Agricultural drought prediction method and system fused with spatio-temporal feature learning

Abstract

The invention provides an agricultural drought prediction method and system fused with spatial-temporal feature learning. The method comprises the following steps: collecting and preprocessing data; constructing a drought prediction model which comprises a spatial feature extraction module, a spatial-temporal feature learning module and a sequence prediction module; constructing a drought event identification and prediction model; and the drought event identification and prediction model analyzes the standardized soil humidity index SMI sequence of the future continuous M days predicted by the sequence prediction module, extracts the drought event, analyzes and obtains the spatial centroid movement characteristics of the serious drought event, and predicts and obtains the time-space change condition of the future drought. According to the method, time and space attributes of drought are comprehensively considered, high-precision grid meteorological data are utilized, a plurality of hydro meteorological elements are input into the trained convolutional neural network and deep learning coupling model to predict future drought indexes, the complexity of the model is kept within a proper range, meanwhile, early warning is efficiently carried out on the drought, and the drought early warning efficiency is improved. And technical guidance is provided for drought control and water resource management.

Description

Technical Field

[0001] The present invention belongs to the technical field of drought prediction, and specifically relates to an agricultural drought prediction method and system integrating spatiotemporal feature learning. Background Art

[0003] Many scholars have proposed different methods to address the difficulty of drought prediction. Early drought prediction was mainly based on dynamic prediction based on hydrological model simulation results. The calculations were relatively complex, and the hydrological model also brought a lot of uncertainty. In recent years, with the continuous development of machine learning, scholars have discovered its potential in drought prediction, so drought prediction based on deep learning has become a research hotspot. However, existing drought prediction methods usually only focus on the temporal changes of regional droughts, ignoring the spatial development process, making it difficult to comprehensively and accurately predict agricultural droughts. Summary of the invention

[0004] In view of the defects of the prior art, the present invention provides an agricultural drought prediction method and system integrating spatiotemporal feature learning, which can effectively solve the above problems.

[0005] The technical solution adopted by the present invention is as follows:

[0006] The present invention provides an agricultural drought prediction method integrating spatiotemporal feature learning, comprising the following steps:

[0007] Step S1, data collection and preprocessing:

[0008] Step S1.1, the prediction area is a gridded area of ​​H*W, where H and W represent the number of grids in the height and width directions of the prediction area respectively;

[0009] Step S1.2, obtain the time series of meteorological and hydrological elements in the forecast area in the recent N days, and perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, which is expressed as: set X = (X1, X2, ..., X N ), where X1, X2, ..., X N , respectively represent the standardized indexes of meteorological and hydrological elements of the forecast area on historical day 1, historical day 2, …, historical day N in chronological order; for X k , k = 1, 2, ..., N, is the standardized index of meteorological and hydrological elements in the forecast area on historical day k; X k In the spatial dimension, there are H*W grid data, each grid data is C = 5-dimensional elements, respectively: 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;

[0010] Step S2, constructing a drought prediction model; the drought prediction model includes a spatial feature extraction module, a spatiotemporal feature learning module and a sequence prediction module;

[0011] Step S2.1, set X = (X1, X2, ..., X N ) as input data, and input it into the spatial feature extraction module, which sequentially extracts the standardized index X of meteorological and hydrological elements of each historical day k. k Perform spatial feature extraction to extract the spatial feature vector SF corresponding to each historical day k k , forming a spatial feature matrix SF = (SF1, SF2, ..., SF N );

[0012] Step S2.2, transform the spatial feature matrix SF = (SF1, SF2, ..., SF N ) is input into the spatiotemporal feature learning module, which sequentially calculates the spatial feature vector SF of each historical day k. k Perform spatiotemporal feature learning to obtain the spatiotemporal feature vector TSF corresponding to each historical day k k , forming the space-time eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N );

[0013] Step S2.3, transform the spatiotemporal eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N ) is input to the sequence prediction module, and the sequence prediction module 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 Represents the predicted standardized soil moisture index SMI for the mth day in the future, m = 1, 2, ..., M, which has standardized soil moisture index SMI corresponding to H*W grids in the spatial dimension;

[0014] Step S3, constructing a drought event identification and prediction model; the drought event identification and prediction model analyzes the standardized soil moisture index SMI sequence of the next M consecutive days predicted by the sequence prediction module, extracts drought events, and analyzes the spatial centroid movement characteristics of severe drought events to predict the future spatiotemporal changes of drought.

[0015] Preferably, before step S1.2, the following steps are also included:

[0016] The ERA5-Land data products and hydrological station data of the forecast area are collected and spatially downscaled to obtain the grid-by-grid time series of meteorological and hydrological elements in the forecast area in the most recent N days.

[0017] Preferably, in step S1.2, formula (1) is used for standardization:

[0018]

[0019] in, It represents the normalized index of the cth meteorological and hydrological element of the grid in the hth row and the wth column of the forecast area on the historical day k; where h = 1, 2, ..., H; w = 1, 2, ..., W; c = 1, 2, ..., C;

[0020] μ represents the average value of the standardized index of the cth meteorological and hydrological element in the history of N days, and σ represents the standard deviation of the standardized index of the cth meteorological and hydrological element in the history of N days;

[0021] represent The value before normalization.

[0022] Preferably, step S2.1 is specifically:

[0023] Step S2.1.1, the spatial feature extraction module includes a Conv2D convolution layer, a BatchNormalization normalization layer and a PReLu activation layer;

[0024] Step S2.1.2, the standardized index X of meteorological and hydrological elements on historical day k k is the data matrix of H*W*C;

[0025] The standardized index X of meteorological and hydrological elements of historical day k k The data is input to the Conv2D convolution layer configured with a 3×3 convolution kernel. The local features are captured on the input data matrix by sliding the convolution kernel. The spatial feature vector with spatial correlation is extracted using formula (2):

[0026]

[0027] Among them, K i,j,c Represents the weight of the convolution kernel;

[0028] The standardized index X of meteorological and hydrological elements representing historical day k k, the value of the standardized index of the cth meteorological and hydrological element of the grid in the h+ith row and the w+jth column; c = 1, 2, ..., C;

[0029] represents the spatial characteristic vector of the standardized index of the cth meteorological and hydrological element of the grid in the hth row and wth column of the forecast area on the historical day k;

[0030] Step S2.1.3, superimpose the spatial characteristic vectors of the C meteorological and hydrological element standardized indexes of the grid of the hth row and the wth column of the prediction area on the historical day k to obtain the spatial characteristic vector of the grid of the hth row and the wth column of the prediction area on the historical day k

[0031] Step S2.1.4, the BatchNormalization normalization layer, for each h-th row and w-th column of the prediction area of ​​historical day k, the spatial feature vector After normalization, it is activated through the PReLu activation layer to obtain the spatial feature vector SF of historical day k k .

[0032] Preferably, step S2.2 is specifically:

[0033] Step S2.2.1, the spatiotemporal feature learning module includes C layers of ConvLSTM2D convolutional layers; for the c-th layer of ConvLSTM2D convolutional layer, c = 1, 2, ..., C, the convolution operation is performed using formula (3):

[0034] output(TSF k,c )=f(conv(Input(SF k,c ),W conv,c ,b c )) (3)

[0035] Among them: Input(SF k,c ) is the spatial feature vector of the standardized index of the cth meteorological and hydrological element of the input prediction area on the historical day k; f represents the activation function; conv represents the convolution operation; W conv,c represents the convolution kernel of the cth layer; b c is the bias term of the cth layer;

[0036] output(TSF k,c ) is the spatiotemporal characteristic vector of the standardized index of the cth meteorological and hydrological element of the input prediction area on the historical day k;

[0037] Step S2.2.2, thus obtaining the spatiotemporal characteristic vectors of the standardized index of meteorological and hydrological elements in the C prediction areas on historical day k, respectively: TSF k,1,TSF k,2 ,...,TSF k,C ; for TSF k,1 ,TSF k,2 ,...,TSF k,C Superposition is performed to obtain the spatiotemporal feature vector TSF corresponding to historical day k k .

[0038] Preferably, step S2.3 is specifically:

[0039] Step S2.3.1, the sequence prediction module includes M ConvLSTM2D convolutional layers and M Conv2D convolutional layers;

[0040] Step S2.3.2, transform the spatiotemporal eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N ) is input to the first ConvLSTM2D convolutional layer ConvLSTM2D1, and the spatiotemporal feature vector TSF of the N+1th day is obtained using formula (4): N+1 :

[0041] output(TSF N+1 )=ConvLSTM2D1(Input(TSF1,TSF2,...,TSF N ),hide1,cell1) (4)

[0042] Among them: hide1 and cell1 represent the hidden state and cell state of the first ConvLSTM2D convolutional layer ConvLSTM2D1 respectively;

[0043] The spatiotemporal feature vector TSF of day N+1 N+1 Input to the first Conv2D convolution layer and output the standardized soil moisture index SMI on the N+1th day N+1 ;

[0044] Step S2.3.3, transform the spatiotemporal eigenvector matrix TSF2,...,TSF N ,TSF N+1 Input to the second ConvLSTM2D convolutional layer ConvLSTM2D2, and use formula (5) to obtain the spatiotemporal feature vector TSF of the N+2th day N+2 :

[0045] output(TSF N+2 )=ConvLSTM2D2(Input(TSF2,...,TSF N ,TSF N+1 ),hide2,cell2)(5)

[0046] Among them: hide2 and cell2 represent the hidden state and cell state of the second ConvLSTM2D convolutional layer ConvLSTM2D2 respectively;

[0047] The spatiotemporal feature vector TSF of day N+2 N+2 Input to the second Conv2D convolution layer and output the standardized soil moisture index SMI on the N+2th day N+2 ;

[0048] Step S2.3.4, keep moving back one position and repeating the cycle, so as to obtain the standardized soil moisture index SMI sequence SMI=(SMI N+1 ,SMI N+2 ,...,SMI N+M ).

[0049] Preferably, step S3 specifically comprises:

[0050] Step S3.1, for each grid in the prediction area, obtain the standardized soil moisture index SMI sequence for the past N days and the next M consecutive days, thus obtaining N+M standardized soil moisture index SMIs, expressed as: sequence SMI N+M =(SMI1, SMI2, ..., SMI N ,SMI N+1 ,SMI N+2 ,...,SMI N+M );

[0051] Step S3.2, for sequence SMI N+M Perform trend analysis on time-varying characteristics to determine whether the trend is significant. If so, the sequence SMI N+M Assign a value to the corresponding grid, otherwise, the grid is assigned a value of 0;

[0052] Step S3.3, analyze the distribution of the standardized soil moisture index SMI of the forecast area in the recent historical M days and the next M consecutive days on a daily basis, determine the centroid center of the severe drought area on a daily basis, and thus obtain the moving characteristics of the centroid center of the severe drought area in the recent historical M days and the next M consecutive days.

[0053] Preferably, step S3.2 is specifically:

[0054] Step S3.2.1, using formula (6), calculate the sequence SMI N+M =(SMI1, SMI2, ..., SMI N ,SMI N+1 ,SMI N+2 ,...,SMI N+M) of the cumulative difference d:

[0055]

[0056] 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 each SMI data in , d greater than 0 means that the SMI data sequence tends to become larger, otherwise it becomes smaller; sign() represents the sign function;

[0057] Step S3.2.2, using formula (7), obtain the concentration Var(d) of the cumulative difference d:

[0058]

[0059] in:

[0060] R stands for SMI1, SMI2, ..., SMI N ,SMI N+1 ,SMI N+2 ,...,SMI N+M There are R different SMIs that appear at least twice in total; for the rth repeated SMI, its number of repetitions is Num r ;

[0061] Step S3.2.3, using formula (8), obtain the statistic St:

[0062]

[0063] Step S3.2.4, if the statistic St exceeds the set threshold, it means that the sequence SMI N+M There is a significant trend; otherwise, it represents the sequence SMI N+M There is no significant trend.

[0064] Preferably, in step S3.3, the method for determining the centroid center of the daily severe drought area is:

[0065] Step S3.3.1, taking each of the recent M historical days and the future M consecutive days as an identification day, and reading the standardized soil moisture index SMI of each grid in the prediction area on each identification day;

[0066] Step S3.3.2, identify whether a drought event occurs in each grid on the identified day:

[0067] The standardized soil moisture index (SMI) of each grid 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 less, and it is recognized that a drought event has occurred in the grid on the identification day, and its SMI value is used as the drought intensity of the grid on the identification day; conversely, if the SMI is greater than or equal to 0, it has a positive variation, indicating that the soil moisture is more, no drought event has occurred, and the grid does not have a drought intensity on the identification day;

[0068] Step S3.3.3, determine the severe drought area in the forecast area on each identified day:

[0069] The prediction area is traversed with a 5×5 window. For each window area, the weighted average of the drought intensity of each grid in the window area on the identification day is calculated to obtain the mean drought intensity corresponding to the window area.

[0070] Determine the window area with the largest mean drought intensity, which is the severe drought area in the forecast area on the identification day;

[0071] Step S3.3.4, using formula (9) and formula (10), obtain the coordinates of the centroid center of the severe drought area in the forecast area on the identification day:

[0072]

[0073] Where: C lon and C lat , respectively representing the longitude and latitude of the center of mass center; Lon z and Lat z , respectively represent the longitude and latitude of the zth grid in the severe drought area.

[0074] The present invention also provides a system for implementing the agricultural drought prediction method integrating spatiotemporal feature learning, comprising:

[0075] The data collection and preprocessing module is used to obtain the time series of meteorological and hydrological elements in the forecast area in the recent N days, and perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, which is expressed as: set X = (X1, X2, ..., X N ), where the prediction area is a gridded area of ​​H*W; where H and W represent the number of grids in the height and width directions of the prediction area, respectively; X1, X2, ..., X N , respectively represent the standardized indexes of meteorological and hydrological elements of the forecast area on historical day 1, historical day 2, …, historical day N in chronological order; for X k , k = 1, 2, ..., N, is the standardized index of meteorological and hydrological elements in the forecast area on historical day k; X kIn the spatial dimension, there are H*W grid data, each grid data is C = 5-dimensional elements, respectively: 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;

[0076] Drought prediction model, including spatial feature extraction module, spatiotemporal feature learning module and sequence prediction module;

[0077] The spatial feature extraction module is used to extract the set X=(X1, X2, ..., X N ) as input data, and then the standardized index X of meteorological and hydrological elements of each historical day k is calculated. k Perform spatial feature extraction to extract the spatial feature vector SF corresponding to each historical day k k , forming a spatial feature matrix SF = (SF1, SF2, ..., SF N );

[0078] The spatiotemporal feature learning module is used to input the spatial feature matrix SF = (SF1, SF2, ..., SF N ), and then the spatial feature vector SF of each historical day k is k Perform spatiotemporal feature learning to obtain the spatiotemporal feature vector TSF corresponding to each historical day k k , forming the space-time eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N );

[0079] Sequence prediction module, used to input spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF N ), the standardized soil moisture index SMI sequence for the next M consecutive days is predicted, expressed as: SMI = (SMI N+1 ,SMI N+2 ,...,SMI N+M ); where M<N, SMI N+m Represents the predicted standardized soil moisture index SMI for the mth day in the future, m = 1, 2, ..., M, which has standardized soil moisture index SMI corresponding to H*W grids in the spatial dimension;

[0080] 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, and analyze the spatial centroid movement characteristics of severe drought events to predict the future spatiotemporal changes of drought.

[0081] The agricultural drought prediction method and system provided by the present invention that integrates spatiotemporal feature learning has the following advantages:

[0082] The present invention provides an agricultural drought prediction method and system integrating spatiotemporal feature learning, which comprehensively considers the temporal and spatial attributes of drought, uses high-precision grid meteorological data, and uses multiple hydrological and meteorological elements to input a trained convolutional neural network and a deep learning coupling model to predict future drought indicators, estimate the spatiotemporal variation characteristics of future drought disasters in the region, fully consider the multidimensional attributes of drought and the important meteorological and hydrological elements that affect drought, and efficiently warn of drought while keeping the model complexity within an appropriate range, providing technical guidance for drought prevention and control and water resources management. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 An overall diagram of an agricultural drought prediction method integrating spatiotemporal feature learning provided by the present invention;

[0084] Figure 2 A structural diagram of the drought prediction model provided by the present invention;

[0085] Figure 3 A diagram of the Standardized Soil Moisture Index (SMI) provided by an embodiment of the present invention;

[0086] Figure 4 This is a diagram showing the prediction results of drought events in the next seven days by the drought event recognition and prediction model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. 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.

[0088] In view of the problem that current drought prediction only considers the drought occurrence process and ignores the spatial variation of drought, a method and system for agricultural drought prediction that integrates spatiotemporal feature learning is proposed, which fully considers the temporal attributes and spatial characteristics of drought. In view of the problem that current drought prediction mostly uses a single variable for prediction, the standardized index of multiple meteorological and hydrological elements that affect drought is used as the model input, which fully considers the impact of multiple variables and reduces the redundant information of the original data, so as to efficiently predict drought.

[0089] Therefore, the agricultural drought prediction method and system provided by the present invention, which integrates spatiotemporal feature learning, comprehensively considers the temporal and spatial attributes of drought, uses high-precision grid meteorological data, and uses multiple hydrological and meteorological elements to input the trained convolutional neural network and deep learning coupling model to predict future drought indicators, estimate the spatiotemporal variation characteristics of regional drought disasters in the future, fully considers the multidimensional attributes of drought and the important meteorological and hydrological elements that affect drought, and effectively warns of drought while keeping the model complexity within an appropriate range, providing technical guidance for drought prevention and water resources management. The present invention has clear ideas, convenient operation, and strong practicality.

[0090] See also Figure 1 The present invention provides an agricultural drought prediction method integrating spatiotemporal feature learning, comprising the following steps:

[0091] Step S1, data collection and preprocessing:

[0092] Step S1.1, the prediction area is a gridded area of ​​H*W, where H and W represent the number of grids in the height and width directions of the prediction area respectively;

[0093] Step S1.2, collect ERA5-Land data products and hydrological station data in the prediction area, and perform spatial downscaling to obtain the grid-by-grid time series of meteorological and hydrological elements in the prediction area in the most recent N days.

[0094] Among them, ERA5-Land is a high-resolution global ground reanalysis data product provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), which provides higher spatial resolution and is more suitable for regional-scale applications and research. Remote sensing meteorological data products such as ERA5-Land data products can be used, or regional hydrological station data can be used for interpolation processing to ultimately obtain a high spatial resolution meteorological raster data sequence. Specifically, the hydrological station data can be rasterized and spatially downscaled using spatial interpolation methods, such as bilinear interpolation and Kriging interpolation.

[0095] Obtain the time series of meteorological and hydrological elements in the forecast area in the recent N days, and perform standardization to obtain the standardized index time series of meteorological and hydrological elements, expressed as: Set X = (X1, X2, ..., X N ), where X1, X2, ..., X N , respectively represent the standardized indexes of meteorological and hydrological elements of the forecast area on historical day 1, historical day 2, …, historical day N in chronological order; for X k , k = 1, 2, ..., N, is the standardized index of meteorological and hydrological elements in the forecast area on historical day k; X kIn the spatial dimension, there are H*W grid data, each grid data is C = 5-dimensional elements, respectively: Standardized maximum temperature index ST max , Standardized minimum temperature index ST min , Standard Precipitation Index SPI (Standard Precipitation Index), Standard Runoff Index SRI (Standard Runoff Index) and Standard Soil Moisture Index SMI (Standard Soil Moisture Index); C represents the dimension of the raster data; Figure 3 , which is a distribution map of the standardized soil moisture index SMI of a region on May 2, 2016 provided by an embodiment.

[0096] In this step, formula (1) is used for standardization:

[0097]

[0098] in, It represents the normalized index of the cth meteorological and hydrological element of the grid in the hth row and the wth column of the forecast area on the historical day k; where h = 1, 2, ..., H; w = 1, 2, ..., W; c = 1, 2, ..., C;

[0099] μ represents the average value of the standardized index of the cth meteorological and hydrological element in the history of N days, and σ represents the standard deviation of the standardized index of the cth meteorological and hydrological element in the history of N days;

[0100] represent The value before normalization.

[0101] Step S2, constructing a drought prediction model; 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 coupling model;

[0102] Step S2.1, set X = (X1, X2, ..., X N ) as input data, and input it into the spatial feature extraction module, which sequentially extracts the standardized index X of meteorological and hydrological elements of each historical day k. k Perform spatial feature extraction to extract the spatial feature vector SF corresponding to each historical day k k , forming a spatial feature matrix SF = (SF1, SF2, ..., SF N );

[0103] Step S2.2, transform the spatial feature matrix SF = (SF1, SF2, ..., SF N ) is input into the spatiotemporal feature learning module, which sequentially calculates the spatial feature vector SF of each historical day k. k Perform spatiotemporal feature learning to obtain the spatiotemporal feature vector TSF corresponding to each historical day k k , forming the space-time eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N );

[0104] Step S2.3, transform the spatiotemporal eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N ) is input to the sequence prediction module, and the sequence prediction module 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 Represents the predicted standardized soil moisture index SMI for the mth day in the future, where m = 1, 2, ..., M, and has standardized soil moisture index SMI corresponding to H*W grids in the spatial dimension.

[0105] The following describes steps S2.1 to S2.3 of the present invention in detail:

[0106] The process of spatial feature extraction by the spatial feature extraction module in step S2.1 is as follows:

[0107] Step S2.1.1, the spatial feature extraction module includes a Conv2D convolution layer, a BatchNormalization normalization layer and a PReLu activation layer;

[0108] Step S2.1.2, the standardized index X of meteorological and hydrological elements on historical day k k is the data matrix of H*W*C;

[0109] The standardized index X of meteorological and hydrological elements of historical day k k The data is input to the Conv2D convolution layer configured with a 3×3 convolution kernel. The local features are captured on the input data matrix by sliding the convolution kernel. The spatial feature vector with spatial correlation is extracted using formula (2):

[0110]

[0111] Among them, K i,j,c Represents the weight of the convolution kernel;

[0112] The standardized index X of meteorological and hydrological elements representing historical day k k , the value of the standardized index of the cth meteorological and hydrological element of the grid in the h+ith row and the w+jth column; c = 1, 2, ..., C;

[0113] represents the spatial characteristic vector of the standardized index of the cth meteorological and hydrological element of the grid in the hth row and wth column of the forecast area on the historical day k;

[0114] The Conv2D convolution layer can use the Conv2D layer wrapped by TimeDistributed, use the convolution kernel size of (3, 3), and use the same padding method to ensure that the spatial dimension of the output feature vector is consistent with the input.

[0115] Step S2.1.3, superimpose the spatial characteristic vectors of the C meteorological and hydrological element standardized indexes of the grid of the hth row and the wth column of the prediction area on the historical day k to obtain the spatial characteristic vector of the grid of the hth row and the wth column of the prediction area on the historical day k

[0116] Step S2.1.4, the BatchNormalization normalization layer, for each h-th row and w-th column of the prediction area of ​​historical day k, the spatial feature vector Normalization is performed to enhance the stability of the model and accelerate the training process, and then activated through the PReLu activation layer to obtain the spatial feature vector SF of the historical day k k .

[0117] The process of spatiotemporal feature extraction by the spatiotemporal feature learning module in step S2.2 is as follows:

[0118] Step S2.2.1, the spatiotemporal feature learning module includes C layers of ConvLSTM2D convolutional layers; for the c-th layer of ConvLSTM2D convolutional layer, c = 1, 2, ..., C, the convolution operation is performed using formula (3):

[0119] output(TSF k,c )=f(conv(Input(SF k,c ),W conv,c ,b c )) (3)

[0120] Among them: Input(SF k,c ) is the spatial feature vector of the standardized index of the cth meteorological and hydrological element of the input prediction area on the historical day k; f represents the activation function; conv represents the convolution operation; W conv,c represents the convolution kernel of the cth layer; b c is the bias term of the cth layer;

[0121] output(TSF k,c ) is the spatiotemporal characteristic vector of the standardized index of the cth meteorological and hydrological element of the input prediction area on the historical day k;

[0122] As an implementation, the spatiotemporal feature learning module includes C stacked ConvLSTM2D convolutional layers, each of which combines the spatial feature extraction capability and time series processing capability of the convolution operation to capture complex patterns and long-term dependencies in spatiotemporal data. The convolution kernel size of all layers is (3, 3), and the same padding is used to maintain the spatial dimension of the feature vector.

[0123] Step S2.2.2, thus obtaining the spatiotemporal characteristic vectors of the standardized index of meteorological and hydrological elements in the C prediction areas on historical day k, respectively: TSF k,1 ,TSF k,2 ,...,TSF k,C ; for TSF k,1 ,TSF k,2 ,...,TSF k,C Superposition is performed to obtain the spatiotemporal feature vector TSF corresponding to historical day k k .

[0124] The process of the sequence prediction module in step S2.3 to predict the SMI for the next M consecutive days is as follows:

[0125] Step S2.3.1, the sequence prediction module includes M ConvLSTM2D convolutional layers and M Conv2D convolutional layers;

[0126] Step S2.3.2, transform the spatiotemporal eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N ) is input to the first ConvLSTM2D convolutional layer ConvLSTM2D1, and the spatiotemporal feature vector TSF of the N+1th day is obtained using formula (4): N+1 :

[0127] output(TSF N+1 )=ConvLSTM2D1(Input(TSF1,TSF2,...,TSF N ),hide1,cell1) (4)

[0128] Among them: hide1 and cell1 represent the hidden state and cell state of the first ConvLSTM2D convolutional layer ConvLSTM2D1 respectively;

[0129] The spatiotemporal feature vector TSF of day N+1N+1 Input to the first Conv2D convolution layer and output the standardized soil moisture index SMI on the N+1th day N+1 ;

[0130] Step S2.3.3, transform the spatiotemporal eigenvector matrix TSF2,...,TSF N ,TSF N+1 Input to the second ConvLSTM2D convolutional layer ConvLSTM2D2, and use formula (5) to obtain the spatiotemporal feature vector TSF of the N+2th day N+2 :

[0131] output(TSF N+2 )=ConvLSTM2D2(Input(TSF2,...,TSF N ,TSF N+1 ),hide2,cell2)(5)

[0132] Among them: hide2 and cell2 represent the hidden state and cell state of the second ConvLSTM2D convolutional layer ConvLSTM2D2 respectively;

[0133] The spatiotemporal feature vector TSF of day N+2 N+2 Input to the second Conv2D convolution layer and output the standardized soil moisture index SMI on the N+2th day N+2 ;

[0134] Step S2.3.4, keep moving back one position and repeating the cycle, so as to obtain the standardized soil moisture index SMI sequence SMI=(SMI N+1 ,SMI N+2 ,...,SMI N+M ).

[0135] Step S3, constructing a drought event identification and prediction model; the drought event identification and prediction model analyzes the standardized soil moisture index SMI sequence of the next M consecutive days predicted by the sequence prediction module, extracts drought events, and analyzes the spatial centroid movement characteristics of severe drought events to predict the future spatiotemporal changes of drought.

[0136] Step S3 is specifically as follows:

[0137] Step S3.1, for each grid in the prediction area, obtain the standardized soil moisture index SMI sequence for the past N days and the next M consecutive days, thus obtaining N+M standardized soil moisture index SMIs, expressed as: sequence SMI N+M =(SMI1, SMI2, ..., SMI N ,SMIN+1 ,SMI N+2 ,...,SMI N+M );

[0138] For example, the standardized soil moisture index (SMI) series for the past 14 days and the next 7 consecutive days can be obtained.

[0139] Step S3.2, for sequence SMI N+M Perform trend analysis on time-varying characteristics to determine whether the trend is significant. If so, the sequence SMI N+M Assign a value to the corresponding grid, otherwise, the grid is assigned a value of 0;

[0140] In step S3.2, the following method is used to determine whether there is a significant trend:

[0141] Step S3.2.1, using formula (6), calculate the sequence SMI N+M =(SMI1, SMI2, ..., SMI N ,SMI N+1 ,SMI N+2 ,...,SMI N+M ) of the cumulative difference d:

[0142]

[0143] 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 each SMI data in , d greater than 0 means that the SMI data sequence tends to become larger, otherwise it becomes smaller; sign() represents the sign function;

[0144] Step S3.2.2, using formula (7), obtain the concentration Var(d) of the cumulative difference d:

[0145]

[0146] in:

[0147] R stands for SMI1, SMI2, ..., SMI N ,SMI N+1 ,SMI N+2 ,...,SMI N+M There are R different SMIs that appear at least twice in total; for the rth repeated SMI, its number of repetitions is Num r ;

[0148] For the sake of convenience, here we only take the sequence of 7 SMI as an example to introduce Num r And the meaning of R:

[0149] Assume that the values ​​of SMI1, SMI2, ..., SMI7 are 1.1, 1.2, 1.3, 1.1, 1.1, 1.2, 1.4 respectively. Then there are two different SMIs that appear repeatedly. For the first repeated SMI, its value is 1.1, and the number of repeated occurrences is 3 times. For the second repeated SMI, its value is 1.2, and the number of repeated occurrences is 2 times.

[0150] Step S3.2.3, using formula (8), obtain the statistic St:

[0151]

[0152] Step S3.2.4, if the statistic St exceeds the set threshold, it means that the sequence SMI N+M There is a significant trend; otherwise, it represents the sequence SMI N+M There is no significant trend.

[0153] As an implementation, if St is greater than 1.65, it means that the trend significance passes the 90% confidence level, which means that the sequence SMI N+M There is a significant trend; otherwise, the sequence SMI N+M There is no significant trend.

[0154] Step S3.3, analyze the distribution of the standardized soil moisture index SMI of the forecast area in the recent historical M days and the next M consecutive days on a daily basis, determine the centroid center of the severe drought area on a daily basis, and thus obtain the moving characteristics of the centroid center of the severe drought area in the recent historical M days and the next M consecutive days.

[0155] In step S3.3, the method for determining the centroid of the daily severe drought area is:

[0156] Step S3.3.1, taking each of the recent M historical days and the future M consecutive days as an identification day, and reading the standardized soil moisture index SMI of each grid in the prediction area on each identification day;

[0157] For example, drought events are identified for the most recent 7 days in history and for the next 7 consecutive days.

[0158] Step S3.3.2, identify whether a drought event occurs in each grid on the identified day:

[0159] The standardized soil moisture index (SMI) of each grid 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 less, and it is recognized that a drought event has occurred in the grid on the identification day, and its SMI value is used as the drought intensity of the grid on the identification day; conversely, if the SMI is greater than or equal to 0, it has a positive variation, indicating that the soil moisture is more, no drought event has occurred, and the grid does not have a drought intensity on the identification day;

[0160] Step S3.3.3, determine the severe drought area in the forecast area on each identified day:

[0161] The prediction area is traversed with a 5×5 window. For each window area, the weighted average of the drought intensity of each grid in the window area on the identification day is calculated to obtain the mean drought intensity corresponding to the window area.

[0162] Determine the window area with the largest mean drought intensity, which is the severe drought area in the forecast area on the identification day;

[0163] Step S3.3.4, using formula (9) and formula (10), obtain the coordinates of the centroid center of the severe drought area in the forecast area on the identification day:

[0164]

[0165] Where: C lon and C lat , respectively representing the longitude and latitude of the center of mass center; Lon z and Lat z , respectively represent the longitude and latitude of the zth grid in the severe drought area.

[0166] As an example, Figure 4 As shown in the figure, it is a distribution cloud map of the standardized soil moisture index SMI of each day in the recent 7 days and the predicted 7 days in the future identified by the model of the present invention; wherein, the recent 7 days are: July 16, 2015 to July 22, 2015. The next 7 days are: July 23, 2015 to July 29, 2015. In addition, in the distribution cloud map of the standardized soil moisture index SMI of each day in the next 7 days, black dots are further used to represent the identified drought events.

[0167] The method of the present invention is used to further calculate the drought events for each of the next 7 days, and the centroid prediction results for the next 7 days are shown in Table 1:

[0168] Table 1: Centroid prediction results for the next 7 days

[0169]

[0170] After testing, the predicted centroid position for the next 7 days has a deviation from the actual location of drought in the predicted area that is less than the set value, which meets the prediction accuracy requirements and can accurately reflect the occurrence and propagation process of drought in the space.

[0171] The present invention also provides an agricultural drought prediction system integrating spatiotemporal feature learning, comprising:

[0172] The data collection and preprocessing module is used to obtain the time series of meteorological and hydrological elements in the forecast area in the recent N days, and perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, which is expressed as: set X = (X1, X2, ..., X N ), where the prediction area is a gridded area of ​​H*W; where H and W represent the number of grids in the height and width directions of the prediction area, respectively; X1, X2, ..., X N , respectively represent the standardized indexes of meteorological and hydrological elements of the forecast area on historical day 1, historical day 2, …, historical day N in chronological order; for X k , k = 1, 2, ..., N, is the standardized index of meteorological and hydrological elements in the forecast area on historical day k; X k In the spatial dimension, there are H*W grid data, each grid data is C = 5-dimensional elements, respectively: 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;

[0173] Drought prediction model, including spatial feature extraction module, spatiotemporal feature learning module and sequence prediction module;

[0174] The spatial feature extraction module is used to extract the set X=(X1, X2, ..., X N ) as input data, and then the standardized index X of meteorological and hydrological elements of each historical day k is calculated. k Perform spatial feature extraction to extract the spatial feature vector SF corresponding to each historical day k k , forming a spatial feature matrix SF = (SF1, SF2, ..., SF N );

[0175] The spatiotemporal feature learning module is used to input the spatial feature matrix SF = (SF1, SF2, ..., SF N ), and then the spatial feature vector SF of each historical day k is k Perform spatiotemporal feature learning to obtain the spatiotemporal feature vector TSF corresponding to each historical day k k, forming the space-time eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N );

[0176] Sequence prediction module, used to input spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF N ), the standardized soil moisture index SMI sequence for the next M consecutive days is predicted, expressed as: SMI = (SMI N+1 ,SMI N+2 ,...,SMI N+M ); where M<N, SMI N+m Represents the predicted standardized soil moisture index SMI for the mth day in the future, m = 1, 2, ..., M, which has standardized soil moisture index SMI corresponding to H*W grids in the spatial dimension;

[0177] 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, and analyze the spatial centroid movement characteristics of severe drought events to predict the future spatiotemporal changes of drought.

[0178] The agricultural drought prediction method and system provided by the present invention that integrates spatiotemporal feature learning has the following advantages:

[0179] The present invention relates to an agricultural drought prediction method and system integrating spatiotemporal feature learning, which adopts high-precision grid meteorological and hydrological data sequences to construct a coupled machine learning model capable of extracting spatiotemporal features of drought indicators, and can fully capture the temporal information and spatial features of drought while controlling the complexity of the model; fully considers the meteorological and hydrological elements related to drought, and adopts standardized indexes to predict future drought indicators, which can not only extract the long-term trend of meteorological and hydrological elements, but also respond to short-term extreme events and hydrological changes, reduce the redundant information of original data, and efficiently predict the occurrence and propagation process of drought areas in space, so as to provide technical guidance for drought prevention and water resources management. The present invention has clear ideas, convenient operation and strong practicality.

[0180] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. An agricultural drought prediction method integrating spatiotemporal feature learning, characterized in that: The following steps are involved: Step S1, data collection and preprocessing: Step S1.1, the prediction area is a gridded area of ​​H*W, where H and W represent the number of grids in the height and width directions of the prediction area respectively; Step S1.2, obtain the time series of meteorological and hydrological elements in the forecast area in the recent N days, and perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, which is expressed as: set X = (X1, X2, ..., X N ), where X1, X2, ..., X N , respectively represent the standardized indexes of meteorological and hydrological elements of the forecast area on historical day 1, historical day 2, …, historical day N in chronological order; for X k , k = 1, 2, ..., N, is the standardized index of meteorological and hydrological elements in the forecast area on historical day k; X k In the spatial dimension, there are H*W grid data, each grid data is C = 5-dimensional elements, respectively: 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; Step S2, constructing 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 X = (X1, X2, ..., X N ) as input data, and input it into the spatial feature extraction module, which sequentially extracts the standardized index X of meteorological and hydrological elements of each historical day k. k Perform spatial feature extraction to extract the spatial feature vector SF corresponding to each historical day k k , forming a spatial feature matrix SF = (SF1, SF2, ..., SF N ); Step S2.2, transform the spatial feature matrix SF = (SF1, SF2, ..., SF N ) is input into the spatiotemporal feature learning module, which sequentially calculates the spatial feature vector SF of each historical day k. k Perform spatiotemporal feature learning to obtain the spatiotemporal feature vector TSF corresponding to each historical day k k , forming the space-time eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N ); Step S2.3, transform the spatiotemporal eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N ) is input to the sequence prediction module, and the sequence prediction module 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 Represents the predicted standardized soil moisture index SMI for the mth day in the future, m = 1, 2, ..., M, which has standardized soil moisture index SMI corresponding to H*W grids in the spatial dimension; Step S3, constructing a drought event identification and prediction model; the drought event identification and prediction model analyzes the standardized soil moisture index SMI sequence of the next M consecutive days predicted by the sequence prediction module, extracts drought events, and analyzes the spatial centroid movement characteristics of severe drought events to predict the future spatiotemporal changes of drought.

2. The agricultural drought prediction method integrating spatiotemporal feature learning according to claim 1 is characterized in that: Before step S1.2, it also includes: The ERA5-Land data products and hydrological station data of the forecast area are collected and spatially downscaled to obtain the grid-by-grid time series of meteorological and hydrological elements in the forecast area in the most recent N days.

3. The agricultural drought prediction method integrating spatiotemporal feature learning according to claim 1 is characterized in that: In step S1.2, formula (1) is used for standardization: in, It represents the normalized index of the cth meteorological and hydrological element of the grid in the hth row and the wth column of the forecast area on the historical day k; where h = 1, 2, ..., H; w = 1, 2, ..., W; c = 1, 2, ..., C; μ represents the average value of the standardized index of the cth meteorological and hydrological element in the history of N days, and σ represents the standard deviation of the standardized index of the cth meteorological and hydrological element in the history of N days; represent The value before normalization.

4. The agricultural drought prediction method integrating spatiotemporal feature learning according to claim 1 is characterized in that: Step S2.1 is specifically as follows: Step S2.1.1, the spatial feature extraction module includes a Conv2D convolution layer, a BatchNormalization normalization layer and a PReLu activation layer; Step S2.1.2, the standardized index X of meteorological and hydrological elements on historical day k k is the data matrix of H*W*C; The standardized index X of meteorological and hydrological elements of historical day k k The data is input to the Conv2D convolution layer configured with a 3×3 convolution kernel. The local features are captured on the input data matrix by sliding the convolution kernel. The spatial feature vector with spatial correlation is extracted using formula (2): Among them, K i,j,c Represents the weight of the convolution kernel; The standardized index X of meteorological and hydrological elements representing historical day k k , the value of the standardized index of the cth meteorological and hydrological element of the grid in the h+ith row and the w+jth column; c = 1, 2, ..., C; represents the spatial characteristic vector of the standardized index of the cth meteorological and hydrological element of the grid in the hth row and wth column of the forecast area on the historical day k; Step S2.1.3, superimpose the spatial characteristic vectors of the C meteorological and hydrological element standardized indexes of the grid of the hth row and the wth column of the prediction area on the historical day k to obtain the spatial characteristic vector of the grid of the hth row and the wth column of the prediction area on the historical day k Step S2.1.4, the BatchNormalization normalization layer, for each h-th row and w-th column of the prediction area of ​​historical day k, the spatial feature vector After normalization, it is activated through the PReLu activation layer to obtain the spatial feature vector SF of historical day k k .

5. The agricultural drought prediction method integrating spatiotemporal feature learning according to claim 1 is characterized in that: Step S2.2 is specifically as follows: Step S2.2.1, the spatiotemporal feature learning module includes C layers of ConvLSTM2D convolutional layers; for the c-th layer of ConvLSTM2D convolutional layer, c = 1, 2, ..., C, the convolution operation is performed using formula (3): output(TSF k,c )=f(conv(Input(SF k,c ),W conv,c ,b c )) (3) Among them: Input(SF k,c ) is the spatial feature vector of the standardized index of the cth meteorological and hydrological element of the input prediction area on the historical day k; f represents the activation function; conv represents the convolution operation; W conv,c represents the convolution kernel of the cth layer; b c is the bias term of the cth layer; output(TSF k,c ) is the spatiotemporal characteristic vector of the standardized index of the cth meteorological and hydrological element of the input prediction area on the historical day k; Step S2.2.2, thus obtaining the spatiotemporal characteristic vectors of the standardized index of meteorological and hydrological elements in the C prediction areas on historical day k, respectively: TSF k,1 ,TSF k,2 ,...,TSF k,C ; for TSF k,1 ,TSF k,2 ,...,TSF k,C Superposition is performed to obtain the spatiotemporal feature vector TSF corresponding to historical day k k .

6. The agricultural drought prediction method integrating spatiotemporal feature learning according to claim 1 is characterized in that: Step S2.3 is specifically as follows: Step S2.3.1, the sequence prediction module includes M ConvLSTM2D convolutional layers and M Conv2D convolutional layers; Step S2.3.2, transform the spatiotemporal eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N ) is input to the first ConvLSTM2D convolutional layer ConvLSTM2D1, and the spatiotemporal feature vector TSF of the N+1th day is obtained using formula (4): N+1 : output(TSF N+1 )=ConvLSTM2D1(Input(TSF1,TSF2,...,TSF N ),hide1,cell1) (4) Among them: hide1 and cell1 represent the hidden state and cell state of the first ConvLSTM2D convolutional layer ConvLSTM2D1 respectively; The spatiotemporal feature vector TSF of day N+1 N+1 Input to the first Conv2D convolution layer and output the standardized soil moisture index SMI on the N+1th day N+1 ; Step S2.3.3, transform the spatiotemporal eigenvector matrix TSF2,...,TSF N ,TSF N+1 Input to the second ConvLSTM2D convolutional layer ConvLSTM2D2, and use formula (5) to obtain the spatiotemporal feature vector TSF of the N+2th day N+2 : output(TSF N+2 )=ConvLSTM2D2(Input(TSF2,...,TSF N ,TSF N+1 ),hide2,cell2) (5) Among them: hide2 and cell2 represent the hidden state and cell state of the second ConvLSTM2D convolutional layer ConvLSTM2D2 respectively; The spatiotemporal feature vector TSF of day N+2 N+2 Input to the second Conv2D convolution layer and output the standardized soil moisture index SMI on the N+2th day N+2 ; Step S2.3.4, keep moving back one position and repeating the cycle, so as to obtain the standardized soil moisture index SMI sequence SMI=(SMI N+1 ,SMI N+2 ,...,SMI N+M ).

7. The agricultural drought prediction method integrating spatiotemporal feature learning according to claim 1 is characterized in that: Step S3 is specifically as follows: Step S3.1, for each grid in the prediction area, obtain the standardized soil moisture index SMI sequence for the past N days and the next M consecutive days, thus obtaining N+M standardized soil moisture index SMIs, expressed as: sequence SMI N+M =(SMI1, SMI2, ..., SMI N ,SMI N+1 ,SMI N+2 ,...,SMI N+M ); Step S3.2, for sequence SMI N+M Perform trend analysis on time-varying characteristics to determine whether the trend is significant. If so, the sequence SMI N+M Assign a value to the corresponding grid, otherwise, the grid is assigned a value of 0; Step S3.3, analyze the distribution of the standardized soil moisture index SMI of the forecast area in the recent historical M days and the next M consecutive days on a daily basis, determine the centroid center of the severe drought area on a daily basis, and thus obtain the moving characteristics of the centroid center of the severe drought area in the recent historical M days and the next M consecutive days.

8. The agricultural drought prediction method integrating spatiotemporal feature learning according to claim 7 is characterized in that: Step S3.2 is specifically as follows: Step S3.2.1, using formula (6), calculate the sequence SMI N+M =(SMI1, SMI2, ..., SMI N ,SMI N+1 ,SMI N+2 ,...,SMI N+M ) of the cumulative difference d: 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 each SMI data in , d greater than 0 means that the SMI data sequence tends to become larger, otherwise it becomes smaller; sign() represents the sign function; Step S3.2.2, using formula (7), obtain the concentration Var(d) of the cumulative difference d: in: R stands for SMI1, SMI2, ..., SMI N ,SMI N+1 ,SMI N+2 ,...,SMI N+M There are R different SMIs that appear at least twice in total; for the rth repeated SMI, its number of repetitions is Num r ; Step S3.2.3, using formula (8), obtain the statistic St: Step S3.2.4, if the statistic St exceeds the set threshold, it means that the sequence SMI N+M There is a significant trend; otherwise, it represents the sequence SMI N+M There is no significant trend.

9. The agricultural drought prediction method integrating spatiotemporal feature learning according to claim 7 is characterized in that: In step S3.3, the method for determining the centroid of the daily severe drought area is: Step S3.3.1, taking each of the recent M historical days and the future M consecutive days as an identification day, and reading the standardized soil moisture index SMI of each grid in the prediction area on each identification day; Step S3.3.2, identify whether a drought event occurs in each grid on the identified day: The standardized soil moisture index (SMI) of each grid 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. It is recognized that the grid has a drought event on the identification day, and its SMI value is used as the drought intensity of the grid on the identification day. On the contrary, if the SMI is greater than or equal to 0, it has positive variation, indicating that the soil moisture is high, no drought event has occurred, and the grid does not have drought intensity on the identification day; Step S3.3.3, determine the severe drought area in the forecast area on each identified day: The prediction area is traversed with a 5×5 window. For each window area, the weighted average of the drought intensity of each grid in the window area on the identification day is calculated to obtain the mean drought intensity corresponding to the window area. Determine the window area with the largest mean drought intensity, which is the severe drought area in the forecast area on the identification day; Step S3.3.4, using formula (9) and formula (10), obtain the coordinates of the centroid center of the severe drought area in the forecast area on the identification day: Where: C lon and C lat , respectively representing the longitude and latitude of the center of mass center; Lon z and Lat z , respectively represent the longitude and latitude of the zth grid in the severe drought area.

10. A system for implementing the agricultural drought prediction method integrating spatiotemporal feature learning as described in any one of claims 1 to 9, characterized in that: include: The data collection and preprocessing module is used to obtain the time series of meteorological and hydrological elements in the forecast area in the recent N days, and perform standardization processing to obtain the standardized index time series of meteorological and hydrological elements, which is expressed as: set X = (X1, X2, ..., X N ), where the prediction area is a gridded area of ​​H*W; where H and W represent the number of grids in the height and width directions of the prediction area, respectively; X1, X2, ..., X N , respectively represent the standardized indexes of meteorological and hydrological elements of the forecast area on historical day 1, historical day 2, …, historical day N in chronological order; for X k , k = 1, 2, ..., N, is the standardized index of meteorological and hydrological elements in the forecast area on historical day k; X k In the spatial dimension, there are H*W grid data, each grid data is C = 5-dimensional elements, respectively: 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; Drought prediction model, including spatial feature extraction module, spatiotemporal feature learning module and sequence prediction module; The spatial feature extraction module is used to extract the set X=(X1, X2, ..., X N ) as input data, and then the standardized index X of meteorological and hydrological elements of each historical day k is calculated. k Perform spatial feature extraction to extract the spatial feature vector SF corresponding to each historical day k k , forming a spatial feature matrix SF = (SF1, SF2, ..., SF N ); The spatiotemporal feature learning module is used to input the spatial feature matrix SF = (SF1, SF2, ..., SF N ), and then the spatial feature vector SF of each historical day k is k Perform spatiotemporal feature learning to obtain the spatiotemporal feature vector TSF corresponding to each historical day k k , forming the space-time eigenvector matrix TSF = (TSF1, TSF2, ..., TSF N ); Sequence prediction module, used to input spatiotemporal feature vector matrix TSF = (TSF1, TSF2, ..., TSF N ), the standardized soil moisture index SMI sequence for the next M consecutive days is predicted, expressed as: SMI = (SMI N+1 ,SMI N+2 ,...,SMI N+M ); where M<N, SMI N+m Represents the predicted standardized soil moisture index SMI for the mth day in the future, m = 1, 2, ..., M, which has standardized soil moisture index SMI corresponding to H*W grids in the spatial dimension; 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, and analyze the spatial centroid movement characteristics of severe drought events to predict the future spatiotemporal changes of drought.

Images