Black river basin drought monitoring method based on space-time convolutional network

Abstract

The invention relates to a black river basin drought monitoring method based on a space-time convolutional network, and the method comprises the steps: obtaining multi-source data of a black river basin from satellite remote sensing, a meteorological station and a ground observation point, and constructing a space-time data set; based on the spatio-temporal data set, capturing a dynamic evolution trend of a drought event by using a drought feature extraction model based on a long and short term memory network, and obtaining a time sequence feature vector; performing feature extraction on the data in the spatial dimension of the spatio-temporal data set by adopting a convolutional neural network, analyzing the spatial distribution rule of drought in the black river basin, and obtaining a spatial feature mapping graph; carrying out feature fusion on the time sequence feature vector and the spatial feature map, and constructing a space-time joint feature matrix; and monitoring the drought of the black river basin based on the spatio-temporal joint feature matrix. According to the method, continuous innovation and development of a drought monitoring technology are promoted, the overall level of drought monitoring is improved, and references are provided for drought monitoring of other regions.

Description

Technical Field

[0001] The present invention relates to the technical field of drought monitoring, and in particular to a drought monitoring method for the Heihe River Basin based on a spatiotemporal convolutional network. Background Art

[0002] Drought monitoring, a crucial aspect of agricultural production and water resources management, is crucial for ensuring regional ecological security and economic development. Accurately monitoring the onset and evolution of drought is particularly crucial in regions like the Heihe River Basin, where water resources are unevenly distributed and the ecological environment is fragile. Drought not only impacts crop growth but can also lead to water shortages and ecological degradation. Therefore, efficient and accurate monitoring methods are urgently needed to support decision-making and early warning. However, current drought monitoring methods often have limitations. Many traditional methods rely on single data sources or simple statistical analyses, which fail to fully capture the complex temporal and spatial variations of drought. These approaches often overlook the dynamic evolution of drought events, resulting in insufficient accuracy and timeliness of monitoring results, making them inadequate to meet the practical needs of complex environments. Against this backdrop, drought monitoring faces significant technical challenges. Primarily, the effective integration of multi-source data, such as information from satellite remote sensing and meteorological stations, presents significant disparities in format, temporal scale, and spatial resolution, making it difficult to develop a unified analytical framework. This difficulty in data integration further hinders the simultaneous consideration of both temporal continuity and spatial distribution when capturing drought events. This fragmentation of spatiotemporal information makes it difficult for monitoring methods to accurately depict the onset, development, and resolution of droughts, which in turn affects the timeliness and reliability of early warnings. Therefore, building an analytical framework based on multi-source data that can simultaneously process both time series and spatial distribution characteristics has become a key issue in improving the accuracy of drought monitoring in the Heihe River Basin. Summary of the Invention

[0003] The purpose of the present invention is to propose a drought monitoring method for the Heihe River Basin based on spatiotemporal convolutional networks to solve the problems existing in the above-mentioned prior art.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A drought monitoring method for the Heihe River Basin based on spatiotemporal convolutional networks, including:

[0006] A spatiotemporal dataset was constructed by acquiring multi-source data of the Heihe River Basin from satellite remote sensing, meteorological stations, and ground observation points;

[0007] Based on the spatiotemporal dataset, a drought feature extraction model based on a long short-term memory network is used to capture the dynamic evolution trend of drought events and obtain time series feature vectors;

[0008] A convolutional neural network was used to extract features from the spatial dimension of the spatiotemporal dataset, analyze the spatial distribution of drought in the Heihe River Basin, and obtain a spatial feature map.

[0009] Performing feature fusion on the time series feature vector and the spatial feature map to construct a spatiotemporal joint feature matrix; wherein the spatiotemporal joint feature matrix is ​​used to reflect the comprehensive variation characteristics of drought events in time and space and determine the evolution pattern of drought events;

[0010] Based on the spatiotemporal joint feature matrix, drought monitoring in the Heihe River Basin is carried out.

[0011] Optionally, multi-source data for the Heihe River Basin can be obtained from satellite remote sensing, meteorological stations, and ground observation points, including:

[0012] NDVI data of preset spatial resolution are obtained from satellites, daily rainfall data are obtained from meteorological stations, and soil moisture data are obtained from ground observation points.

[0013] Optionally, building a spatiotemporal dataset includes:

[0014] Data preprocessing methods are used to standardize the format and time align data from different sources, address differences in spatial resolution, and generate data in a unified format;

[0015] For data in a unified format, spatial interpolation technology is used to improve the resolution of data with a spatial resolution lower than a preset resolution threshold. At the same time, a time series smoothing algorithm is combined to fill in the missing data values ​​in the time dimension to obtain the spatiotemporal dataset.

[0016] Optionally, obtaining the time series feature vector includes:

[0017] Based on the spatiotemporal dataset, continuous data in the time dimension is obtained, data segments related to drought events are extracted, and the data are arranged in chronological order using a segmented processing method to obtain an initial time series dataset;

[0018] Based on the pre-trained drought feature extraction model, the dynamic evolution information in the continuous data is obtained for the initial time series data set, the dependency relationship between the data is captured through the memory mechanism of the network, and the evolution characteristics of the drought event are determined;

[0019] According to the captured evolution characteristics of drought events, the time dimension characteristic value related to trend capture is extracted. If the time dimension characteristic value exceeds a preset threshold, the data of the time period exceeding the threshold is weighted to obtain the time series characteristic vector.

[0020] Optionally, obtaining the spatial feature map includes:

[0021] Use pre-trained convolutional neural networks to extract features from the spatial dimension of spatiotemporal datasets;

[0022] Based on the extracted features, cluster analysis was used to analyze the spatial distribution of drought in the Heihe River Basin, and the high-incidence areas, concentrated areas and spatial distribution patterns of drought were identified.

[0023] The spatial distribution characteristics obtained by analysis are mapped to the geographical space of the Heihe River Basin to generate a spatial characteristic map.

[0024] Optionally, the drought feature extraction model based on the long short-term memory network includes:

[0025] The dual-channel input layer is used to receive the original time series data and the time series data after feature enhancement processing respectively;

[0026] A multi-scale time decomposition module is used to perform wavelet transform on the data of each channel and decompose it into high-frequency subsequences and low-frequency subsequences; wherein the high-frequency subsequence is used to reflect short-term change characteristics, and the low-frequency subsequence is used to reflect long-term trend characteristics;

[0027] A hybrid GRU-LSTM layer, in which the GRU layer receives high-frequency subsequence data and uses its short-term feature extraction capability to process short-term change features. The LSTM layer receives low-frequency subsequence data and uses its long-distance dependency capture capability to process long-term trend features. The output feature vectors of the GRU layer and the LSTM layer are concatenated to obtain a fusion of short-term and long-term feature vectors.

[0028] The attention mechanism enhancement layer receives the fused feature vector, calculates the attention weight of each time step, multiplies the attention weight by the integrated feature vector, and sums it along the time dimension to obtain the weighted feature vector;

[0029] The multi-task learning output layer is used to set up multi-task output nodes and flow the weighted feature vectors to each output node to predict different drought indicators.

[0030] Optionally, the convolutional neural network includes:

[0031] The input layer is used to accept data in the spatial dimension of the spatiotemporal dataset;

[0032] The multi-scale feature extraction module is used to set up multiple convolution branches, each branch uses convolution kernels of different sizes to extract features of different scales, and splices the output feature maps of each branch;

[0033] The attention mechanism module is used to calculate the importance weight of each spatial position in the feature map, multiply the spatial attention weight by the feature map, and obtain the weighted feature map.

[0034] Optionally, performing drought monitoring in the Heihe River Basin based on the spatiotemporal joint feature matrix includes:

[0035] Acquire eigenvalue sequence data from the spatiotemporal joint feature matrix, analyze the temporal features of each region one by one, and obtain preliminary eigenvalue distribution results;

[0036] According to the eigenvalue distribution results, if the eigenvalue sequence of a certain area exceeds the preset drought threshold, the area is judged to be in a drought warning state and the warning mark of the area is determined;

[0037] For areas with identified warning signs, the corresponding spatial feature data is extracted from the space-time matrix, and spatial mapping technology is used to generate a spatial feature map of the area to obtain preliminary visualization results of the drought impact range;

[0038] By conducting boundary analysis on the spatial feature map and combining the relationship between spatial features and the impact range, the specific boundaries of the drought impact range can be determined and quantitative data for range assessment can be obtained;

[0039] Based on the quantitative data of the scope assessment and combined with the changing trends of time characteristics, the persistence of the drought warning status is analyzed and the dynamic evolution of the drought impact is determined.

[0040] Optionally, the spatiotemporal joint feature matrix is ​​a spatial grid point of a preset dimension, each grid point corresponds to a time series, the time span is a preset number of days in the past, and the characteristic values ​​of each day include NDVI data, rainfall and soil moisture.

[0041] The beneficial effects of the present invention are:

[0042] By integrating multiple data sets, such as rainfall, soil moisture, and NDVI, this new system can comprehensively assess drought conditions from multiple dimensions, improving the accuracy and reliability of monitoring results. This system can more comprehensively reflect the onset, development, and severity of drought, providing a more scientific basis for drought response decision-making.

[0043] Combining this data with geospatial information to generate spatial feature maps and spatiotemporal joint feature matrices enables refined spatial monitoring of the Heihe River Basin. This can pinpoint the specific location, scope, and severity of drought, supporting the rational allocation of drought relief resources and targeted drought response efforts.

[0044] By constructing a comprehensive drought feature vector and setting drought warning thresholds, we can timely identify drought trends and provide early warnings. This provides the government and relevant departments with sufficient preparation time to develop drought mitigation plans in advance and reduce the risk of drought disasters.

[0045] This invention provides a new technical approach and method for drought monitoring, enriching the means of drought monitoring. It helps promote the continuous innovation and development of drought monitoring technology, improves the overall level of drought monitoring, and provides a reference and reference for drought monitoring in other regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 The figure is a flow chart of a drought monitoring method for the Heihe River Basin based on a spatiotemporal convolutional network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] like Figure 1 As shown, this embodiment proposes a drought monitoring method for the Heihe River Basin based on a spatiotemporal convolutional network, including:

[0051] A spatiotemporal dataset was constructed by acquiring multi-source data of the Heihe River Basin from satellite remote sensing, meteorological stations, and ground observation points;

[0052] Based on spatiotemporal datasets, a drought feature extraction model based on long short-term memory networks is used to capture the dynamic evolution trend of drought events and obtain time series feature vectors.

[0053] A convolutional neural network was used to extract features from the spatial dimension of the spatiotemporal dataset, analyze the spatial distribution of drought in the Heihe River Basin, and obtain a spatial feature map.

[0054] The time series feature vectors and spatial feature maps are fused to construct a spatiotemporal joint feature matrix. The spatiotemporal joint feature matrix is ​​used to reflect the comprehensive temporal and spatial variation characteristics of drought events and determine the evolution pattern of drought events.

[0055] Drought monitoring in the Heihe River Basin is carried out based on the spatiotemporal joint feature matrix.

[0056] Furthermore, multi-source data of the Heihe River Basin were obtained from satellite remote sensing, meteorological stations and ground observation points, including:

[0057] NDVI data of preset spatial resolution are obtained from satellites, daily rainfall data are obtained from meteorological stations, and soil moisture data are obtained from ground observation points.

[0058] Specifically, in this embodiment, in response to the drought monitoring needs in the Heihe River Basin, the monitoring goals are achieved through multi-source data fusion and processing. First, data is obtained from satellite remote sensing, meteorological stations and ground observation points. For example, NDVI data with a spatial resolution of 30 meters is obtained from the Landsat-8 satellite and is updated every 16 days; daily rainfall data is obtained from meteorological stations, which are unevenly distributed in space and the station spacing is about 50 kilometers; soil moisture data is obtained from ground observation points, which is updated daily, but the coverage is only for key areas in the basin.

[0059] Furthermore, constructing a spatiotemporal dataset includes:

[0060] Data preprocessing methods are used to standardize the format and time align data from different sources, address differences in spatial resolution, and generate data in a unified format;

[0061] For data in a unified format, spatial interpolation technology is used to improve the resolution of data with a spatial resolution lower than the preset resolution threshold. At the same time, the time series smoothing algorithm is combined to fill in the missing data values ​​in the time dimension to obtain a spatiotemporal dataset.

[0062] Specifically, in the present embodiment, in the data preprocessing stage, the data from different sources are formatted and standardized, such as the GeoTIFF format of satellite data, the CSV format of weather stations, and the TXT format of ground observation points are uniformly converted to NetCDF format, and Python's xarray library is used to read and convert to ensure that the data fields are consistent, such as timestamps are unified in UTC format with accuracy to seconds. Then time alignment is performed, and for the frequency of satellite data once every 16 days, it is aligned with the daily updated weather and soil data, and a linear interpolation algorithm is used to calculate the NDVI value of the middle date, with an error control within 0.05 to ensure time series integrity. Then the spatial resolution difference is processed, and the satellite 30-meter resolution data and the weather station 50-kilometer spacing data are spatially interpolated, and the Kriging interpolation method is used, based on the correlation weight calculation between sites, to generate a unified 1-kilometer resolution grid data, and the interpolation is verified to have an error standard deviation less than 0.1, ensuring accuracy. Finally, a multi-source data set in a unified format is generated, comprising three variables: NDVI, rainfall, and soil moisture.

[0063] A preliminary analysis of spatial resolution and time series values ​​was conducted to determine the distribution of areas with missing data and insufficient spatial precision. Based on the preliminary analysis results, spatial interpolation was used to increase the resolution of areas with lower spatial dimensions, resulting in an enhanced spatial data distribution. Based on this enhanced spatial data distribution and the missing data in the time dimension, a temporal smoothing method was used to fill in the time series values, yielding a padded time data series.

[0064] Furthermore, obtaining the time series feature vector includes:

[0065] Based on the spatiotemporal dataset, continuous data in the time dimension are obtained, data segments related to drought events are extracted, and the data are arranged in chronological order using a segmented processing method to obtain the initial time series dataset;

[0066] Based on the pre-trained drought feature extraction model, the dynamic evolution information in the continuous data is obtained for the initial time series dataset. The dependency relationship between the data is captured through the network's memory mechanism to determine the evolution characteristics of the drought event.

[0067] According to the captured evolution characteristics of drought events, the time dimension characteristic values ​​related to trend capture are extracted. If the time dimension characteristic value exceeds the preset threshold, the data of the time period exceeding the threshold are weighted to obtain the time series characteristic vector.

[0068] Specifically, in this embodiment, continuous data along the temporal dimension is acquired, and data segments related to drought events are extracted from the complete spatiotemporal data matrix. Using a segmented processing approach, the data is arranged chronologically to obtain a preliminary time series dataset. A long short-term memory network (LSTM) is used to construct a model for this preliminary time series dataset, processing the dynamic evolution information within the continuous data. The network's memory mechanism captures inter-data dependencies and determines the evolutionary characteristics of drought events. Based on the captured drought event evolution characteristics, temporal dimension features relevant to trend capture are extracted. If the feature value exceeds a preset threshold, the data for that time period is weighted to obtain an enhanced feature set.

[0069] Furthermore, the drought feature extraction model based on the long short-term memory network includes:

[0070] The dual-channel input layer is used to receive the original time series data and the time series data after feature enhancement processing respectively;

[0071] The multi-scale time decomposition module is used to perform wavelet transform on the data of each channel and decompose it into high-frequency subsequences and low-frequency subsequences. The high-frequency subsequences are used to reflect short-term change characteristics, while the low-frequency subsequences are used to reflect long-term trend characteristics.

[0072] A hybrid GRU-LSTM layer, in which the GRU layer receives high-frequency subsequence data and uses its short-term feature extraction capability to process short-term change features. The LSTM layer receives low-frequency subsequence data and uses its long-distance dependency capture capability to process long-term trend features. The output feature vectors of the GRU layer and the LSTM layer are concatenated to obtain a fusion of short-term and long-term feature vectors.

[0073] The attention mechanism enhancement layer receives the fused feature vector, calculates the attention weight of each time step, multiplies the attention weight by the integrated feature vector, and sums it along the time dimension to obtain the weighted feature vector;

[0074] The multi-task learning output layer is used to set up multi-task output nodes and flow the weighted feature vectors to each output node to predict different drought indicators.

[0075] Among them, the dual-channel input layer: Input channel 1: receives raw time series data, such as rainfall, temperature, etc., with a shape of [batch_size, timesteps, raw_features];

[0076] Input channel 2: receives time series data that has been processed with feature enhancement, such as SPI index, vegetation index, etc., with a shape of [batch_size, timesteps, enhanced_features];

[0077] Multi-scale time decomposition module:

[0078] Wavelet transform or moving average: Perform wavelet transform or use moving average method on the data of each channel to decompose it into high-frequency and low-frequency subsequences;

[0079] Hybrid GRU-LSTM layer:

[0080] The GRU layer is used to process high-frequency subsequences: it receives high-frequency subsequence data and uses its efficient short-term feature extraction capability to process short-term variation features. The shape is maintained as [batch_size, timesteps, gru_units1];

[0081] Activation function: ReLU activation function is used to enhance nonlinear expression capabilities;

[0082] Regularization: Add L2 regularization to prevent overfitting;

[0083] The LSTM layer is used to process low-frequency subsequence data. It receives low-frequency subsequence data and uses its powerful ability to capture long-term dependencies to process long-term trend features. The shape is maintained as [batch_size, timesteps, lstm_units1];

[0084] Activation function: tanh activation function is used to maintain stable information transmission;

[0085] Regularization: Add L2 regularization to prevent overfitting;

[0086] The output feature vectors of the GRU layer and the LSTM layer are concatenated, and the integrated feature shape is [batch_size, timesteps, gru_units1+lstm_units1] to fuse short-term and long-term features;

[0087] Attention-enhanced output layer:

[0088] Attention calculation:

[0089] Attention layer: Receives the integrated feature vector and calculates the attention weight for each time step. Input shape: [batch_size, timesteps, gru_units2 + lstm_units2]; Output shape: [batch_size, timesteps, 1], representing the weight of each time step.

[0090] Attention activation function: Use the Softmax activation function to make the sum of weights equal to 1.

[0091] Feature weighted summation: Multiply the attention weights by the integrated feature vector and sum along the time dimension to obtain the weighted feature vector; input shape: [batch_size, timesteps, gru_units2+lstm_units2] and [batch_size, timesteps, 1]; output shape: [batch_size, gru_units2+lstm_units2], representing the weighted feature vector;

[0092] Multi-task learning output layer: Set up multi-task output nodes:

[0093] Output node 1: predicted drought-related indicator 1 (such as SPI), shape is [batch_size, 1];

[0094] Output node 2: predict drought-related indicator 2 (such as soil moisture), shape is [batch_size, 1];

[0095] Output node 3: predict drought-related indicators 3 (such as vegetation index), the shape is [batch_size, 1;

[0096] Activation function: Use sigmoid or softmax activation function.

[0097] The LSTM model in the prior art usually only processes time series data of a single input channel, while the model of this embodiment can simultaneously receive original time series data and enhanced feature data through dual-channel input; introduce a multi-scale time decomposition module to extract short-term and long-term features respectively; combine the advantages of both LSTM and GRU to form a hybrid structure, taking into account the efficiency and ability of long-term and short-term feature extraction; this model adds an attention mechanism before the output layer to dynamically assign weights to different time steps and highlight key features. The use of a multi-task learning output layer can simultaneously predict multiple drought-related indicators; dual-channel input and multi-scale time decomposition enable the model to more comprehensively extract drought characteristics and improve monitoring accuracy; the hybrid GRU-LSTM structure and attention mechanism enhance the model's ability to express the dynamic evolution characteristics of drought events; the multi-task learning output provides a richer dimension for drought assessment, which helps to fully understand the drought situation; the innovation of the model structure enhances its generalization ability in different regions and under different drought conditions, making it more widely applicable.

[0098] Furthermore, obtaining the spatial feature map includes:

[0099] Use pre-trained convolutional neural networks to extract features from the spatial dimension of spatiotemporal datasets;

[0100] Based on the extracted features, cluster analysis was used to analyze the spatial distribution of drought in the Heihe River Basin, and the high-incidence areas, concentrated areas and spatial distribution patterns of drought were identified.

[0101] The spatial distribution characteristics obtained by analysis are mapped to the geographical space of the Heihe River Basin to generate a spatial characteristic map.

[0102] Furthermore, the convolutional neural network includes:

[0103] The input layer is used to accept data in the spatial dimension of the spatiotemporal dataset;

[0104] The multi-scale feature extraction module is used to set up multiple convolution branches, each branch uses convolution kernels of different sizes to extract features of different scales, and splices the output feature maps of each branch;

[0105] The attention mechanism module is used to calculate the importance weight of each spatial position in the feature map, multiply the spatial attention weight by the feature map, and obtain the weighted feature map.

[0106] Specifically, in this embodiment, the input layer receives spatial distribution data from a complete temporal-spatial data matrix in the format of [batch_size, height, width, channels], where height and width represent the dimensions of the spatial grid, and channels represents different feature channels (such as rainfall, soil moisture, vegetation index, etc.);

[0107] Multi-scale feature extraction module: set up multiple convolution branches, each branch uses convolution kernels of different sizes (such as 3×3, 5×5, 7×7) to extract features of different scales; multi-scale convolution branches: convolution layer 1: use 3×3 convolution kernel to extract fine-grained features; convolution layer 2: use 5×5 convolution kernel to extract medium-scale features; convolution layer 3: use 7×7 convolution kernel to extract large-scale features; depth-wise separable convolution: use depth-wise separable convolution in each convolution branch to reduce computational complexity while maintaining feature extraction capabilities; activation function: use ReLU activation function to enhance nonlinear feature expression.

[0108] Attention mechanism module: Spatial attention layer: calculates the importance weight of each spatial position in the feature map, input shape: [batch_size, height, width, filters_sum]; output shape: [batch_size, height, width, 1], representing the weight of each spatial position; activation function: uses the Sigmoid activation function to normalize the weight to between 0 and 1; feature weighted summation: multiplies the spatial attention weight with the feature map to enhance important features and suppress unimportant features.

[0109] Traditional CNNs usually use convolution kernels of a single scale. This application adopts multi-scale convolution branches to extract spatial features of different scales. This model uses depthwise separable convolution to reduce computational complexity and improve computational efficiency. Multi-scale feature extraction and depthwise separable convolution enable the model to more accurately extract drought spatial features and improve monitoring accuracy. The attention mechanism highlights key spatial location features through weight distribution, improving the model's ability to identify drought core areas.

[0110] Furthermore, based on the spatiotemporal joint feature matrix, drought monitoring in the Heihe River Basin includes:

[0111] Obtain eigenvalue sequence data from the spatiotemporal joint feature matrix, analyze the temporal characteristics of each region one by one, and obtain preliminary eigenvalue distribution results;

[0112] According to the eigenvalue distribution results, if the eigenvalue sequence of a certain area exceeds the preset drought threshold, the area is judged to be in a drought warning state and the warning mark of the area is determined;

[0113] For areas with identified warning signs, the corresponding spatial feature data is extracted from the space-time matrix, and spatial mapping technology is used to generate a spatial feature map of the area to obtain preliminary visualization results of the drought impact range;

[0114] By conducting boundary analysis on the spatial feature map and combining the relationship between spatial features and the impact range, the specific boundaries of the drought impact range can be determined and quantitative data for range assessment can be obtained;

[0115] Based on the quantitative data of the scope assessment and combined with the changing trends of time characteristics, the persistence of the drought warning status is analyzed and the dynamic evolution of the drought impact is determined.

[0116] Furthermore, the spatiotemporal joint feature matrix is ​​a spatial grid point of a preset dimension, each grid point corresponds to a time series with a time span of a preset number of days in the past, and the daily feature values ​​include NDVI data, rainfall, and soil moisture.

[0117] Specifically, in this embodiment, the spatiotemporal joint feature matrix is ​​a spatial grid point with a matrix dimension of 100×100, each grid point corresponds to a time series with a time span of the past 30 days, and the daily feature values ​​include NDVI data, rainfall, and soil moisture.

[0118] Based on historical drought data from the Heihe River Basin and field research, a reasonable drought warning threshold was established. This threshold represents the boundary between drought and non-drought conditions, based on the comprehensive drought characteristic vector. The threshold setting needs to take into account seasonal and regional variations, as well as the gradual nature of drought development. For example, the threshold can be appropriately raised during the rainy season and lowered during the dry season.

[0119] Using a weighted summation method, rainfall, soil moisture, and NDVI data are multiplied by their corresponding weight coefficients and then added together to obtain a comprehensive drought characteristic value. The formula is: Comprehensive drought characteristic value = (rainfall weight × rainfall data) + (soil moisture weight × soil moisture data) + (NDVI weight × NDVI data).

[0120] The calculated comprehensive drought characteristic value is compared with a pre-set drought warning threshold. If the comprehensive drought characteristic value is lower than the drought warning threshold, the region is deemed to be in a drought warning state. The severity of the drought is further categorized into mild, moderate, and severe levels, depending on the degree of the drought below the threshold, so that appropriate drought mitigation measures can be implemented. If the comprehensive drought characteristic value is higher than or equal to the drought warning threshold, the region is deemed to be non-drought, but continuous monitoring of relevant data is still necessary to detect signs of drought. Timely and accurate monitoring of drought conditions will help ensure stable agricultural production in the Heihe River Basin and reduce crop yield losses and economic losses caused by drought. It will also help protect the ecological environment within the basin, maintain ecological balance, and promote sustainable development in the region.

[0121] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A drought monitoring method for the Heihe River Basin based on spatiotemporal convolutional networks, characterized in that: include: A spatiotemporal dataset was constructed by acquiring multi-source data of the Heihe River Basin from satellite remote sensing, meteorological stations, and ground observation points; Based on the spatiotemporal dataset, a drought feature extraction model based on a long short-term memory network is used to capture the dynamic evolution trend of drought events and obtain time series feature vectors; A convolutional neural network was used to extract features from the spatial dimension of the spatiotemporal dataset, analyze the spatial distribution of drought in the Heihe River Basin, and obtain a spatial feature map. Performing feature fusion on the time series feature vector and the spatial feature map to construct a spatiotemporal joint feature matrix; wherein the spatiotemporal joint feature matrix is ​​used to reflect the comprehensive variation characteristics of drought events in time and space and determine the evolution pattern of drought events; Based on the spatiotemporal joint feature matrix, drought monitoring in the Heihe River Basin is carried out.

2. The Heihe River Basin drought monitoring method based on spatiotemporal convolutional network according to claim 1 is characterized in that: The multi-source data of the Heihe River Basin obtained from satellite remote sensing, meteorological stations and ground observation points include: NDVI data of preset spatial resolution are obtained from satellites, daily rainfall data are obtained from meteorological stations, and soil moisture data are obtained from ground observation points.

3. The Heihe River Basin drought monitoring method based on spatiotemporal convolutional network according to claim 1 is characterized in that: Building a spatiotemporal dataset includes: Data preprocessing methods are used to standardize the format and time align data from different sources, address differences in spatial resolution, and generate data in a unified format; For data in a unified format, spatial interpolation technology is used to improve the resolution of data with a spatial resolution lower than a preset resolution threshold. At the same time, a time series smoothing algorithm is combined to fill in the missing data values ​​in the time dimension to obtain the spatiotemporal dataset.

4. The Heihe River Basin drought monitoring method based on spatiotemporal convolutional network according to claim 1 is characterized in that: Obtaining the time series feature vector includes: Based on the spatiotemporal dataset, continuous data in the time dimension is obtained, data segments related to drought events are extracted, and the data are arranged in chronological order using a segmented processing method to obtain an initial time series dataset; Based on the pre-trained drought feature extraction model, the dynamic evolution information in the continuous data is obtained for the initial time series data set, the dependency relationship between the data is captured through the memory mechanism of the network, and the evolution characteristics of the drought event are determined; According to the captured evolution characteristics of drought events, the time dimension characteristic value related to trend capture is extracted. If the time dimension characteristic value exceeds a preset threshold, the data of the time period exceeding the threshold is weighted to obtain the time series characteristic vector.

5. The Heihe River Basin drought monitoring method based on spatiotemporal convolutional network according to claim 1 is characterized in that: Acquiring the spatial feature map includes: Use pre-trained convolutional neural networks to extract features from the spatial dimension of spatiotemporal datasets; Based on the extracted features, cluster analysis was used to analyze the spatial distribution of drought in the Heihe River Basin, and the high-incidence areas, concentrated areas and spatial distribution patterns of drought were identified. The spatial distribution characteristics obtained by analysis are mapped to the geographical space of the Heihe River Basin to generate a spatial characteristic map.

6. The Heihe River Basin drought monitoring method based on spatiotemporal convolutional network according to claim 1 is characterized in that: The drought feature extraction model based on the long short-term memory network includes: The dual-channel input layer is used to receive the original time series data and the time series data after feature enhancement processing respectively; A multi-scale time decomposition module is used to perform wavelet transform on the data of each channel and decompose it into high-frequency subsequences and low-frequency subsequences; wherein the high-frequency subsequence is used to reflect short-term change characteristics, and the low-frequency subsequence is used to reflect long-term trend characteristics; A hybrid GRU-LSTM layer, in which the GRU layer receives high-frequency subsequence data and uses its short-term feature extraction capability to process short-term change features. The LSTM layer receives low-frequency subsequence data and uses its long-distance dependency capture capability to process long-term trend features. The output feature vectors of the GRU layer and the LSTM layer are concatenated to obtain a fusion of short-term and long-term feature vectors. The attention mechanism enhancement layer receives the fused feature vector, calculates the attention weight of each time step, multiplies the attention weight by the integrated feature vector, and sums it along the time dimension to obtain the weighted feature vector; The multi-task learning output layer is used to set up multi-task output nodes and flow the weighted feature vectors to each output node to predict different drought indicators.

7. The Heihe River Basin drought monitoring method based on spatiotemporal convolutional network according to claim 1 is characterized in that: The convolutional neural network includes: The input layer is used to accept data in the spatial dimension of the spatiotemporal dataset; The multi-scale feature extraction module is used to set up multiple convolution branches, each branch uses convolution kernels of different sizes to extract features of different scales, and splices the output feature maps of each branch; The attention mechanism module is used to calculate the importance weight of each spatial position in the feature map, multiply the spatial attention weight by the feature map, and obtain the weighted feature map.

8. The Heihe River Basin drought monitoring method based on spatiotemporal convolutional network according to claim 1 is characterized in that: Based on the spatiotemporal joint feature matrix, drought monitoring in the Heihe River Basin includes: Acquire eigenvalue sequence data from the spatiotemporal joint feature matrix, analyze the temporal features of each region one by one, and obtain preliminary eigenvalue distribution results; According to the eigenvalue distribution results, if the eigenvalue sequence of a certain area exceeds the preset drought threshold, the area is judged to be in a drought warning state and the warning mark of the area is determined; For areas with identified warning signs, the corresponding spatial feature data is extracted from the space-time matrix, and spatial mapping technology is used to generate a spatial feature map of the area to obtain preliminary visualization results of the drought impact range; By conducting boundary analysis on the spatial feature map and combining the relationship between spatial features and the impact range, the specific boundaries of the drought impact range can be determined and quantitative data for range assessment can be obtained; Based on the quantitative data of the scope assessment and combined with the changing trends of time characteristics, the persistence of the drought warning status is analyzed and the dynamic evolution of the drought impact is determined.

Images