Steppe drought monitoring method based on space-time convolutional network

Abstract

The invention discloses a grassland drought monitoring method based on a space-time convolutional network, and belongs to the technical field of space-time convolutional networks, and the method comprises the steps: carrying out the dynamic multi-source heterogeneous data fusion and structured coding through the fusion of remote sensing, weather and human activity agent data, generating a unified data cube, and carrying out the dynamic multi-source heterogeneous data fusion and structured coding; according to the method, through dynamic multi-source data fusion, double-flow feature extraction, attention mechanism decoupling, drought feature enhancement, causal correction and dynamic threshold visualization, accurate extraction and non-interference attribution of drought signals are realized, fixed threshold segmentation and single data source limitation are broken through, the problem of confusion of human activities and meteorological drought is solved, and the accuracy of drought detection is improved. The spatial resolution, the time precision and the ecological interpretation capacity of drought monitoring are remarkably improved, and the finally generated grassland true drought monitoring graph can directly support ecological management decisions and adapt to dynamic requirements of different grassland ecological regions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of space-time convolution network, and particularly relates to a grassland drought monitoring method based on a space-time convolution network. BACKGROUND

[0002] With the intensification of climate change, grassland ecosystems are facing increasingly frequent drought stress; traditional drought monitoring mainly relies on single remote sensing vegetation index or meteorological indicators, which is difficult to distinguish natural climate drought from vegetation degradation caused by human activities, resulting in misjudgment and decision deviation.

[0003] However, the existing grassland drought monitoring method still has certain defects, the prior art uses a single flow network or a simple linear separation method to process multi-source data, which cannot effectively distinguish the drought signal driven by meteorology from the vegetation degradation caused by human activities, and relies on fixed time window smoothing or spatial filtering in the feature enhancement stage, without adaptive optimization combined with specific ecological background; such methods are prone to misjudgment of short-term climate events as persistent drought trends, or loss of key drought turning points due to excessive smoothing, lack of explicit causal correction logic, and most of them directly use original or preliminary enhanced features as drought indices without quantifying and deducting human disturbance; although some methods attempt to introduce a threshold to remove high interference areas, a fixed threshold is used, which cannot adapt to the intensity gradient changes of human activities in different grassland areas, and a preset fixed threshold is used for grading, ignoring the sensitivity differences of different grassland types to drought; the same threshold may be too sensitive in humid areas and sluggish in response in arid areas, resulting in a disconnection between the grading results and the actual ecological conditions, therefore, the grassland drought monitoring method based on the space-time convolution network is proposed. SUMMARY

[0004] The purpose of the present application is to provide a grassland drought monitoring method based on a space-time convolution network to solve the problems raised in the background.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a grassland drought monitoring method based on a space-time convolution network, comprising the following steps: S1, dynamic multi-source heterogeneous data fusion and structured coding are performed by fusing remote sensing, meteorological and human activity proxy data to generate a unified data cube; S2, based on the generated data cube, double-flow space-time feature preliminary extraction is performed to output meteorological vegetation flow features and human vegetation flow features; S3, based on the extracted preliminary features, feature decoupling and interaction are performed through an attention mechanism to separate meteorological exclusive features and human exclusive features; S4, drought feature enhancement based on meteorological exclusive features is performed to generate enhanced drought features; S5. Integrate enhanced drought characteristics with anthropogenic exclusive characteristics, and perform causal characteristic fusion and true drought index generation; S6. Dynamic threshold segmentation and drought level visualization are performed on the generated true drought index to finally generate a grassland true drought monitoring map.

[0006] Preferably, in step S1, hyperspectral remote sensing image sequences, meteorological station time-series data, and anthropogenic activity proxy data are collected. Integrity checks are performed on each type of data, outliers are detected and marked, radiometric correction is applied to the remote sensing images, converting the raw digital pixel values ​​of the remote sensing images to reflectance, atmospheric correction is performed, time-series interpolation is performed on the meteorological data, spatial standardization is performed on the anthropogenic activity data, all data are resampled at a unified time granularity and spatial resolution, geometric correction is performed on the remote sensing images, accuracy correction is performed using ground control points, spatial coordinate unification is performed on data from different sources, preliminary data fusion is performed using principal component analysis, feature encoding is performed on the fused data, and remote sensing vegetation indices, meteorological parameters, and anthropogenic activity indicators are encoded into multi-dimensional feature vectors to construct a unified data cube.

[0007] Preferably, in step S2, a unified data cube is received, and the data is split according to feature categories: meteorological vegetation flow cube: extracting 5 feature dimensions including vegetation index, surface temperature, precipitation, air temperature, and evapotranspiration; anthropogenic vegetation flow cube: extracting 4 feature dimensions including vegetation index, nighttime light intensity, grazing intensity, and land use change. Meteorological Vegetation Flow Network: The input layer receives a meteorological vegetation flow cube, constructs a 3D convolutional layer, extracts local spatiotemporal correlations layer by layer, adds ReLU activation function and batch normalization after each convolution to suppress noise and accelerate convergence, and compresses redundant time information through 2 layers of time dimension downsampling to output the initially extracted meteorological feature tensor. The meteorological vegetation flow cube is input into the meteorological vegetation flow network. After feature extraction, the meteorological features are output, as shown in the formula: , In the formula, Indicates time t and position Meteorological vegetation flow characteristics, Indicates time t and position The daily precipitation intensity, Indicates time t and position The rate of temperature change Indicates time t and position The normalized vegetation index, where C represents the normalization constant. Preferably, the S2, the artificial vegetation flow network: the input layer receives the artificial vegetation flow cube, captures the dynamic correlation of the artificial activity vegetation through the same 3-layer 3D convolution structure as the meteorological flow, and also adds ReLU activation and batch normalization, and preliminarily extracts the artificial feature tensor; The artificial vegetation flow cube is input into the artificial vegetation flow network, and the artificial feature is output after feature extraction, and the realization formula is: , In the formula, represents the artificial vegetation flow feature of time t, location , represents the nighttime light intensity of time t, location , represents the grazing intensity index of time t, location , represents the normalized vegetation index of time t, location , represents the grassland area ratio of location .

[0008] Preferably, the S3, the meteorological vegetation flow feature and the artificial vegetation flow feature are obtained, the meteorological vegetation flow feature is taken as a query, the artificial vegetation flow feature is taken as a key and a value, and the attention weight of the meteorological feature to the artificial feature is calculated The artificial vegetation flow feature is taken as a query, and the meteorological vegetation flow feature is taken as a key and a value, and the attention weight of the artificial feature to the meteorological feature is calculated The weighted average of the meteorological vegetation flow feature and the artificial vegetation flow feature is taken as a shared feature representation, and the realization is: , In the formula, S represents a shared feature, represents a normalization constant, represents the attention weight of the meteorological feature to the artificial feature, , represents the attention weight of the artificial feature to the meteorological feature, .

[0009] Preferably, the S3, the shared feature is subtracted from the meteorological vegetation flow feature to obtain a meteorological exclusive feature, and the realization is: , In the formula, represents the meteorological exclusive feature, represents a gating factor; The shared information is subtracted from the artificial vegetation flow feature to obtain an artificial exclusive feature, and the realization is: ​​ , In the formula, represents the artificial monopoly feature.

[0010] Preferably, the S4 acquires the weather monopoly feature, the weather monopoly feature contains short-term fluctuation noise, and the enhanced persistent drought is implemented as: , In the formula, represents the time-weighted weather feature, represents the time weight coefficient, represents the weather monopoly feature, and k represents the historical days.

[0011] Preferably, the S4 enhances the spatial consistency based on the time-weighted weather feature, generates the enhanced drought feature, and is implemented as: , In the formula, represents the enhanced drought feature, represents the gating factor, represents the spatial neighborhood, represents the spatial weight coefficient, represents the center point reservation weight.

[0012] Preferably, the S5 acquires the enhanced drought feature and the artificial monopoly feature, takes the artificial monopoly feature as the interference weight, dynamically adjusts the contribution of the enhanced drought feature, controls the correction strength through the gating factor γ, and implements the interference correction factor as: , In the formula, represents the interference correction factor.

[0013] Preferably, the S5 is based on the interference correction factor, and the true drought index is implemented as: , In the formula, represents the true drought index.

[0014] Preferably, in the S5; the output true drought index timestamp and spatial resolution are consistent with the grassland geographical boundary, automatically eliminate invalid data points, and verify through historical drought event backtracking: check whether the distribution of the true drought index in the known drought year matches the spatial range recorded by the meteorological department, exclude data anomalies, calculate the empirical distribution function of the true drought index value of all spatial positions, identify the key dividing point in the distribution, instead of a fixed threshold, and perform drought grade division, four-level dynamic classification: No drought: the true drought index value is lower than the preset first dividing point, and the vegetation state is normal.

[0015] Light drought: the true drought index value is between the preset first and second boundary points, and the vegetation is slightly stressed.

[0016] Moderate drought: the true drought index value is between the preset second and third boundary points, and the vegetation is significantly degraded.

[0017] Severe drought: the true drought index value is higher than the preset third boundary point, and the vegetation is on the verge of collapse; Assign a unique color to each level to form a continuous color band, and perform spatial smoothing: fine-tune the levels of adjacent grids, overlay the grassland administrative boundary, major rivers and roads on the base map, mark the time stamp, add the legend and scale, and improve the readability.

[0018] Compared with the prior art, the beneficial effects of the present application are: 1. The present application realizes accurate extraction and non-interference attribution of drought signals by dynamic multi-source data fusion, double-flow feature extraction, attention mechanism decoupling, drought feature enhancement, causal correction and dynamic threshold visualization, breaks through the limitations of fixed threshold segmentation and single data source, solves the confusion problem of human activities and meteorological drought, significantly improves the spatial resolution, time accuracy and ecological interpretation of drought monitoring, and finally generates a grassland true drought monitoring map that can directly support ecological management decisions and adapt to the dynamic needs of different grassland ecological zones; 2. The present application realizes dynamic decoupling of meteorological and human features through the attention mechanism, calculates the attention weight of meteorological features on human features, and extracts shared features through weighted average, and obtains pure meteorological exclusive features and pure human exclusive features by subtracting the shared part from the original features through the gating factor, ensuring the integrity and adaptability of the decoupling process, avoiding signal loss caused by fixed threshold or simple subtraction, making the contribution of drought signals and human activities quantifiable and separated, and improving the fitting ability of the model to nonlinear ecological relationships; 3. The present application strengthens the persistence of drought signals by time weighting and spatial aggregation for short-term noise (such as single-day rainfall interference) in meteorological exclusive features. In the time dimension, the historical similarity weight is dynamically calculated to retain long-term trends. In the spatial dimension, the neighborhood gradient similarity is aggregated to enhance regional consistency. Combined with time and space dual context perception, the purity of the drought signal is ensured. 4. The present application generates a pure meteorological driven true drought index through dynamic interference correction, uses the output human exclusive features as interference weight, combines with the shared gating factor, dynamically adjusts and enhances the contribution of drought features, realizes continuous correction instead of binary threshold, adapts to the gradient change of human interference intensity, and significantly improves the accuracy of drought attribution analysis. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The operation flow of the grassland drought monitoring method based on the spatio-temporal convolution network of the present application Figure One ; Figure 2 The operation flow of the grassland drought monitoring method based on the spatio-temporal convolution network of the present application Figure Two ; Figure 3 The operation flow of the grassland drought monitoring method based on the spatio-temporal convolution network of the present application Figure Three ; Figure 4 The operation flow of the grassland drought monitoring method based on the spatio-temporal convolution network of the present application Figure Four ; Figure 5 The operation flow of the grassland drought monitoring method based on the spatio-temporal convolution network of the present application Figure Five . DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application. EMBODIMENT

[0021] Please refer to Figures 1-5 , the present application provides a technical solution: comprising the following steps: S1, by fusing remote sensing, meteorological and human activity agent data, dynamic multi-source heterogeneous data fusion and structured coding are carried out, and a unified data cube is generated; S2, based on the generated data cube, double-flow spatio-temporal feature preliminary extraction is carried out, and meteorological vegetation flow feature and human vegetation flow feature are respectively output; S3, based on the extracted preliminary features, feature decoupling and interaction are carried out through attention mechanism, and meteorological exclusive features and human exclusive features are separated; S4, based on the meteorological exclusive features, drought feature enhancement of spatio-temporal context perception is carried out, and enhanced drought features are generated; S5, the enhanced drought features and human exclusive features are fused, and causal feature fusion and true drought index generation are carried out; S6, the generated true drought index is subjected to dynamic threshold segmentation and drought grade visualization, and finally a grassland true drought monitoring map is generated.

[0022] In this embodiment, the S1 collects hyperspectral remote sensing image sequences, meteorological station time series data, and human activity agent data, performs integrity checking on each type of data, detects outliers and performs marking processing, performs radiation correction on the remote sensing image, converts the original digital pixel value of the remote sensing image into reflectivity, performs atmospheric correction, performs time series interpolation on the meteorological data, performs spatial standardization on the human activity data, resamples all data according to a unified time granularity and spatial resolution, performs geometric correction on the remote sensing image, performs precision correction using ground control points, unifies the spatial coordinates of data from different sources, aligns the time of different phase data, and accurately matches the acquisition time of the remote sensing image with the recording time of the meteorological station; the human activity data is processed in time synchronization, the historical grazing intensity data is associated with the acquisition time of the remote sensing image, the preliminary data fusion is performed through the principal component analysis method, the fused data is feature coded, the remote sensing vegetation index, meteorological parameters, and human activity indicators are coded into a multi-dimensional feature vector, a unified data cube is constructed, and the data is organized into a three-dimensional structure: spatial dimension, time dimension, and feature dimension.

[0023] In this embodiment, the S2 receives the unified data cube and splits the data according to the feature categories: meteorological vegetation flow cube: extracts vegetation index, surface temperature, precipitation, air temperature, and evapotranspiration as five feature dimensions, and constructs a cube with time × space × 5 channels; human vegetation flow cube: extracts vegetation index, nighttime light intensity, grazing intensity, and land use change as four feature dimensions, and constructs a cube with time × space × 4 channels. Meteorological vegetation flow network: the input layer receives the meteorological vegetation flow cube, constructs three 3D convolution layers, extracts local spatiotemporal correlation layer by layer, adds ReLU activation function and batch normalization after each convolution to suppress noise and accelerate convergence, performs 2-layer time dimension downsampling to compress redundant time information, and outputs the preliminary extracted meteorological feature tensor; The meteorological vegetation flow cube is input into the meteorological vegetation flow network, and the meteorological feature is output after feature extraction, and the formula is: , In the formula, represents the meteorological vegetation flow feature at time t and location characterizes the comprehensive influence of meteorological factors on vegetation, represents the daily precipitation intensity at time t and location , represents the air temperature change rate at time t and location , which is derived from the current day air temperature and the previous day air temperature , , represents the meteorological vegetation flow feature at time t and location NDVI, C represents a normalization constant, In this embodiment, the S2, the artificial vegetation flow network: the input layer receives the artificial vegetation flow cube, captures the dynamic correlation of the artificial activity vegetation through the same 3-layer 3D convolution structure as the meteorological flow, such as the fluctuation rule of grazing intensity-vegetation index, and adds ReLU activation and batch normalization, carries out 2-layer time downsampling, and preliminarily extracts the artificial feature tensor; The artificial vegetation flow cube is input into the artificial vegetation flow network, and the artificial feature is output after feature extraction, and the realization formula is: , In the formula, represents the artificial vegetation flow feature of time t, position , which represents the interference degree of human activity on vegetation, represents the night light intensity of time t, position , represents the grazing intensity index of time t, position , represents the normalized vegetation index of time t, position , characterizes the decline amplitude of vegetation coverage, and is positively correlated with the damage effect of human activity, represents the grassland area proportion of position .

[0024] In this embodiment, the S3, the meteorological vegetation flow feature and the artificial vegetation flow feature are obtained, the meteorological vegetation flow feature is taken as a query, the artificial vegetation flow feature is taken as a key and a value, and the attention weight of the meteorological feature to the artificial feature is calculated The artificial vegetation flow feature is taken as a query, and the meteorological vegetation flow feature is taken as a key and a value, and the attention weight of the artificial feature to the meteorological feature is calculated The weighted average of the meteorological vegetation flow feature and the artificial vegetation flow feature is taken as the shared feature representation, and the realization is: , In the formula, S represents the shared feature, which represents the common spatiotemporal pattern in the meteorological and artificial features, represents a normalization constant, represents the attention weight of the meteorological feature to the artificial feature, , represents the attention weight of the artificial feature to the meteorological feature, .

[0025] ​​In this embodiment, S3, subtract the shared features from the weather vegetation flow features to obtain weather-exclusive features, only retain weather-driven vegetation anomaly signals, and is implemented as: , In the formula, represents the weather-exclusive feature, represents the gating factor, dynamically adjusting the decoupling strength ; Subtract the shared information from the artificial vegetation flow features to obtain artificial-exclusive features, and only retain vegetation anomaly signals driven by human activities, which is implemented as: , In the formula, represents the artificial-exclusive feature.

[0026] In this embodiment, S4, obtain the weather-exclusive feature, the weather-exclusive feature contains short-term fluctuation noise, and perform enhanced persistent drought, time-weighted weather feature is implemented as: , In the formula, represents the time-weighted weather feature, representing the persistent drought signal, represents the time weight coefficient, dynamically calculating the historical pattern similarity , represents the weather-exclusive feature, represents the value of the weather-exclusive feature at time point t-k, and k represents the number of historical days.

[0027] In this embodiment, S4, based on the time-weighted weather feature, perform enhanced spatial consistency, generate enhanced drought features, which is implemented as: , In the formula, represents the enhanced drought feature, represents the gating factor, represents the spatial neighborhood, the adjacent position set centered on , such as to , represents the spatial weight coefficient, dynamically calculating the spatial gradient , represents the center point retention weight.

[0028] In this embodiment, S5, obtain the enhanced drought feature and the artificial-exclusive feature, take the artificial-exclusive feature as the interference weight, dynamically adjust the contribution of the enhanced drought feature, control the correction strength through the gating factor γ, and implement the interference correction factor as: , In the formula, represents the interference correction factor.

[0029] Preferably, in the S5, the true drought index is realized based on the interference correction factor as follows: , In the formula, represents the true drought index, represents the correction coefficient, when H=0, no interference, Q=D, retaining all weather signals, when H=1, strong interference, Q=0, completely excluding human influence.

[0030] Preferably, in the S5, the time stamp and spatial resolution of the output true drought index are consistent with the grassland geographical boundary, invalid data points are automatically removed, and the true drought index is verified by historical drought events: checking whether the distribution of the true drought index in the known drought years matches the spatial range recorded by the meteorological department, excluding data anomalies, counting the true drought index values of all spatial positions, calculating the empirical distribution function, identifying the key dividing points in the distribution instead of fixed thresholds, and classifying the drought levels, four dynamic levels: No drought: the true drought index value is lower than the preset first dividing point, and the vegetation state is normal.

[0031] Mild drought: the true drought index value is between the preset first and second dividing points, and the vegetation is slightly stressed.

[0032] Moderate drought: the true drought index value is between the preset second and third dividing points, and the vegetation is significantly degraded.

[0033] Severe drought: the true drought index value is higher than the preset third dividing point, and the vegetation is on the verge of collapse. Assign a dedicated color to each level to form a continuous gradient color band, and perform spatial smoothing processing: fine-tune the level of adjacent grids, overlay the grassland administrative boundary, main rivers and roads on the base map, mark the time stamp, add the legend and scale, and improve the readability.

[0034] Working principle: by integrating hyperspectral remote sensing, meteorological stations and human activity proxy data, the problem of heterogeneity of multi-source data is solved; the remote sensing image is subjected to radiation correction, atmospheric correction and geometric correction to eliminate sensor noise and spatial deviation; the meteorological data is subjected to time series interpolation to fill in missing values; the human activity data is subjected to spatial standardization to unify the expression form; finally, the multi-source information is preliminarily fused through principal component analysis, and is encoded into a unified data cube containing vegetation index, precipitation, grazing intensity and other dimensions; Two types of features are extracted by double-flow network architecture; the meteorological vegetation flow network takes 5-dimensional features such as vegetation index, precipitation, and air temperature as input, captures local spatio-temporal correlation layer by layer through 3 layers of 3D convolution, and uses ReLU activation and batch normalization to suppress noise, and outputs meteorological driving features; the human vegetation flow network takes 4-dimensional features such as nighttime light and grazing intensity as input, and uses the same structure to extract the influence of human activities on vegetation; the two flow networks are processed in parallel; the dynamic decoupling of meteorological and human features is realized through a bidirectional attention mechanism; the meteorological features are taken as queries to calculate the attention weight of human features, and vice versa, forming a shared feature representation; then the shared part is subtracted from the original features through a gating factor to obtain meteorological exclusive features and human exclusive features, and the persistence and consistency of drought signals are enhanced through time weighting and spatial aggregation; in the time dimension, the weight is dynamically calculated based on the similarity of historical meteorological vegetation patterns to retain long-term trends; in the spatial dimension, the isolated noise points are suppressed through neighborhood gradient similarity aggregation; the enhanced drought features generated finally integrate time trends and spatial consistency, and the true drought index driven by pure meteorological factors is generated through dynamic interference correction; the contribution of the enhanced drought features is dynamically adjusted by taking the human exclusive features as the interference weight and combining the shared gating factor; the data-driven dynamic threshold segmentation replaces the fixed threshold to ensure that the drought grade division matches the regional ecological sensitivity; based on the empirical distribution function of the true drought index, the key dividing points are automatically identified to divide the drought into four grades; the monitoring map is generated through color gradient and spatial smoothing processing, and geographical elements such as administrative boundaries and rivers are superimposed, and the time stamp and legend are marked.

[0035] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

[0036] The above describes the present application and its embodiments, which are not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In summary, if a person skilled in the art is inspired by it, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution should belong to the protection scope of the present application.

Claims

1. A grassland drought monitoring method based on spatiotemporal convolutional networks, characterized in that, Includes the following steps: S1. By integrating remote sensing, meteorological, and human activity proxy data, dynamic multi-source heterogeneous data fusion and structured coding are performed to generate a unified data cube; S2. Based on the generated data cube, perform preliminary extraction of the spatiotemporal features of the dual-flow system, and output the meteorological vegetation flow features and the anthropogenic vegetation flow features respectively; S3. Based on the extracted preliminary features, feature decoupling and interaction are performed through an attention mechanism to separate meteorological exclusive features and human exclusive features. S4. Enhance drought features based on unique meteorological characteristics and spatiotemporal context awareness to generate enhanced drought features; S5. Integrate enhanced drought characteristics with anthropogenic exclusive characteristics, and perform causal characteristic fusion and true drought index generation; S6. Dynamic threshold segmentation and drought level visualization are performed on the generated true drought index to finally generate a grassland true drought monitoring map.

2. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 1, characterized in that: In step S1, hyperspectral remote sensing image sequences, meteorological station time-series data, and anthropogenic activity proxy data are collected. Integrity checks are performed on each type of data, outliers are detected and marked, radiometric correction is applied to the remote sensing images, converting the raw digital pixel values ​​of the remote sensing images to reflectance, atmospheric correction is performed, time-series interpolation is performed on the meteorological data, spatial standardization is performed on the anthropogenic activity data, all data are resampled at a unified time granularity and spatial resolution, geometric correction is performed on the remote sensing images, accuracy correction is performed using ground control points, spatial coordinate unification is performed on data from different sources, preliminary data fusion is performed using principal component analysis, feature encoding is performed on the fused data, and remote sensing vegetation indices, meteorological parameters, and anthropogenic activity indicators are encoded into multi-dimensional feature vectors to construct a unified data cube.

3. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 1, characterized in that: In S2, a unified data cube is received, and the data is split according to feature categories, including meteorological vegetation flow cubes and anthropogenic vegetation flow cubes. Meteorological Vegetation Flow Network: The input layer receives a meteorological vegetation flow cube, constructs a 3D convolutional layer, extracts local spatiotemporal correlations layer by layer, adds ReLU activation function and batch normalization after each convolution to suppress noise and accelerate convergence, and compresses redundant time information through 2 layers of time dimension downsampling to output the initially extracted meteorological feature tensor. The meteorological vegetation flow cube is input into the meteorological vegetation flow network. After feature extraction, the meteorological features are output, as shown in the formula: ， In the formula, Indicates time t and position Meteorological vegetation flow characteristics, Indicates time t and position The daily precipitation intensity, Indicates time t and position The rate of temperature change Indicates time t and position The normalized vegetation index is denoted by C, where C represents the normalization constant.

4. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 3, characterized in that: In S2, the anthropogenic vegetation flow network: the input layer receives the anthropogenic vegetation flow cube, and captures the dynamic correlation of anthropogenic vegetation through the same 3-layer 3D convolutional structure as meteorological flow. Similarly, ReLU activation and batch normalization are added to initially extract the anthropogenic feature tensor. The anthropogenic vegetation flow cube is input into the anthropogenic vegetation flow network. After feature extraction, the anthropogenic features are output, as shown in the formula: ， In the formula, Indicates time t and position Anthropogenic vegetation flow characteristics Indicates time t and position The intensity of nighttime lights, Indicates time t and position Grazing intensity index Indicates time t and position Normalized Difference Vegetation Index (NDVI) Indicates position The proportion of grassland area.

5. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 1, characterized in that: In step S3, meteorological vegetation flow features and anthropogenic vegetation flow features are obtained. The meteorological vegetation flow features are used as the query, and the anthropogenic vegetation flow features are used as the key and value. The attention weight of meteorological features on anthropogenic features is calculated. Using anthropogenic vegetation flow features as queries and meteorological vegetation flow features as keys and values, the attention weights of anthropogenic features on meteorological features are calculated. The weighted average of meteorological vegetation flow characteristics and anthropogenic vegetation flow characteristics is used as the shared feature representation, which is implemented as follows: ， In the formula, S represents the shared feature. Represents the normalization constant. This represents the attention weight of meteorological features on anthropogenic features. This represents the attention weight of human features on meteorological features.

6. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 5, characterized in that: In step S3, the shared features are subtracted from the meteorological vegetation flow features to obtain the meteorological exclusive features, which is implemented as follows: ， In the formula, Indicates a unique meteorological characteristic. Indicates the gating factor; Subtracting shared information from anthropogenic vegetation flow characteristics yields anthropogenic exclusive characteristics, implemented as follows: ， In the formula, This indicates a characteristic of human-created exclusive ownership.

7. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 1, characterized in that: In step S4, meteorological exclusive features are acquired. These meteorological exclusive features include short-term fluctuation noise. The sustained drought is reinforced, and the time-weighted meteorological features are implemented as follows: ， In the formula, Indicates time-weighted meteorological characteristics. This represents the time weighting coefficient. This indicates a unique meteorological characteristic, and k represents the number of historical days.

8. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 7, characterized in that: In step S4, based on time-weighted meteorological characteristics, spatial consistency is enhanced to generate enhanced drought characteristics, which is achieved as follows: ， In the formula, This indicates an enhanced drought characteristic. Indicates the gating factor. Represents spatial neighborhood, Indicates the spatial weighting coefficient. This indicates that the center point retains its weight.

9. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 1, characterized in that: In step S5, enhanced drought features and anthropogenic exclusive features are obtained. The anthropogenic exclusive features are used as interference weights, and the contribution of the enhanced drought features is dynamically adjusted. The correction intensity is controlled by a gating factor γ. The interference correction factor is implemented as follows: ， In the formula, This represents the interference correction factor.

10. The grassland drought monitoring method based on spatiotemporal convolutional networks according to claim 9, characterized in that: In S5, the true drought index is realized based on the interference correction factor as follows: ， In the formula, This indicates the true drought index.

Images