Reservoir drawdown zone evolution early warning method based on time-series remote sensing images and intelligent ai

Abstract

The application discloses a reservoir drawdown zone evolution early warning method based on time sequence remote sensing images and intelligent AI, relates to the technical field of big data analysis, and comprises the following steps: acquiring historical reservoir drawdown zone remote sensing image data, dividing the historical reservoir drawdown zone change trend vector according to the known drawdown zone geological disaster risk event development preference, and constructing a plurality of reservoir drawdown zone change trend entity-risk event relationship graph; acquiring real-time target reservoir drawdown zone remote sensing image data, constructing an entity-risk event development path transmission tracking model, and generating a real-time target reservoir drawdown zone attribution entity-risk event development path; predicting an upstream and downstream environment vector of a future target reservoir, dynamically compensating the real-time target reservoir drawdown zone attribution entity-risk event development path, and generating a future target reservoir drawdown zone attribution entity-risk event triggering probability. The application provides strong action-oriented forward-looking decision support for reservoir safety management, and realizes a change from passive response to active prevention and control.

Description

Technical Field

[0001] This invention relates to the field of big data analysis technology, specifically to a method for early warning of reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI. Background Technology

[0002] Current methods for early warning of reservoir drawdown zone evolution mainly rely on manual inspections and single-point monitoring, and qualitative judgment based on expert experience, lacking systematic quantitative analysis of massive amounts of time-series remote sensing data; static threshold early warning models cannot capture the nonlinear characteristics in the dynamic evolution of drawdown zones; geological disaster risk assessment is disconnected from meteorological and hydrological conditions, failing to establish a prediction model coupled with multi-source data; early warning information is limited to point-like danger areas, lacking the ability to spatially extrapolate the risk chain transmission path; and response mechanisms lag behind the disaster development process, making it difficult to achieve the transformation from passive emergency response to proactive prevention and control. Summary of the Invention

[0003] To address the aforementioned technical issues, this paper provides an early warning method for the evolution of reservoir drawdown zones based on time-series remote sensing imagery and intelligent AI. This technical solution resolves the problems mentioned above.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A reservoir drawdown zone evolution early warning method based on temporal remote sensing imagery and intelligent AI includes: S1. Based on the remote sensing image database, acquire remote sensing image data of the drawdown zones of several historical reservoirs, analyze the changing trends of the drawdown zones of several historical reservoirs, generate the changing trend vectors of the drawdown zones of several historical reservoirs, divide the changing trend vectors of the drawdown zones of several historical reservoirs according to the known development preferences of geological disaster risk events in the drawdown zones, and construct the entity-risk event relationship map of the changing trend of several reservoirs' drawdown zones. S2. Obtain remote sensing image data of the drawdown zone of the real-time target reservoir, and perform clustering according to the entity-risk event relationship map of the drawdown zone change trend of several reservoirs to obtain the entity-risk event to which the drawdown zone of the real-time target reservoir belongs. Construct an entity-risk event development path transmission tracking model to generate the entity-risk event development path to which the drawdown zone of the real-time target reservoir belongs. S3. Obtain meteorological data of the upstream and downstream of the target reservoir, predict the future environmental vectors of the upstream and downstream of the target reservoir, and dynamically compensate for the development path of the risk event of the entity belonging to the drawdown zone of the real-time target reservoir according to the prior knowledge of the environmental vector of the entity belonging to the drawdown zone of the real-time target reservoir. Generate the trigger probability of the risk event of the entity belonging to the drawdown zone of the future target reservoir, and generate the corresponding early warning scheme for the risk event.

[0005] Preferably, step S1 specifically includes: Based on the remote sensing image database, remote sensing image data of the drawdown zones of several historical reservoirs were acquired. Using the water body index method and threshold segmentation method, the instantaneous water body boundaries in the remote sensing image data of the drawdown zones of several historical reservoirs were extracted. Combined with water level data, the theoretical range of the drawdown zones of several historical reservoirs was determined. Based on the theoretical range of drawdown zones of several historical reservoirs, the data is gridded into fixed-size analysis units according to rules. A sliding window is used, with each unit of time as the observation window. The spectral features, texture features, and topographic features of each analysis unit are used as the observation objects to construct a temporal feature sequence of drawdown zones of several historical reservoirs. Based on the time-series feature sequences of drawdown zones of several historical reservoirs, a long short-term memory network is trained. By stacking multiple LSTM layers, long-term dependencies in the time-series feature sequences of drawdown zones of several historical reservoirs are extracted. The hidden state at the last time step of each stacked layer is substituted into a fully connected layer and mapped to a low-dimensional vector of fixed length to obtain the change trend vector of drawdown zones of several historical reservoirs.

[0006] Preferably, step S1 further includes: Based on the bi-column correlation coefficient, the correlation coefficient between each low-dimensional vector in the historical trend vector of the drawdown zone of several reservoirs and the known geological disaster risk events in the drawdown zone is verified. Substituted into PCA principal component analysis, several low-dimensional vectors in the historical trend vector of the drawdown zone of several reservoirs that are most correlated with the known geological disaster risks in the drawdown zone are selected. Based on the binary scatter plot, several low-latitude vectors that are most relevant to the geological disaster risk of known drawdown zones from the historical trend vectors of drawdown zones of several reservoirs are projected onto a two-dimensional plane to obtain a binary scatter plot of the historical trend vectors of drawdown zones of several reservoirs. Based on the binary scatter plot of the historical drawdown zone change trend vector of several reservoirs, the DBSCAN density clustering algorithm is used to cluster the binary scatter plot of the historical drawdown zone change trend vector according to the known geological disaster risk event development preference of the drawdown zone, and the nearest neighbor distance and minimum spatial unit number are preset. This results in a set of historical drawdown zone change trend entity clusters. By statistically analyzing the mean vector of the trend vector of all analytical units in the entity cluster set of historical drawdown zone changes of several reservoirs and their spatial topographic attributes, a physical profile of the historical drawdown zone changes of several reservoirs is established. Based on Bayesian causal networks, entity nodes are represented by historical images of the changing trends of the drawdown zones of several reservoirs, and risk event nodes are represented by known geological disaster risk events in the drawdown zones. Each entity node is marked as having contained an analysis unit that has experienced a known geological disaster risk event in the drawdown zone. Undirected edges between entity nodes and risk event nodes are constructed. Based on the entropy weight method, according to the undirected edges between entity nodes and risk event nodes, the proportion of analysis units in each entity node that have experienced known drawdown zone geological disaster risk events relative to the entity node is calculated, and the undirected edges between entity nodes and risk event nodes are assigned weights to obtain the weighted edges between entity nodes and risk event nodes. Based on cosine similarity, the similarity between entity nodes and risk event nodes in the weighted directed network is calculated, and all entity nodes with positive similarity are connected to obtain the similarity edges between entity nodes. Based on the weighted edges of entity nodes and risk event nodes, and the similarity edges of entity nodes, construct several entity-risk event relationship graphs of the changing trends of reservoir drawdown zones.

[0007] Preferably, step S2 specifically includes: Acquire real-time remote sensing image data of the drawdown zone of the target reservoir, and use the water index method and threshold segmentation method to extract the instantaneous water boundary in the real-time target reservoir drawdown zone remote sensing image data. Combined with water level data, determine the theoretical range of the real-time target drawdown zone. Based on the theoretical range of the real-time target drawdown zone, the data is gridded into fixed-size analysis units according to rules. A sliding window is used, with each unit of time as the observation window. The spectral features, texture features, and topographic features of each analysis unit are used as the observation objects to construct the temporal feature sequence of the real-time target reservoir drawdown zone. Substitute the real-time target reservoir drawdown zone temporal feature sequence into a long short-term memory network to generate a real-time target reservoir drawdown zone change trend vector; Based on the entity-risk event relationship map of several reservoir drawdown zone change trends, the Euclidean distance formula is used to calculate the Euclidean distance between the real-time target reservoir drawdown zone change trend vector and the center vector of each entity node. The entity-risk event relationship path of the drawdown zone change trend pointed to by the maximum Euclidean distance of the real-time target reservoir drawdown zone change trend vector is selected, and the entity-risk event to which the real-time target reservoir drawdown zone belongs is determined.

[0008] Preferably, step S2 further includes: Based on the entity-risk event relationship map of the drawdown zone change trend of several reservoirs, the entity-risk event sub-map of the drawdown zone of the target reservoir is extracted according to the entity-risk event to which the drawdown zone belongs; Based on the entity-risk event relationship graph of several reservoir drawdown zone change trends, a pre-trained R-GCN relationship graph convolutional network is used to construct an entity-risk event development path transmission tracking model. Taking the entity-risk event relationship edges of several reservoir drawdown zone change trends as input, the model learns to predict missing connection edges by constructing a masking graph of several reservoir drawdown zone change trends entity-risk event relationships, obtains the potential structure and relationship edge paths of the masking graph, and generates the development path of each entity-risk event in several reservoir drawdown zone change trends entity-risk event relationship graphs. Using the entity-risk event subgraph of the drawdown zone of the target reservoir, and substituting it into the entity-risk event development path propagation tracking model, the entity-risk event development path of the drawdown zone of the target reservoir is generated.

[0009] Preferably, step S3 specifically includes: Based on meteorological data from upstream and downstream of the target reservoir, a local environmental state prediction model for the target reservoir is constructed using an RNN recurrent neural network to predict the future environmental vectors upstream and downstream of the target reservoir. Based on the relationship map of several reservoir drawdown zone change trend entities and risk events, and using the Bayesian prior conditional probability formula, given the probability of risk event triggering when environmental data within the time window before the occurrence of a risk event occurs in the analysis unit of each historical reservoir drawdown zone change trend entity, a likelihood function is constructed to generate the Gaussian distribution probability density of environmental data within the time window before the occurrence of a risk event in the analysis unit of each historical reservoir drawdown zone change trend entity.

[0010] Preferably, step S3 further includes: Based on the development path of the target reservoir's drawdown zone attribution entity-risk event, the environmental evidence likelihood value of the target reservoir's drawdown zone attribution entity-risk event sensitive time window is calculated for the future upstream and downstream environmental vectors of the target reservoir. Based on the Gaussian distribution probability density of environmental data within the time window before the occurrence of risk events in the analysis unit of the historical drawdown zone change trend entity of each reservoir, and substituted into Logistic regression, the basic occurrence probability of the historical drawdown zone change trend entity-risk event development path is generated. Based on the probability of risk event triggering within the time window prior to a risk event in the analysis unit of the historical drawdown trend entity of each reservoir, the basic occurrence probability of the risk event development path of the historical drawdown trend entity of each reservoir, and the environmental evidence likelihood value of the target reservoir's drawdown entity-risk event sensitive time window, a posterior probability is constructed. Dynamic compensation is then performed on the real-time target reservoir's drawdown entity-risk event development path to generate the future target reservoir's drawdown entity-risk event triggering probability, as follows:

[0011] in, Let the drawdown zone of the target reservoir be assigned to the entity along the i-th development path - the probability of a risk event triggering. Given the historical drawdown trend of each reservoir entity, the probability of occurrence of the risk event is the basic occurrence probability of the i-th development path entity within the time window prior to the occurrence of a risk event in the analysis unit. Belongs to path Each entity-risk event pair is iterated over. The environmental evidence likelihood value for assigning the drawdown zone of the target reservoir to the entity along the i-th development path and the sensitive time window of the risk event. The local adjustment coefficient for the risk event of the entity along the i-th development path, representing the drawdown zone of the target reservoir.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a reservoir drawdown zone evolution early warning method based on time-series remote sensing imagery and intelligent AI. By constructing an intelligent knowledge graph that associates historical evolution patterns with geological hazards, it achieves automatic identification and accurate extraction of precursors to reservoir bank instability. The system intelligently matches real-time monitoring data with the knowledge graph, uses graph neural networks to deduce risk transmission paths consistent with geomechanical principles, and combines meteorological forecast data with a Bayesian dynamic update mechanism to generate probabilistic multi-scenario early warning schemes. This scheme significantly improves the objectivity and efficiency of risk identification, compressing the traditional manual monitoring cycle from weekly to hourly, breaking through the limitations of single-point threshold early warning, revealing the chain transmission process of risks, and providing highly action-oriented forward-looking decision support for reservoir safety management, realizing a shift from passive response to proactive prevention and control. Attached Figure Description

[0013] Figure 1 This is a flowchart of a method for early warning of reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI. Detailed Implementation

[0014] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0015] Reference Figure 1 As shown, the early warning method for reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI includes: S1. Based on the remote sensing image database, acquire remote sensing image data of the drawdown zones of several historical reservoirs, analyze the changing trends of the drawdown zones of several historical reservoirs, generate the changing trend vectors of the drawdown zones of several historical reservoirs, divide the changing trend vectors of the drawdown zones of several historical reservoirs according to the known development preferences of geological disaster risk events in the drawdown zones, and construct the entity-risk event relationship map of the changing trend of several reservoirs' drawdown zones. Step S1 specifically includes: Based on the remote sensing image database, remote sensing image data of the drawdown zones of several historical reservoirs were acquired. Using the water body index method and threshold segmentation method, the instantaneous water body boundaries in the remote sensing image data of the drawdown zones of several historical reservoirs were extracted. Combined with water level data, the theoretical range of the drawdown zones of several historical reservoirs was determined. Based on the theoretical range of drawdown zones of several historical reservoirs, the data is gridded into fixed-size analysis units according to rules. A sliding window is used, with each unit of time as the observation window. The spectral features, texture features, and topographic features of each analysis unit are used as the observation objects to construct a temporal feature sequence of drawdown zones of several historical reservoirs. Based on the time-series feature sequences of the drawdown zones of several historical reservoirs, a long short-term memory network is trained. By stacking multiple LSTM layers, the long-term dependencies in the time-series feature sequences of the drawdown zones of several historical reservoirs are extracted. The hidden state at the last time step of each stacked layer is substituted into a fully connected layer and mapped to a low-dimensional vector of fixed length to obtain the change trend vector of the drawdown zones of several historical reservoirs. Step S1 also includes: Based on the bi-column correlation coefficient, the correlation coefficient between each low-dimensional vector in the historical trend vector of the drawdown zone of several reservoirs and the known geological disaster risk events in the drawdown zone is verified. Substituted into PCA principal component analysis, several low-dimensional vectors in the historical trend vector of the drawdown zone of several reservoirs that are most correlated with the known geological disaster risks in the drawdown zone are selected. Based on the binary scatter plot, several low-latitude vectors that are most relevant to the geological disaster risk of known drawdown zones from the historical trend vectors of drawdown zones of several reservoirs are projected onto a two-dimensional plane to obtain a binary scatter plot of the historical trend vectors of drawdown zones of several reservoirs. Based on the binary scatter plot of the historical drawdown zone change trend vector of several reservoirs, the DBSCAN density clustering algorithm is used to cluster the binary scatter plot of the historical drawdown zone change trend vector according to the known geological disaster risk event development preference of the drawdown zone, and the nearest neighbor distance and minimum spatial unit number are preset. This results in a set of historical drawdown zone change trend entity clusters. By statistically analyzing the mean vector of the trend vector of all analytical units in the entity cluster set of historical drawdown zone changes of several reservoirs and their spatial topographic attributes, a physical profile of the historical drawdown zone changes of several reservoirs is established. Based on Bayesian causal networks, entity nodes are represented by historical images of the changing trends of the drawdown zones of several reservoirs, and risk event nodes are represented by known geological disaster risk events in the drawdown zones. Each entity node is marked as having contained an analysis unit that has experienced a known geological disaster risk event in the drawdown zone. Undirected edges between entity nodes and risk event nodes are constructed. Based on the entropy weight method, according to the undirected edges between entity nodes and risk event nodes, the proportion of analysis units in each entity node that have experienced known drawdown zone geological disaster risk events relative to the entity node is calculated, and the undirected edges between entity nodes and risk event nodes are assigned weights to obtain the weighted edges between entity nodes and risk event nodes. Based on cosine similarity, the similarity between entity nodes and risk event nodes in the weighted directed network is calculated, and all entity nodes with positive similarity are connected to obtain the similarity edges between entity nodes. Based on the weighted edges of entity nodes and risk event nodes and the similarity edges of entity nodes, construct several entity-risk event relationship graphs of the changing trends of reservoir drawdown zones; When using it, please refer to the steps outlined above: As a further development, an intelligent knowledge graph is constructed based on the correlation between historical evolution patterns and geological hazards. Long Short-Term Memory (LSTM) networks analyze long-term remote sensing image data to learn and compress land cover and topographic change patterns in the reservoir drawdown zone, generating fingerprint vectors representing unique evolutionary patterns. Density clustering combined with Bayesian causal networks classifies these evolution vectors into risk entity categories with engineering significance based on historical hazard event development preferences. This method quantitatively characterizes the causal correlation strength between entities and geological hazards, as well as the similarity between entities, forming a structured entity-risk event relationship graph. This enables the automatic extraction of precursory patterns of reservoir bank instability from massive amounts of remote sensing imagery, improving the objectivity and efficiency of risk identification.

[0016] S2. Obtain remote sensing image data of the drawdown zone of the target reservoir in real time, and perform clustering according to the entity-risk event relationship map of the drawdown zone change trend of several reservoirs to obtain the entity-risk event to which the drawdown zone of the target reservoir belongs in real time. Construct an entity-risk event development path transmission tracking model to generate the entity-risk event development path to which the drawdown zone of the target reservoir belongs in real time. Step S2 specifically includes: Acquire real-time remote sensing image data of the drawdown zone of the target reservoir, and use the water index method and threshold segmentation method to extract the instantaneous water boundary in the real-time target reservoir drawdown zone remote sensing image data. Combined with water level data, determine the theoretical range of the real-time target drawdown zone. Based on the theoretical range of the real-time target drawdown zone, the data is gridded into fixed-size analysis units according to rules. A sliding window is used, with each unit of time as the observation window. The spectral features, texture features, and topographic features of each analysis unit are used as the observation objects to construct the temporal feature sequence of the real-time target reservoir drawdown zone. Substitute the real-time target reservoir drawdown zone temporal feature sequence into a long short-term memory network to generate a real-time target reservoir drawdown zone change trend vector; Based on the entity-risk event relationship map of several reservoir drawdown zone change trends, the Euclidean distance formula is used to calculate the Euclidean distance between the real-time target reservoir drawdown zone change trend vector and the center vector of each entity node. The entity-risk event relationship path of the drawdown zone change trend pointed to by the maximum Euclidean distance of the real-time target reservoir drawdown zone change trend vector is selected, and the entity-risk event to which the real-time target reservoir drawdown zone belongs is determined. Step S2 also includes: Based on the entity-risk event relationship map of the drawdown zone change trend of several reservoirs, the entity-risk event sub-map of the drawdown zone of the target reservoir is extracted according to the entity-risk event to which the drawdown zone belongs; Based on the entity-risk event relationship graph of several reservoir drawdown zone change trends, a pre-trained R-GCN relationship graph convolutional network is used to construct an entity-risk event development path transmission tracking model. Taking the entity-risk event relationship edges of several reservoir drawdown zone change trends as input, the model learns to predict missing connection edges by constructing a masking graph of several reservoir drawdown zone change trends entity-risk event relationships, obtains the potential structure and relationship edge paths of the masking graph, and generates the development path of each entity-risk event in several reservoir drawdown zone change trends entity-risk event relationship graphs. Using the entity-risk event subgraph of the drawdown zone of the target reservoir, and substituting it into the entity-risk event development path propagation tracking model, the entity-risk event development path of the drawdown zone of the target reservoir is generated. When using it, please refer to the steps outlined above: As a further development, risk path deduction is based on knowledge graph matching and graph neural network inference. Real-time reservoir drawdown zone remote sensing data is processed by an LSTM network isomorphic to a historical knowledge base to extract features and generate standardized evolution feature vectors. The Euclidean distance algorithm calculates the spatial distance between the real-time vectors and the central vectors of each entity in the pre-built knowledge graph, achieving accurate matching of risk patterns. A relational graph convolutional network trained through occlusion learns complex causal relationships between nodes through self-supervised learning, performs context-aware inference on the matched entity subgraphs, and simulates multiple risk transmission paths conforming to geomechanical principles. This achieves automated mapping of massive real-time data to risk patterns, compressing the monitoring cycle from weekly to hourly. Graph neural network path deduction overcomes the limitations of single-point early warning, reveals the chain-like transmission mechanism of risks, outputs probabilistic multi-scenario early warning schemes, and forms reservoir safety decision support with interpretability and strong action orientation.

[0017] S3. Obtain meteorological data of the upstream and downstream of the target reservoir, predict the environmental vector of the upstream and downstream of the target reservoir in the future, and dynamically compensate for the development path of the risk event of the entity belonging to the drawdown zone of the target reservoir in the real time according to the prior knowledge of the environmental vector of the entity belonging to the drawdown zone of the target reservoir in the real time, generate the trigger probability of the risk event of the entity belonging to the drawdown zone of the target reservoir in the future, and generate the early warning plan for the corresponding risk event. Step S3 specifically includes: Based on meteorological data from upstream and downstream of the target reservoir, a local environmental state prediction model for the target reservoir is constructed using an RNN recurrent neural network to predict the future environmental vectors upstream and downstream of the target reservoir. Based on the relationship map of several reservoir drawdown zone change trend entities and risk events, using the Bayesian prior conditional probability formula, given the probability of risk event triggering when environmental data within the time window before the occurrence of risk events in the analysis unit of each historical reservoir drawdown zone change trend entity, a likelihood function is constructed to generate the Gaussian distribution probability density of environmental data within the time window before the occurrence of risk events in the analysis unit of each historical reservoir drawdown zone change trend entity. Step S3 also includes: Based on the development path of the target reservoir's drawdown zone attribution entity-risk event, the environmental evidence likelihood value of the target reservoir's drawdown zone attribution entity-risk event sensitive time window is calculated for the future upstream and downstream environmental vectors of the target reservoir. Based on the Gaussian distribution probability density of environmental data within the time window before the occurrence of risk events in the analysis unit of the historical drawdown zone change trend entity of each reservoir, and substituted into Logistic regression, the basic occurrence probability of the historical drawdown zone change trend entity-risk event development path is generated. Based on the probability of risk event triggering within the time window prior to a risk event in the analysis unit of the historical drawdown trend entity of each reservoir, the basic occurrence probability of the risk event development path of the historical drawdown trend entity of each reservoir, and the environmental evidence likelihood value of the target reservoir's drawdown entity-risk event sensitive time window, a posterior probability is constructed. Dynamic compensation is then performed on the real-time target reservoir's drawdown entity-risk event development path to generate the future target reservoir's drawdown entity-risk event triggering probability, as follows:

[0018] in, Let the drawdown zone of the target reservoir be assigned to the entity along the i-th development path - the probability of a risk event triggering. Given the historical drawdown trend of each reservoir entity, the probability of occurrence of the risk event is the basic occurrence probability of the i-th development path entity within the time window prior to the occurrence of a risk event in the analysis unit. Belongs to path Each entity-risk event pair is iterated over. The environmental evidence likelihood value for assigning the drawdown zone of the target reservoir to the entity along the i-th development path and the sensitive time window of the risk event. The local adjustment coefficient for the risk event of the entity along the i-th development path, representing the drawdown zone of the target reservoir.

[0019] When using it, please refer to the steps outlined above: As a further development, a Bayesian dynamically updated meteorological-geological coupled early warning decision-making mechanism is constructed. A recurrent neural network predicts future meteorological and hydrological environmental vectors, and this prediction serves as input for new evidence. Historical data analysis has established a priori knowledge base of environmental responses for entities evolving in different drawdown zones, quantified using a Gaussian probability density function. A logistic regression model generates the normal baseline probability of risk development paths. An improved Bayesian update formula dynamically compensates the likelihood values ​​of environmental evidence for specific sensitive time windows in the future to the baseline probability, calculating the posterior probability of each risk path under specific future weather scenarios. Based on probability thresholds, a tiered and targeted early warning action plan is automatically generated. This method achieves a shift from static state assessment to dynamic process prediction in early warning, coupling the forward-looking nature of meteorological forecasts with the regularity of geological risks, extending the early warning lead time, and providing quantifiable and operable decision-making basis through probabilistic output and multi-scenario extrapolation, transforming intelligent early warning into a forward-looking risk management capability.

[0020] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for early warning of reservoir drawdown zone evolution based on temporal remote sensing imagery and intelligent AI, characterized in that, include: S1. Based on the remote sensing image database, acquire remote sensing image data of the drawdown zones of several historical reservoirs, analyze the changing trends of the drawdown zones of several historical reservoirs, generate the changing trend vectors of the drawdown zones of several historical reservoirs, divide the changing trend vectors of the drawdown zones of several historical reservoirs according to the known development preferences of geological disaster risk events in the drawdown zones, and construct the entity-risk event relationship map of the changing trend of several reservoirs' drawdown zones. S2. Obtain remote sensing image data of the drawdown zone of the real-time target reservoir, and perform clustering according to the entity-risk event relationship map of the drawdown zone change trend of several reservoirs to obtain the entity-risk event to which the drawdown zone of the real-time target reservoir belongs. Construct an entity-risk event development path transmission tracking model to generate the entity-risk event development path to which the drawdown zone of the real-time target reservoir belongs. S3. Obtain meteorological data of the upstream and downstream of the target reservoir, predict the future environmental vectors of the upstream and downstream of the target reservoir, and dynamically compensate for the development path of the risk event of the entity belonging to the drawdown zone of the real-time target reservoir according to the prior knowledge of the environmental vector of the entity belonging to the drawdown zone of the real-time target reservoir. Generate the trigger probability of the risk event of the entity belonging to the drawdown zone of the future target reservoir, and generate the corresponding early warning scheme for the risk event.

2. The method for early warning of reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI according to claim 1, characterized in that, Step S1 specifically includes: Based on the remote sensing image database, remote sensing image data of the drawdown zones of several historical reservoirs were acquired. Using the water body index method and threshold segmentation method, the instantaneous water body boundaries in the remote sensing image data of the drawdown zones of several historical reservoirs were extracted. Combined with water level data, the theoretical range of the drawdown zones of several historical reservoirs was determined. Based on the theoretical range of drawdown zones of several historical reservoirs, the data is gridded into fixed-size analysis units according to rules. A sliding window is used, with each unit of time as the observation window. The spectral features, texture features, and topographic features of each analysis unit are used as the observation objects to construct a temporal feature sequence of drawdown zones of several historical reservoirs. Based on the time-series feature sequences of drawdown zones of several historical reservoirs, a long short-term memory network is trained. By stacking multiple LSTM layers, long-term dependencies in the time-series feature sequences of drawdown zones of several historical reservoirs are extracted. The hidden state at the last time step of each stacked layer is substituted into a fully connected layer and mapped to a low-dimensional vector of fixed length to obtain the change trend vector of drawdown zones of several historical reservoirs.

3. The method for early warning of reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI according to claim 2, characterized in that, Step S1 also includes: Based on the bi-column correlation coefficient, the correlation coefficient between each low-dimensional vector in the historical trend vector of the drawdown zone of several reservoirs and the known geological disaster risk events in the drawdown zone is verified. Substituted into PCA principal component analysis, several low-dimensional vectors in the historical trend vector of the drawdown zone of several reservoirs that are most correlated with the known geological disaster risks in the drawdown zone are selected. Based on the binary scatter plot, several low-latitude vectors that are most relevant to the geological disaster risk of known drawdown zones from the historical trend vectors of drawdown zones of several reservoirs are projected onto a two-dimensional plane to obtain a binary scatter plot of the historical trend vectors of drawdown zones of several reservoirs. Based on the binary scatter plot of the historical drawdown zone change trend vector of several reservoirs, the DBSCAN density clustering algorithm is used to cluster the binary scatter plot of the historical drawdown zone change trend vector according to the known geological disaster risk event development preference of the drawdown zone, and the nearest neighbor distance and minimum spatial unit number are preset. This results in a set of historical drawdown zone change trend entity clusters. By statistically analyzing the mean vector of the trend vector of all analytical units in the entity cluster set of historical drawdown zone changes of several reservoirs and their spatial topographic attributes, a physical profile of the historical drawdown zone changes of several reservoirs is established. Based on Bayesian causal networks, entity nodes are represented by historical images of the changing trends of the drawdown zones of several reservoirs, and risk event nodes are represented by known geological disaster risk events in the drawdown zones. Each entity node is marked as having contained an analysis unit that has experienced a known geological disaster risk event in the drawdown zone. Undirected edges between entity nodes and risk event nodes are constructed. Based on the entropy weight method, according to the undirected edges between entity nodes and risk event nodes, the proportion of analysis units in each entity node that have experienced known drawdown zone geological disaster risk events relative to the entity node is calculated, and the undirected edges between entity nodes and risk event nodes are assigned weights to obtain the weighted edges between entity nodes and risk event nodes. Based on cosine similarity, the similarity between entity nodes and risk event nodes in the weighted directed network is calculated, and all entity nodes with positive similarity are connected to obtain the similarity edges between entity nodes. Based on the weighted edges of entity nodes and risk event nodes, and the similarity edges of entity nodes, construct several entity-risk event relationship graphs of the changing trends of reservoir drawdown zones.

4. The method for early warning of reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI according to claim 3, characterized in that, Step S2 specifically includes: Acquire real-time remote sensing image data of the drawdown zone of the target reservoir, and use the water index method and threshold segmentation method to extract the instantaneous water boundary in the real-time target reservoir drawdown zone remote sensing image data. Combined with water level data, determine the theoretical range of the real-time target drawdown zone. Based on the theoretical range of the real-time target drawdown zone, the data is gridded into fixed-size analysis units according to rules. A sliding window is used, with each unit of time as the observation window. The spectral features, texture features, and topographic features of each analysis unit are used as the observation objects to construct the temporal feature sequence of the real-time target reservoir drawdown zone. Substitute the real-time target reservoir drawdown zone temporal feature sequence into a long short-term memory network to generate a real-time target reservoir drawdown zone change trend vector; Based on the entity-risk event relationship map of several reservoir drawdown zone change trends, the Euclidean distance formula is used to calculate the Euclidean distance between the real-time target reservoir drawdown zone change trend vector and the center vector of each entity node. The entity-risk event relationship path of the drawdown zone change trend pointed to by the maximum Euclidean distance of the real-time target reservoir drawdown zone change trend vector is selected, and the entity-risk event to which the real-time target reservoir drawdown zone belongs is determined.

5. The method for early warning of reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI according to claim 4, characterized in that, Step S2 also includes: Based on the entity-risk event relationship map of the drawdown zone change trend of several reservoirs, the entity-risk event sub-map of the drawdown zone of the target reservoir is extracted according to the entity-risk event to which the drawdown zone belongs; Based on the entity-risk event relationship graph of several reservoir drawdown zone change trends, a pre-trained R-GCN relationship graph convolutional network is used to construct an entity-risk event development path transmission tracking model. Taking the entity-risk event relationship edges of several reservoir drawdown zone change trends as input, the model learns to predict missing connection edges by constructing a masking graph of several reservoir drawdown zone change trends entity-risk event relationships, obtains the potential structure and relationship edge paths of the masking graph, and generates the development path of each entity-risk event in several reservoir drawdown zone change trends entity-risk event relationship graphs. Using the entity-risk event subgraph of the drawdown zone of the target reservoir, and substituting it into the entity-risk event development path propagation tracking model, the entity-risk event development path of the drawdown zone of the target reservoir is generated.

6. The method for early warning of reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI according to claim 5, characterized in that, Step S3 specifically includes: Based on meteorological data from upstream and downstream of the target reservoir, a local environmental state prediction model for the target reservoir is constructed using an RNN recurrent neural network to predict the future environmental vectors upstream and downstream of the target reservoir. Based on the relationship map of several reservoir drawdown zone change trend entities and risk events, and using the Bayesian prior conditional probability formula, given the probability of risk event triggering when environmental data within the time window before the occurrence of a risk event occurs in the analysis unit of each historical reservoir drawdown zone change trend entity, a likelihood function is constructed to generate the Gaussian distribution probability density of environmental data within the time window before the occurrence of a risk event in the analysis unit of each historical reservoir drawdown zone change trend entity.

7. The method for early warning of reservoir drawdown zone evolution based on time-series remote sensing imagery and intelligent AI according to claim 6, characterized in that, Step S3 also includes: Based on the development path of the target reservoir's drawdown zone attribution entity-risk event, the environmental evidence likelihood value of the target reservoir's drawdown zone attribution entity-risk event sensitive time window is calculated for the future upstream and downstream environmental vectors of the target reservoir. Based on the Gaussian distribution probability density of environmental data within the time window before the occurrence of risk events in the analysis unit of the historical drawdown zone change trend entity of each reservoir, and substituted into Logistic regression, the basic occurrence probability of the historical drawdown zone change trend entity-risk event development path is generated. Based on the probability of risk event triggering within the time window prior to a risk event in the analysis unit of the historical drawdown trend entity of each reservoir, the basic occurrence probability of the risk event development path of the historical drawdown trend entity of each reservoir, and the environmental evidence likelihood value of the target reservoir's drawdown entity-risk event sensitive time window, a posterior probability is constructed. Dynamic compensation is then performed on the real-time target reservoir's drawdown entity-risk event development path to generate the future target reservoir's drawdown entity-risk event triggering probability, as follows: ； in, Let the drawdown zone of the target reservoir be assigned to the entity along the i-th development path - the probability of a risk event triggering. Given the historical drawdown trend of each reservoir entity, the probability of occurrence of the risk event is the basic occurrence probability of the i-th development path entity within the time window prior to the occurrence of a risk event in the analysis unit. Belongs to path Each entity-risk event pair is iterated over. The environmental evidence likelihood value for assigning the drawdown zone of the target reservoir to the entity along the i-th development path and the sensitive time window of the risk event. The local adjustment coefficient for the risk event of the entity along the i-th development path, representing the drawdown zone of the target reservoir.

Images