, SMI N+m Let S represent the predicted standardized soil moisture index (SMI) for the m-th future day, where m = 1, 2, ..., M, and it has H*W grid cells corresponding to the standardized soil moisture index SMI in the spatial dimension.

[0104] The steps S2.1 to S2.3 of the present invention are described in detail below:

[0105] The spatial feature extraction module in step S2.1 performs spatial feature extraction as follows:

[0106] Step S2.1.1, the spatial feature extraction module includes a Conv2D convolutional layer, a BatchNormalization normalization layer, and a PReLu activation layer;

[0107] Step S2.1.2, Standardized index X of meteorological and hydrological elements for historical day k k A data matrix of size H*W*C;

[0108] Standardized index of meteorological and hydrological elements of historical day k X k The input is fed into a Conv2D convolutional layer configured with 3×3 kernels. Local features are captured on the input data matrix by sliding the convolution kernel. Spatial feature vectors with spatial correlation are extracted using formula (2).

[0109]

[0110] Among them, K i,j,c The weights represent the convolution kernel;

[0111] Standardized index X of meteorological and hydrological elements representing historical day k k In the diagram, the value of the standardized index of the c-th meteorological and hydrological element in the raster of the (h+i)-th row and (w+j)-th column; c = 1, 2, ..., C;

[0112] This represents the spatial feature vector of the standardized index of the c-th meteorological and hydrological element in the h-th row and w-th column of the raster in the prediction area on historical day k.

[0113] The Conv2D convolutional layer can be a Conv2D layer wrapped by TimeDistributed, with a kernel size of (3,3) and padding of "same" to ensure that the spatial dimension of the output feature vector is consistent with that of the input.

[0114] Step S2.1.3: Superimpose the spatial feature vectors of the standardized indices of the C meteorological and hydrological elements in the h-th row and w-th column of the raster for historical day k in the prediction area to obtain the spatial feature vector of the h-th row and w-th column of the raster for historical day k in the prediction area.

[0115] Step S2.1.4, the BatchNormalization normalization layer, for each h-th row and w-th column of the prediction region of historical day k, the spatial feature vectors... Normalization is performed to enhance model stability and accelerate the training process, followed by activation using a PReLU activation layer to obtain the spatial feature vector SF for historical day k. k .

[0116] The process of spatiotemporal feature extraction in step S2.2 of the spatiotemporal feature learning module is as follows:

[0117] Step S2.2.1, the spatiotemporal feature learning module includes C layers of ConvLSTM2D convolutional layers; for the c-th ConvLSTM2D convolutional layer, c = 1, 2, ..., C, the convolution operation is performed using formula (3):

[0118] output(TSF k,c ) = f(conv(Input(SF) k,c ),W conv,c ,b c (3)

[0119] Where: Input(SF) k,c ) represents the spatial feature vector of the standardized index of the c-th meteorological and hydrological element in the historical day k of the input prediction region; f represents the activation function; conv represents the convolution operation; W conv,c b represents the convolutional kernel of layer c; c It is the bias term of the c-th layer;

[0120] output(TSF k,c ) is the spatiotemporal feature vector of the standardized index of the c-th meteorological and hydrological element in the historical day k of the prediction region;

[0121] As one implementation, the spatiotemporal feature learning module includes C stacked ConvLSTM2D convolutional layers. Each ConvLSTM2D convolutional layer combines the spatial feature extraction capability of convolution operations with the time series processing capability, enabling it to capture complex patterns and long-term dependencies in spatiotemporal data. The kernel size of all layers is (3, 3), and same padding is used to maintain the spatial dimension of the feature vector.

[0122] Step S2.2.2 yields the spatiotemporal feature vectors of the standardized meteorological and hydrological element indices for C prediction regions on historical day k, namely: TSF k,1 ,TSF k,2 ,...,TSF k,C ; For TSF k,1 ,TSF k,2 ,...,TSF k,C By superimposing the vectors, we obtain the spatiotemporal feature vector TSF corresponding to historical day k. k .

[0123] The process of the sequence prediction module in step S2.3 predicting SMI for the next M consecutive days is as follows:

[0124] Step S2.3.1, the sequence prediction module includes M ConvLSTM2D convolutional layers and M Conv2D convolutional layers;

[0125] Step S2.3.2, convert the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF) N The input is fed into the first ConvLSTM2D convolutional layer ConvLSTM2D1, and the spatiotemporal feature vector TSF of day N+1 is obtained by using formula (4). N+1 :

[0126] output(TSF N+1 )=ConvLSTM2D1(Input(TSF1,TSF2,...,TSF N ),hide1,cell1) (4)

[0127] Where: hide1 and cell1 represent the hidden state and cell state of the first ConvLSTM2D convolutional layer ConvLSTM2D1, respectively;

[0128] The spatiotemporal feature vector TSF of day N+1N+1 The input is fed into the first Conv2D convolutional layer, and the output is the standardized soil moisture index (SMI) for day N+1. N+1 ;

[0129] Step S2.3.3: Convert the spatiotemporal feature vector matrix TSF2,...,TSF N ,TSF N+1 The input is fed into the second ConvLSTM2D convolutional layer, ConvLSTM2D2, and Equation (5) is used to obtain the spatiotemporal feature vector TSF for the (N+2)th day. N+2 :

[0130] output(TSF N+2 )=ConvLSTM2D2(Input(TSF2,...,TSF N ,TSF N+1 ),hide2,cell2)(5)

[0131] Where: hide2 and cell2 represent the hidden state and cell state of the second ConvLSTM2D convolutional layer, respectively;

[0132] The spatiotemporal feature vector TSF of day N+2 N+2 The input is fed into the second Conv2D convolutional layer, and the output is the standardized soil moisture index (SMI) for day N+2. N+2 ;

[0133] Step S2.3.4, continue shifting one position to the right in this cycle to obtain the standardized soil moisture index (SMI) sequence for the next M consecutive days: SMI = (SMI... N+1 SMI N+2 ,...,SMI N+M ).

[0134] Step S3: Construct a drought event identification and prediction model; the drought event identification and prediction model analyzes the standardized soil moisture index (SMI) sequence for the next M consecutive days predicted by the sequence prediction module, extracts drought events, analyzes the spatial centroid movement characteristics of severe drought events, and predicts the future spatiotemporal changes of drought.

[0135] Step S3 is as follows:

[0136] Step S3.1: For each grid cell in the prediction area, obtain the standardized soil moisture index (SMI) sequence for the historical N days and the next M consecutive days. Therefore, N+M standardized soil moisture indices (SMI) are obtained, represented as: Sequence SMI N+M =(SMI1,SMI2,...,SMI N SMIN+1 SMI N+2 ,...,SMI N+M );

[0137] For example, standardized soil moisture index (SMI) sequences for the past 14 days and the next 7 consecutive days can be obtained.

[0138] Step S3.2, for sequence SMI N+M Perform trend analysis on the time-varying characteristics to determine whether the trend is significant. If so, then the series is SMI (Search Engine Management). N+M Assign the value to the corresponding grid cell; otherwise, assign the value 0 to the grid cell.

[0139] In step S3.2, the following method is used to determine whether a trend is significant:

[0140] Step S3.2.1: Using formula (6), the sequence SMI is calculated. N+M =(SMI1,SMI2,...,SMI N SMI N+1 SMI N+2 ,...,SMI N+M The cumulative difference d:

[0141]

[0142] Where: q = 1, 2, ..., N+M-1; p = q+1, q+2, ..., N+M; d represents the sequence SMI. N+M The cumulative difference between the various SMI data points, d > 0 indicates that the SMI data sequence tends to increase, and vice versa; sign() represents the sign function;

[0143] Step S3.2.2, using formula (7), the concentration Var(d) of the cumulative difference d is obtained:

[0144]

[0145] in:

[0146] R represents SMI1, SMI2, ..., SMI N SMI N+1 SMI N+2 ,...,SMI N+M In the given information, there are R distinct SMIs that repeat at least twice; for the r-th repeating SMI, its number of repetitions is Num. r ;

[0147] For ease of explanation, this section will only use a sequence of 7 SMIs as an example to introduce Num. r The meaning of R:

[0148] Suppose the values ​​of SMI1, SMI2, ..., SMI7 are 1.1, 1.2, 1.3, 1.1, 1.1, 1.2, and 1.4, respectively. Then there are two distinct SMIs that repeat. The first repeating SMI has a value of 1.1 and repeats 3 times. The second repeating SMI has a value of 1.2 and repeats 2 times.

[0149] Step S3.2.3, using formula (8), obtain the statistic St:

[0150]

[0151] Step S3.2.4: If the statistic St exceeds the set threshold, it indicates that the sequence SMI has occurred. N+M It has a significant trend; otherwise, it represents the SMI sequence. N+M It does not show a significant trend.

[0152] As one implementation, if St is greater than 1.65, it indicates that the trend significance has passed the 90% confidence level, which represents a series SMI. N+M It has a significant trend; otherwise, the sequence is considered SMI. N+M It does not show a significant trend.

[0153] Step S3.3: Perform daily analysis of the distribution of the standardized soil moisture index (SMI) of the predicted area for the most recent M days and the next M consecutive days to determine the centroid of the severely drought area each day, thereby obtaining the movement characteristics of the centroid of the severely drought area over the most recent M days and the next M consecutive days.

[0154] In step S3.3, the method for determining the centroid of the severely drought-stricken area each day is as follows:

[0155] Step S3.3.1: Take each day in the most recent historical M days and the next consecutive M days as the identification day. On each identification day, read the standardized soil moisture index SMI of each grid in the prediction area on the identification day.

[0156] For example, drought events can be identified based on the most recent 7 days and the next 7 consecutive days.

[0157] Step S3.3.2: Identify whether a drought event occurred on the identification day for each grid cell:

[0158] The standardized soil moisture index (SMI) of each grid cell on the identification day is compared with 0. If the SMI is less than 0, it has negative variation, indicating that the soil moisture is low and a drought event has occurred on the identification day. The SMI value is then used as the drought intensity of the grid cell on the identification day. Conversely, if the SMI is greater than or equal to 0, it has positive variation, indicating that the soil moisture is high and no drought event has occurred. The grid cell does not have drought intensity on the identification day.

[0159] Step S3.3.3: Determine the severely drought-stricken areas in the prediction region on each identification day:

[0160] The prediction area is traversed in a 5×5 window. For each window area, the weighted average drought intensity of each grid cell in the window area on the identification day is calculated to obtain the average drought intensity corresponding to the window area.

[0161] The window area with the highest average drought intensity is identified as the severely drought-stricken area on the identification date in the prediction region;

[0162] Step S3.3.4: Using formulas (9) and (10), the coordinates of the centroid center of the predicted region on the identification date are obtained:

[0163]

[0164] Where: C lon and C lat , representing the longitude and latitude of the center of mass, respectively; Lon z and Lat z , representing the longitude and latitude of the z-th grid in the severely drought-stricken area, respectively.

[0165] As an example, such as Figure 4 The image shows the distribution cloud map of the Standardized Soil Moisture Index (SMI) for the most recent 7 days and the predicted next 7 days, as identified by the model of this invention. The most recent 7 days are from July 16, 2015 to July 22, 2015. The next 7 days are from July 23, 2015 to July 29, 2015. Furthermore, in the distribution cloud map of the Standardized Soil Moisture Index (SMI) for the next 7 days, black dots are used to represent identified drought events.

[0166] Using the method of this invention, the drought events for each day of the next 7 days were further calculated, and the centroid prediction results for the next 7 days are shown in Table 1:

[0167] Table 1: Centroid Prediction Results for the Next 7 Days

[0168]

[0169] The test results showed that the predicted centroid position for the next 7 days deviated less than the actual location of drought in the predicted area, meeting the prediction accuracy requirements and accurately reflecting the spatial occurrence and spread of drought.

[0170] This invention also provides an agricultural drought prediction system that integrates spatiotemporal feature learning, comprising:

[0171] The data collection and preprocessing module is used to acquire the time series of meteorological and hydrological elements in the prediction area over the most recent N days, and to perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, represented as: set X = (X1, X2, ..., X...). N ); where the prediction region is a rasterized region of H*W; where H and W represent the number of grid cells in the height and width directions of the prediction region, respectively; X1, X2, ..., X N , representing the standardized indices of meteorological and hydrological elements for the forecast area arranged chronologically on historical day 1, historical day 2, ..., historical day N; for X k , k = 1, 2, ..., N, where X is the standardized index of meteorological and hydrological elements in the prediction area on historical day k; k The spatial dimension contains H*W raster data, and each raster data has C=5 dimensional features, namely: the standardized maximum temperature index ST. max Standardized minimum temperature index ST min Standardized precipitation index (SPI), standardized runoff index (SRI), and standardized soil moisture index (SMI); C represents the dimension of the raster data.

[0172] The drought prediction model includes a spatial feature extraction module, a spatiotemporal feature learning module, and a sequence prediction module;

[0173] The spatial feature extraction module is used to extract the set X = (X1, X2, ..., X...). N Using this as input data, the standardized index X of meteorological and hydrological elements for each historical day k is sequentially processed. k Spatial feature extraction is performed to extract the spatial feature vector SF corresponding to each historical day k. k This forms the spatial feature matrix SF = (SF1, SF2, ..., SF2). N );

[0174] The spatiotemporal feature learning module is used to input the spatial feature matrix SF = (SF1, SF2, ..., SF3). N ), sequentially processing the spatial eigenvectors SF of each historical day k k Spatiotemporal feature learning is performed to obtain the spatiotemporal feature vector (TSF) corresponding to each historical day k. kThis forms the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF). N );

[0175] The sequence prediction module is used to input the spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF) N The standardized soil moisture index (SMI) sequence for the next M consecutive days is predicted and expressed as: SMI = (SMI N+1 SMI N+2 ,...,SMI N+M ); where M < N, SMI N+m The standardized soil moisture index SMI is the predicted standard soil moisture index for the m-th future day, where m = 1, 2, ..., M, and has H*W grids corresponding to the standardized soil moisture index SMI in the spatial dimension.

[0176] The drought event identification and prediction model is used to analyze the standardized soil moisture index (SMI) sequence for the next M consecutive days predicted by the sequence prediction module, extract drought events, analyze the spatial centroid movement characteristics of severe drought events, and predict the future spatiotemporal changes of drought.

[0177] The agricultural drought prediction method and system that integrates spatiotemporal feature learning provided by this invention has the following advantages:

[0178] This invention relates to an agricultural drought prediction method and system that integrates spatiotemporal feature learning. It employs high-precision raster meteorological and hydrological data sequences to construct a coupled machine learning model capable of extracting the spatiotemporal features of drought indicators. While controlling model complexity, it fully captures the temporal and spatial characteristics of drought. By fully considering drought-related meteorological and hydrological elements and using standardized indices to predict future drought indicators, it can extract long-term trends of meteorological and hydrological elements and respond to short-term extreme events and hydrological changes. This reduces redundant information in the original data and efficiently predicts the spatial occurrence and spread of drought in drought-stricken areas, providing technical guidance for drought prevention and water resource management. This invention has a clear concept, is easy to operate, and is highly practical.

[0179] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An agricultural drought prediction method integrating spatiotemporal feature learning, characterized in that, Includes the following steps: Step S1, Data Collection and Preprocessing: Step S1.1, the prediction region is a rasterized region of H*W; where H and W represent the number of grid cells in the height and width directions of the prediction region, respectively; Step S1.2: Obtain the time series of meteorological and hydrological elements for the prediction area over the most recent N days, and perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, represented as: set ;in, , representing the standardized indices of meteorological and hydrological elements for the forecast area arranged chronologically on historical day 1, historical day 2, ..., historical day N; for , To predict the region's historical days Standardized indices of meteorological and hydrological elements; There are H*W raster data in the spatial dimension, and each raster data is... =5-dimensional elements, namely: Standardized Maximum Temperature Index (ST) max Standardized minimum temperature index ST min Standardized precipitation index (SPI), standardized runoff index (SRI), and standardized soil moisture index (SMI); Indicates the dimension of raster data; Step S2: Construct a drought prediction model; the drought prediction model includes a spatial feature extraction module, a spatiotemporal feature learning module, and a sequence prediction module; Step S2.1, set As input data, the data is fed into the spatial feature extraction module, which sequentially standardizes the meteorological and hydrological element indices for each historical day k. Spatial feature extraction is performed to extract the spatial feature vector corresponding to each historical day k. Forming a spatial feature matrix ; Step S2.2, convert the spatial feature matrix The input is fed into the spatiotemporal feature learning module, which sequentially processes the spatial feature vectors of each historical day k. Spatiotemporal feature learning is performed to obtain the spatiotemporal feature vector corresponding to each historical day k. To form a spatiotemporal feature vector matrix ; Step S2.3, convert the spatiotemporal feature vector matrix The input is given to the sequence prediction module, which then predicts future consecutive sequences. The standardized soil moisture index (SMI) sequence for 1 day is represented as follows: ;in, , Represents the predicted future. The Standardized Soil Moisture Index (SMI) of Japan It has H*W grids in the spatial dimension corresponding to the standardized soil moisture index (SMI); Step S3: Construct a drought event identification and prediction model; the drought event identification and prediction model is used to predict future continuous drought events obtained by the sequence prediction module. The standardized soil moisture index (SMI) sequence of the day was analyzed to extract drought events, and the spatial centroid movement characteristics of severe drought events were analyzed to predict the spatiotemporal changes of future drought.

2. The agricultural drought prediction method based on spatiotemporal feature learning according to claim 1, characterized in that, The steps preceding step S1.2 also include: Collect ERA5-Land data products and hydrological station data for the forecast area, and perform spatial downscaling to obtain the time series of meteorological and hydrological elements for the forecast area over the most recent N days.

3. The agricultural drought prediction method based on spatiotemporal feature learning according to claim 1, characterized in that, In step S1.2, standardization is performed using formula (1): （1）， in, Indicates on historical days The first prediction region Line number The first column of the grid Standardized indices for meteorological and hydrological elements; among them... ; ; ; μ represents the first The standardized index of a meteorological and hydrological element over N historical days, σ represents the average value of the Nth meteorological and hydrological element. Standard deviation of a meteorological and hydrological element standardized index over N historical days; represent Values ​​before standardization.

4. The agricultural drought prediction method based on spatiotemporal feature learning according to claim 1, characterized in that, Step S2.1 is as follows: Step S2.1.1, the spatial feature extraction module includes a Conv2D convolutional layer, a BatchNormalization normalization layer, and a PReLu activation layer; Step S2.1.2, Standardized index of meteorological and hydrological elements for historical day k A data matrix of size H*W*C; Standardized index of meteorological and hydrological elements of historical days The input is fed into a Conv2D convolutional layer configured with 3×3 kernels. Local features are captured on the input data matrix by sliding the convolution kernel. Spatial feature vectors with spatial correlation are extracted using formula (2). （2）， in, The weights represent the convolution kernel; Standardized index of meteorological and hydrological elements representing historical day k In the middle, the first Line number The first column of the grid The values ​​of standardized indices for meteorological and hydrological elements; ; Indicates on historical days The first prediction region Line number The first column of the grid Spatial characteristic vectors of standardized indices of meteorological and hydrological elements; Step S2.1.3, transfer historical dates The first prediction region Line number Column grid By superimposing the spatial characteristic vectors of standardized indices of meteorological and hydrological elements, historical daily data can be obtained. The first prediction region Line number Spatial feature vector of a column grid ; Step S2.1.4, the BatchNormalization normalization layer, for historical days The prediction regions of each Line number Spatial feature vector of a column grid After normalization, the vector is activated by a PReLU activation layer to obtain the spatial feature vector of historical day k. .

5. The agricultural drought prediction method based on spatiotemporal feature learning according to claim 1, characterized in that, Step S2.2 specifically includes: Step S2.2.1, the spatiotemporal feature learning module includes Layer ConvLSTM2D convolutional layer; for the th ConvLSTM2D convolutional layers of the layer, Convolution is performed using formula (3): （3）， in: The input prediction region in historical days The Spatial feature vectors of standardized indices of meteorological and hydrological elements; f represents the activation function; conv represents the convolution operation; Indicates the first Convolutional kernels of layers; It is the first Layer bias terms; The input prediction region in historical days The Spatiotemporal characteristic vectors of standardized indices of meteorological and hydrological elements; Step S2.2.2, thus obtaining The predicted region is in historical days The spatiotemporal characteristic vectors of the standardized indices of meteorological and hydrological elements are as follows: ;right By superimposing the vectors, we obtain the spatiotemporal feature vectors corresponding to historical day k. .

6. The agricultural drought prediction method based on spatiotemporal feature learning according to claim 1, characterized in that, Step S2.3 specifically includes: Step S2.3.1, the sequence prediction module includes One ConvLSTM2D convolutional layer and One Conv2D convolutional layer; Step S2.3.2, convert the spatiotemporal feature vector matrix Input to the first ConvLSTM2D convolutional layer Using formula (4), we obtain the first... Spatiotemporal feature vector of day : （4）， in: and , respectively representing the first ConvLSTM2D convolutional layer The implicit state and cellular state; The first Spatiotemporal feature vector of day The input is fed into the first Conv2D convolutional layer, and the output is the first Conv2D convolutional layer. Standardized Soil Moisture Index (SMO) ; Step S2.3.3, convert the spatiotemporal feature vector matrix Input to the second ConvLSTM2D convolutional layer Using formula (5), we obtain the first... Spatiotemporal feature vector of day : （5）， in: and , respectively representing the second ConvLSTM2D convolutional layer The implicit state and cellular state; The first Spatiotemporal feature vector of day The input is fed into the second Conv2D convolutional layer, and the output is the... Standardized Soil Moisture Index (SMO) ; Steps S2, 3, and 4 are repeated, shifting one position to the right in a loop, to obtain the future sequence. Standardized Soil Moisture Index (SMI) Series .

7. The agricultural drought prediction method based on spatiotemporal feature learning according to claim 1, characterized in that, Step S3 is as follows: Step S3.1: For each grid cell in the prediction area, obtain the historical N days and the future continuous N days. The standardized soil moisture index (SMI) sequence for the day was used to obtain N+. A standardized soil moisture index (SMI) is represented as: sequence ; Step S3.2, for the sequence Perform trend analysis on the time-varying characteristics to determine whether the trend is significant. If so, then classify the series... Assign the value to the corresponding grid cell; otherwise, assign the value 0 to the grid cell. Step S3.3, for the most recent history Tianhe Future Continuous The distribution of the standardized soil moisture index (SMI) in the predicted area was analyzed daily to determine the centroid of the severely drought-stricken area each day. This yields the centroid of the severely arid region. In recent history Tianhe Future Continuous The movement characteristics of the sky.

8. The agricultural drought prediction method based on spatiotemporal feature learning according to claim 7, characterized in that, Step S3.2 specifically includes: Step S3.2.1: Calculate the sequence using formula (6). Cumulative difference d: （6）， in: ; d represents a sequence The cumulative difference between the various SMI data points is d. A value greater than 0 indicates that the SMI data sequence tends to increase, and vice versa; sign() represents the sign function. Step S3.2.2, using formula (7), the concentration Var(d) of the cumulative difference d is obtained: (7), in: represent In China, there are a total of A different It repeats at least twice; for the first A recurring Its recurrence count is ; Step S3.2.3: Use formula (8) to obtain the statistic. : （8） Step S3.2.4, if the statistic If the threshold is exceeded, it represents a sequence. It has a significant trend; otherwise, it represents a sequence. It does not show a significant trend.

9. The agricultural drought prediction method based on spatiotemporal feature learning according to claim 7, characterized in that, In step S3.3, the daily centroid of the severely drought-stricken area The method for determining it is as follows: Step S3.3.1, retrieve the most recent history Tianhe Future Continuous Each day in the sky is used as the identification day. On each identification day, the standardized soil moisture index (SMI) of each grid cell in the prediction area is read. Step S3.3.2: Identify whether a drought event occurred on the identification day for each grid cell: The standardized soil moisture index (SMI) of each grid cell on the identification day is compared with 0. If the SMI is less than 0, it has a negative variation, indicating that the soil moisture is low. The grid cell is identified as having experienced a drought event on the identification day, and its SMI value is used as the drought intensity of the grid cell on the identification day. Conversely, if SMI is greater than or equal to 0, it has positive variation, indicating that the soil moisture is relatively high and no drought event has occurred. The grid does not have drought intensity on the identification day. Step S3.3.3: Determine the severely drought-stricken areas in the prediction region on each identification day: The prediction area is traversed in a 5×5 window. For each window area, the weighted average drought intensity of each grid cell in the window area on the identification day is calculated to obtain the average drought intensity corresponding to the window area. The window area with the highest average drought intensity is identified as the severely drought-stricken area on the identification date in the prediction region; Step S3.3.4: Using formulas (9) and (10), the centroid of the predicted region in the severely drought-stricken area on the identification date is obtained. Coordinates: （9）， （10）， in: and , respectively representing the center of mass Longitude and latitude; and , respectively representing the severe drought areas The longitude and latitude of each grid cell.

10. A system for implementing the agricultural drought prediction method based on spatiotemporal feature learning as described in any one of claims 1-9, characterized in that, include: The data collection and preprocessing module is used to acquire the time series of meteorological and hydrological elements in the prediction area over the most recent N days, and to perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, represented as: set The prediction region is a rasterized region of size H*W; where H and W represent the number of grid cells in the height and width directions of the prediction region, respectively. , representing the standardized indices of meteorological and hydrological elements for the forecast area arranged chronologically on historical day 1, historical day 2, ..., historical day N; for , To predict the region's historical days Standardized indices of meteorological and hydrological elements; There are H*W raster data in the spatial dimension, and each raster data is... =5-dimensional elements, namely: Standardized Maximum Temperature Index (ST) max Standardized minimum temperature index ST min Standardized precipitation index (SPI), standardized runoff index (SRI), and standardized soil moisture index (SMI); Indicates the dimension of raster data; The drought prediction model includes a spatial feature extraction module, a spatiotemporal feature learning module, and a sequence prediction module; The spatial feature extraction module is used to extract the set As input data, the standardized indices of meteorological and hydrological elements for each historical day k are processed sequentially. Spatial feature extraction is performed to extract the spatial feature vector corresponding to each historical day k. Forming a spatial feature matrix ; The spatiotemporal feature learning module is used to input the spatial feature matrix. The spatial feature vectors of each historical day k are processed sequentially. Spatiotemporal feature learning is performed to obtain the spatiotemporal feature vector corresponding to each historical day k. To form a spatiotemporal feature vector matrix ; The sequence prediction module is used to input the spatiotemporal feature vector matrix. Predicting future continuous The standardized soil moisture index (SMI) sequence for 1 day is represented as follows: ;in, , Represents the predicted future. The Standardized Soil Moisture Index (SMI) of Japan It has H*W grids in the spatial dimension corresponding to the standardized soil moisture index (SMI); A drought event identification and prediction model is used to predict future consecutive drought events from the sequence prediction module. The standardized soil moisture index (SMI) sequence of the day was analyzed to extract drought events, and the spatial centroid movement characteristics of severe drought events were analyzed to predict the spatiotemporal changes of future drought.

Images