A deep learning-based dynamic risk plot evaluation method and system

Abstract

The application discloses a kind of dynamic risk plot evaluation method and system based on deep learning, comprising the following steps: step one: obtain historical disaster data and real-time weather data;Step two: construct plot disturbance response tensor and plot relationship diagram, generate dynamic heterogeneous plot unit set;Step three: build weather-driven causal trigger path encoder, output causal guide feature sequence;Step four: generate plot-level cross-enhanced time series feature representation;Step five: based on the improved Chronos time series model, by constructing multi-scale time window and performing causal constraint low-rank decomposition, output plot risk prediction sequence and risk residual sequence;Step six: generate plot risk chain migration sequence through plot risk migration perception mechanism;Step seven: generate dynamic risk change curve.The application combines weather-driven causal trigger path encoder and improved Chronos time series model to realize dynamic evaluation of plot risk.

Description

Technical Field

[0001] This invention relates to the field of disaster risk assessment technology, and in particular to a method and system for dynamic risk assessment of land parcels based on deep learning. Background Technology

[0002] With the increasing frequency of extreme weather events and the accelerating pace of urbanization, the impact of geological and meteorological disaster risks on regional security and social operations is becoming increasingly prominent. Therefore, disaster risk assessment and dynamic prediction at the plot scale have attracted widespread attention. Existing plot risk assessment methods largely rely on historical disaster statistics or analysis of single meteorological indicators, typically limiting risk characterization to static or short-term prediction levels. In practical applications, these methods generally suffer from the following problems: Traditional methods often treat land parcels as independent analytical units, making risk assessments based solely on the historical disaster frequency or current meteorological conditions of a single parcel. This lacks systematic modeling of the time lag effects of multiple meteorological variables, making it difficult to accurately reflect the cumulative effects of meteorological disturbances over time and their varying impacts on disaster triggering. Existing technologies rarely consider both the spatial proximity and disaster response similarity between land parcels simultaneously. Land parcel grouping and risk analysis are often based on geographical location or empirical divisions, making it difficult to identify spatially discontinuous sets of land parcels with similar risk evolution characteristics. Meanwhile, in multi-plot scenarios, the migration and superposition effects of risks between different plots objectively exist. However, existing methods mostly remain at the level of empirical analysis in describing the risk transmission process between plots. They lack a quantitative identification mechanism that combines the predicted risk results with uncertainty information, making it difficult to accurately describe the chain reaction process of risks, thus limiting the reliability and practicality of dynamic risk assessment results.

[0003] Therefore, how to provide a dynamic risk assessment method and system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a dynamic risk assessment method and system for land parcels based on deep learning. This invention combines a weather-driven causal triggering path encoder and an improved Chronos time series model to achieve dynamic risk assessment of land parcels. By constructing a dynamic heterogeneous set of land parcel units and introducing causal triggering paths, the evolution mechanism of disaster risk driven by meteorology is structurally modeled. Based on this, causal guided feature sequences are used to cross-enhance geographic attributes and meteorological time series, and multi-scale time windows and causal-constrained low-rank decomposition are used to effectively extract the dominant features of risk evolution while retaining risk residual information. Furthermore, by combining a land parcel risk migration perception mechanism, the continuous transmission process of risk between land parcels is depicted, ultimately forming a dynamic risk change curve, thereby significantly improving the accuracy, temporal continuity, and interpretability of land parcel risk prediction.

[0005] A method and system for dynamic risk assessment of land parcels based on deep learning according to an embodiment of the present invention includes the following steps: Step 1: Acquire historical disaster data and real-time meteorological data, and establish corresponding historical disaster data sets and meteorological time series data sets for each plot of land; Step 2: Construct a land parcel disturbance response tensor based on the historical disaster data set and meteorological time series data set, and construct a land parcel relationship graph by combining the spatial connectivity relationship of land parcels and the similarity relationship of disaster response. Perform heterogeneous clustering on the land parcels to generate a dynamic heterogeneous land parcel unit set. Step 3: Taking dynamic heterogeneous land parcel units as the processing object, construct a weather-driven causal triggering path encoder based on the corresponding meteorological time series data set, generate the triggering path structure between meteorological variables and land parcels, and output the causal guidance feature sequence; Step 4: Encode the geographic attribute data and meteorological time series data of the land parcels, and calculate the gating coefficient based on the causal guided feature sequence to generate a land parcel-level cross-enhanced time series feature representation; Step 5: Input the plot-level cross-enhanced time series feature representation into the improved Chronos time series model, and output the plot risk prediction sequence and risk residual sequence by constructing a multi-scale time window and performing causal constraint low-rank decomposition. Step Six: Based on the land parcel risk prediction sequence and risk residual sequence, identify the risk chain reaction process between land parcels through the land parcel risk migration perception mechanism, and generate a land parcel risk chain migration sequence; Step 7: Based on the risk chain migration sequence of the land parcels, calculate the comprehensive risk value of each land parcel at each predicted time step, and generate a dynamic risk change curve.

[0006] Optionally, the historical disaster data specifically includes the time of disaster occurrence and the type of disaster, and the real-time meteorological data specifically includes precipitation, temperature, humidity and wind speed data collected in chronological order, and organized into a meteorological time series data set corresponding to each land parcel according to a uniform time interval.

[0007] Optionally, step two specifically includes: The historical disaster data of each plot are arranged in chronological order to establish a disaster status time series corresponding to the plot. Each time point in the disaster status time series corresponds to a binary identifier. The identifier is 1, which indicates that a disaster occurred at the corresponding time point, and the identifier is 0, which indicates that no disaster occurred at the corresponding time point. Based on the meteorological time series data set of each plot, taking the time point of each disaster as a reference, a time window of a set length is extracted forward, and the precipitation, temperature, humidity and wind speed data within the time window are extracted and arranged in chronological order to form a set of disturbed meteorological segments. The disturbance meteorological segments corresponding to all disaster events in the same plot are stacked on the event dimension to construct a disturbance response tensor. The disturbance response tensor includes three dimensions: the first dimension represents the number of disaster events, the second dimension represents the number of time steps, and the third dimension represents the meteorological variable dimension. The Euclidean distance between all plots is calculated based on their two-dimensional coordinates on the map. A spatial connectivity threshold is set. When the Euclidean distance between any two plots is less than the spatial connectivity threshold, the corresponding position in the spatial connectivity matrix is ​​assigned a value of 1; otherwise, it is assigned a value of 0. This results in the spatial connectivity matrix. For any two plots of land, the disturbance response tensors are standardized along the time dimension for each meteorological variable. The absolute difference of each variable and each time step is used as the basic difference value. The basic difference value is accumulated at all time steps and along the meteorological variable dimension to obtain the disturbance response difference degree value. The disturbance response difference degree value is input into a monotonically decreasing kernel function to convert the disturbance response difference degree value into a similarity value with a numerical range of 0 to 1, and a disaster response similarity matrix is ​​constructed. The spatial connectivity matrix is ​​added to the disaster response similarity matrix to obtain the joint weight matrix; Construct a land parcel relationship graph, which is an undirected weighted graph, where each node corresponds to a land parcel, and the weight of the edge is the corresponding element value in the joint weight matrix; The Louvain algorithm is used to cluster the nodes in the land parcel relationship graph, and the land parcels in each cluster subgraph are marked as the same class, generating a dynamic heterogeneous land parcel unit set.

[0008] Optionally, the Louvain algorithm is used to cluster the nodes in the land parcel relationship graph, marking the land parcels in each cluster subgraph as belonging to the same class, generating a dynamic heterogeneous land parcel unit set, specifically as follows: Each plot node in the plot relationship graph is initialized as an independent community, and each community initially contains only one plot node. The total weight of all edges in the entire plot relationship graph is recorded. For each node, iterate through all the different communities to which all adjacent nodes belong, and attempt to move the current node to the community of an adjacent node in turn. Calculate the modularity increment during each move. The calculation of the modularity increment includes the following steps: Obtain the sum of edge weights between the current node and all nodes in the target community, denoted as the local connection weight; Get the sum of the weights of all adjacent edges of the current node, denoted as the total edge weight of the node; Obtain the sum of edge weights of all nodes in the target community, and denote it as the target community edge weight; Add the target community edge weight to the total edge weight of the node, square it, and divide it by four times the total weight of all edges in the graph to obtain the structural contribution value after the move. Divide the square of the edge weight of the target community by four times the total weight of all edges in the graph to obtain the structural contribution value before the move; Subtract the structural contribution value before the move from the structural contribution value after the move to obtain the structural contribution difference; The modularity increment is obtained by dividing the local connection weight by twice the total weight of all edges in the graph and subtracting the structural contribution difference from the result. If the modularity increment is greater than zero, the current node will be moved to the target community that brings the largest modularity increment; if the modularity increment is less than or equal to zero, the current node will remain in its original community and will not be moved; if multiple target communities have the same modularity increment, one of them will be randomly selected for merging. Repeat the node movement process until the community division of all nodes no longer changes after one round of traversal, and obtain the first stage community division result; Based on the results of the first stage, each community is regarded as a new node. If there are multiple connecting edges between two communities in the original graph, the weights of all edges are added together as the edge weight between the two new nodes, and a new simplified plot relationship graph is constructed. The process of node movement and modularity increment calculation is repeated in the new simplified plot relationship graph. Community aggregation is performed iteratively until the community division results of all nodes in two consecutive rounds of traversal are consistent and the overall graph structure no longer changes, thus generating the final community division result. The land parcel nodes contained in each community in the final community division result are marked as the same category, resulting in multiple dynamic heterogeneous land parcel units, which constitute a dynamic heterogeneous land parcel unit set.

[0009] Optionally, step three specifically includes: Taking each dynamic heterogeneous plot unit as the processing object, the meteorological time series data sets of each plot within the same dynamic heterogeneous plot unit are aggregated, and the precipitation, temperature, humidity and wind speed data are time-aligned according to a unified time axis to obtain a multi-meteorological variable time series. For each meteorological variable in the multi-meteorological variable time series, the values ​​of the current time and several consecutive times before it are selected and arranged in chronological order to form a time lag sequence of the meteorological variable. For each time lag sequence of a meteorological variable, the numerical change between adjacent moments is calculated step by step, and the numerical change is accumulated within the time lag sequence to obtain the change intensity value of the meteorological variable under the corresponding lag condition. The change intensity value is aligned step by step with the disaster state time series of the corresponding plot within the dynamic heterogeneous plot unit. The time point when the disaster state changes from 0 to 1 is marked, and the change intensity value is extracted at the corresponding time point. The change intensity values ​​at all marked time points are accumulated to obtain the cumulative value of the impact of meteorological variables on the change of plot disaster state under the corresponding lag conditions. A weather-driven causal triggering path encoder is constructed, and a set of triggering path nodes is established. The set of triggering path nodes includes meteorological variable nodes, lag condition nodes, and plot nodes. The meteorological variable nodes correspond to precipitation, temperature, humidity, and wind speed, respectively. The lag condition nodes correspond to different lag step sizes, and the plot nodes correspond to different plots within a dynamic heterogeneous plot unit. The change intensity values ​​of the same meteorological variable obtained under different lag conditions are normalized, and the normalized change intensity values ​​are used as the first weight; the cumulative impact values ​​of the same meteorological variable obtained under different lag conditions are normalized, and the normalized cumulative impact values ​​are used as the second weight. Construct a first directed connection from meteorological variable nodes to lag condition nodes, and assign the corresponding first weight to the first directed connection weight; construct a second directed connection from lag condition nodes to land parcel nodes, and assign the corresponding second weight to the second directed connection weight; the first directed connection and the second directed connection are connected in series to form a trigger path structure between meteorological variables and land parcels. For all trigger path structures corresponding to the same plot node, calculate the path weight of each trigger path. The path weight is the product of the first directed connection weight and the second directed connection weight in the trigger path. The path weights corresponding to precipitation, temperature, humidity and wind speed are arranged in ascending order of time lag step size to obtain the causal guided feature sequence and output it.

[0010] Optionally, step four specifically includes: Obtain the geographic attribute data corresponding to the land parcel, including the land parcel's elevation, slope, and impermeable surface coverage ratio. Then, perform vectorization processing on the geographic attribute data to obtain the geographic attribute encoding result. The meteorological time series data corresponding to the plots are encoded according to time steps to obtain meteorological time series coding results; The summation and average of all elements in the causal guidance feature sequence are used to obtain the causal guidance value. The causal guidance value is then input into the Sigmoid mapping function to obtain a gating coefficient with a value range of 0 to 1. The gating coefficient is multiplied one by one with each element in the geographic attribute coding result to obtain the geographic attribute enhancement feature; the gating coefficient is multiplied one by one with each element corresponding to each time step in the meteorological time series coding result to obtain the meteorological time series enhancement feature. The geographic attribute enhancement features and the meteorological time series enhancement features are concatenated along the feature dimension to form a plot-level cross-enhanced time series feature representation.

[0011] Optionally, the improved Chronos time series model includes a time embedding module, a multi-scale time window construction module, a causal constraint low-rank decomposition module, a time series state update module, and a result output module. The plot-level cross-enhanced temporal feature representation is input into the temporal embedding module, and the feature vector of each time step is linearly mapped to obtain the temporal embedding sequence. In the multi-scale time window construction module, the time-embedded sequence is slidably segmented based on the set short-term, medium-term and long-term window lengths to obtain multiple time window sub-sequences with different time lengths. Each time window subsequence is input into the causal constraint low-rank decomposition module. Based on the causal guided feature sequence, the time lag step length corresponding to each time step in the time window subsequence is determined, and the path weights corresponding to each meteorological variable under the time lag step length are extracted. The path weights are summed and averaged to obtain the causal weight value corresponding to each time step. Features are extracted and low-rank decomposition is performed on time steps where the causal weight value is greater than the set weight threshold to obtain a causal-constrained low-rank principal component representation; The causal constraint low-rank principal component representations obtained under different time window scales are aligned and merged in the time dimension to form a multi-scale principal risk time series representation. The multi-scale master risk time series representation is input into the time series state update module. The time series state update module uses a recurrent neural network structure for recursive processing. Specifically, at the prediction start time, the land parcel risk state is initialized. At each time step, the multi-scale master risk time series representation corresponding to the current time step and the land parcel risk state of the previous time step are used as inputs. The land parcel risk state of the current time step is calculated by the state update unit. The land parcel risk state of the current time step is used as the previous land parcel risk state of the next time step. The state update of all time steps is completed in chronological order. The land parcel risk states obtained in each time step are arranged in chronological order to obtain the land parcel risk prediction sequence. In the causal constraint low-rank decomposition module, the original time window subsequence is subtracted element-wise from the corresponding causal constraint low-rank principal component representation to obtain the risk residual representation. The risk residuals obtained at each time window scale are aligned and merged in the time dimension to obtain the risk residual sequence corresponding to the land parcel risk prediction sequence. The result output module outputs the land parcel risk prediction sequence and the risk residual sequence, respectively.

[0012] Optionally, step six specifically includes: Using a dynamic heterogeneous land parcel set as the object of risk migration analysis, for each parcel within each dynamic heterogeneous land parcel unit, the land parcel risk prediction sequence and the corresponding risk residual sequence are obtained respectively; Within the same dynamic heterogeneous land parcel unit, the land parcel risk prediction sequence of each parcel is time-aligned according to the prediction time step. For any two different land parcels, when the risk prediction values ​​of the first land parcel and the second land parcel both show an upward change in adjacent prediction time steps, and the difference between the risk prediction value of the second land parcel in the next prediction time step and the risk prediction value of the first land parcel in the current prediction time step is greater than the set change threshold, it is recorded as a valid risk change association. Obtain the residual value in the risk residual sequence of the second plot at the corresponding time step. When the residual value is less than the set residual threshold, the effective risk change is determined to be associated with a stable risk change. When the residual value is greater than or equal to the set residual threshold, the effective risk change is determined to be associated with an unstable risk change and is removed. When the first plot and the second plot both meet the effective risk change association conditions and are both determined to be stable risk changes within a continuous period of no less than the set prediction time steps, it is determined that there is a risk migration relationship between the first plot and the second plot, and the first plot is marked as the risk source plot and the second plot is marked as the risk-bearing plot, thus forming a risk migration path from the risk source plot to the risk-bearing plot. Multiple risk migration paths formed within the same dynamic heterogeneous land parcel unit within the prediction time range are summarized and linked together according to the time sequence of risk migration and the connection relationship between land parcels to obtain a land parcel risk chain migration sequence that reflects the continuous risk transmission process between land parcels. This sequence is used to characterize the risk chain reaction process within the dynamic heterogeneous land parcel unit.

[0013] Optionally, step seven specifically includes: For each land parcel, at each prediction time step, the risk prediction value of the land parcel at the corresponding prediction time step in the land parcel risk prediction sequence is extracted and used as the benchmark risk value at the prediction time step. The risk migration path recorded in the risk chain migration sequence of the land parcel is traversed in the prediction time step. When a land parcel is marked as a risk-bearing land parcel in the prediction time step, the risk prediction values ​​of all risk source land parcels corresponding to the risk-bearing land parcel are extracted in the prediction time step. The predicted risk values ​​of all risk source plots are summed to obtain the migration risk value at the predicted time step. The migration risk value is then added to the baseline risk value to obtain the comprehensive risk value at the predicted time step. When a land parcel is not marked as a risk-bearing land parcel at the prediction time step, the benchmark risk value is used as the comprehensive risk value at the prediction time step. The comprehensive risk values ​​obtained at each prediction time step are arranged in chronological order to form a dynamic risk change curve corresponding to the time and the comprehensive risk value.

[0014] A method and system for dynamic risk assessment of land parcels based on deep learning according to an embodiment of the present invention includes the following modules: The data acquisition module is used to acquire historical disaster data and real-time meteorological data, and to establish a corresponding historical disaster data set and meteorological time series data set for each plot of land. The land parcel relationship construction and clustering module is used to construct a land parcel disturbance response tensor based on the historical disaster data set and meteorological time series data set, construct a land parcel relationship graph by combining the spatial connectivity relationship of land parcels and the similarity relationship of disaster response, and perform heterogeneous clustering of land parcels to generate a dynamic heterogeneous land parcel unit set; The weather-driven causal triggering path encoding module is used to construct a weather-driven causal triggering path encoder with dynamic heterogeneous plot units as the processing object, generate the triggering path structure between meteorological variables and plots, and output the causal guidance feature sequence. The cross-enhanced feature generation module is used to encode the geographic attribute data and meteorological time series data of the land parcels, and calculate the gating coefficient based on the causal guided feature sequence to generate a land parcel-level cross-enhanced time series feature representation. The risk time series prediction module is used to input the plot-level cross-enhanced time series feature representation into the improved Chronos time series model, and output the plot risk prediction sequence and risk residual sequence by constructing a multi-scale time window and performing causal constraint low-rank decomposition. The risk migration perception module is used to identify the risk chain reaction process between land parcels based on the land parcel risk prediction sequence and risk residual sequence, and generate a land parcel risk chain migration sequence through the land parcel risk migration perception mechanism. The risk result output module is used to calculate the comprehensive risk value of each plot at each prediction time step based on the plot risk chain migration sequence, and generate a dynamic risk change curve.

[0015] The beneficial effects of this invention are: This invention addresses the problems of static risk characterization, weak inter-plot relationships, and difficulty in quantifying risk transmission processes in existing land risk assessments. It introduces collaborative modeling of historical disaster data and meteorological time series data. At the data level, it constructs a land disturbance response tensor and integrates spatial connectivity and disaster response similarity relationships between land plots to form a land plot relationship map. The Louvain algorithm is then used to generate a dynamic set of heterogeneous land plot units. Structurally, this invention breaks through the limitation of relying solely on geographical proximity for land plot analysis and achieves effective aggregation of land plots with similar disaster response patterns.

[0016] In the causal modeling stage, this invention uses a weather-driven causal triggering path encoder to structure and encode the impact of multiple meteorological variables on the changes in the disaster state of land parcels under different time lags, generating a causal guided feature sequence that provides clear causal constraints for feature fusion and risk prediction. In the feature fusion stage, a gating coefficient is calculated based on the causal guided feature sequence, and cross-enhanced encoding is performed on geographic attribute data and meteorological time series data to effectively improve the sensitivity of land parcel-level feature representation to key risk drivers.

[0017] Furthermore, this invention introduces a multi-scale time window construction and a causal constraint low-rank decomposition mechanism into the improved Chronos time series model, enabling the model to focus on key time steps strongly correlated with risk evolution at different time scales, suppress noise interference, and simultaneously output land parcel risk prediction sequences and risk residual sequences. Combined with a land parcel risk migration perception mechanism, stable risk migration paths are identified within dynamic heterogeneous land parcel units, constructing a land parcel risk chain migration sequence, and finally generating a dynamic risk change curve reflecting the superposition effect of the land parcel's own risk evolution and risk transmission, thereby significantly improving the continuity, reliability, and interpretability of dynamic risk land parcel assessment results. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based dynamic risk assessment method for land parcels proposed in this invention. Figure 2 This is a schematic diagram of the structure of a dynamic risk assessment system for land parcels based on deep learning proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1A dynamic risk assessment method for land parcels based on deep learning includes the following steps: Step 1: Acquire historical disaster data and real-time meteorological data, and establish corresponding historical disaster data sets and meteorological time series data sets for each plot of land; Step 2: Construct a land parcel disturbance response tensor based on the historical disaster data set and meteorological time series data set, and construct a land parcel relationship graph by combining the spatial connectivity relationship of land parcels and the similarity relationship of disaster response. Perform heterogeneous clustering on the land parcels to generate a dynamic heterogeneous land parcel unit set. Step 3: Taking dynamic heterogeneous land parcel units as the processing object, construct a weather-driven causal triggering path encoder based on the corresponding meteorological time series data set, generate the triggering path structure between meteorological variables and land parcels, and output the causal guidance feature sequence; Step 4: Encode the geographic attribute data and meteorological time series data of the land parcels, and calculate the gating coefficient based on the causal guided feature sequence to generate a land parcel-level cross-enhanced time series feature representation; Step 5: Input the plot-level cross-enhanced time series feature representation into the improved Chronos time series model, and output the plot risk prediction sequence and risk residual sequence by constructing a multi-scale time window and performing causal constraint low-rank decomposition. Step Six: Based on the land parcel risk prediction sequence and risk residual sequence, identify the risk chain reaction process between land parcels through the land parcel risk migration perception mechanism, and generate a land parcel risk chain migration sequence; Step 7: Based on the risk chain migration sequence of the land parcels, calculate the comprehensive risk value of each land parcel at each predicted time step, and generate a dynamic risk change curve.

[0021] In this embodiment, the historical disaster data specifically includes the time of disaster occurrence and the type of disaster, and the real-time meteorological data specifically includes precipitation, temperature, humidity and wind speed data collected in chronological order, and is organized into a meteorological time series data set corresponding to each land parcel according to a uniform time interval.

[0022] In this embodiment, step two specifically includes: The historical disaster data of each plot are arranged in chronological order to establish a disaster status time series corresponding to the plot. Each time point in the disaster status time series corresponds to a binary identifier. The identifier is 1, which indicates that a disaster occurred at the corresponding time point, and the identifier is 0, which indicates that no disaster occurred at the corresponding time point. Based on the meteorological time series data set of each plot, taking the time point of each disaster as a reference, a time window of a set length is extracted forward, and the precipitation, temperature, humidity and wind speed data within the time window are extracted and arranged in chronological order to form a set of disturbed meteorological segments. The disturbance meteorological segments corresponding to all disaster events in the same plot are stacked on the event dimension to construct a disturbance response tensor. The disturbance response tensor includes three dimensions: the first dimension represents the number of disaster events, the second dimension represents the number of time steps, and the third dimension represents the meteorological variable dimension. The Euclidean distance between all plots is calculated based on their two-dimensional coordinates on the map. A spatial connectivity threshold is set. If the Euclidean distance between any two plots is less than the spatial connectivity threshold, the corresponding position in the spatial connectivity matrix is ​​assigned a value of 1; otherwise, it is assigned a value of 0. The spatial connectivity matrix is ​​obtained by setting the spatial connectivity threshold to 500 meters. For any two plots of land, the disturbance response tensors are standardized along the time dimension for each meteorological variable. The absolute difference of each variable and each time step is used as the basic difference value. The basic difference value is accumulated at all time steps and along the meteorological variable dimension to obtain the disturbance response difference degree value. The disturbance response difference degree value is input into a monotonically decreasing kernel function to convert the disturbance response difference degree value into a similarity value with a numerical range of 0 to 1, and a disaster response similarity matrix is ​​constructed. The kernel function is: ; in, This value represents the degree of difference in disturbance response between two plots. Represents an exponential function; The spatial connectivity matrix is ​​added to the disaster response similarity matrix to obtain the joint weight matrix; Construct a land parcel relationship graph, which is an undirected weighted graph, where each node corresponds to a land parcel, and the weight of the edge is the corresponding element value in the joint weight matrix; The Louvain algorithm is used to cluster the nodes in the land parcel relationship graph, and the land parcels in each cluster subgraph are marked as the same class, generating a dynamic heterogeneous land parcel unit set. This invention effectively improves the scientific rigor and robustness of land parcel clustering by constructing a disturbance response tensor and a land parcel relationship graph, combining spatial connectivity and disaster response similarity. In the construction of the disturbance response tensor, meteorological variables within the corresponding time window are extracted using the disaster occurrence point as the anchor point, comprehensively preserving the multivariate disturbance patterns before the disaster and achieving targeted modeling. Furthermore, through standardization and variable interpolation, the true degree of difference in disturbance response among multiple land parcels is obtained. Compared with traditional methods that only consider spatial location or meteorological data similarity, the proposed joint weight matrix integrates the spatial proximity of land parcels and the dynamic characteristics of disaster meteorological response. The constructed land parcel relationship graph has higher structural expressiveness and risk propagation semantics. By introducing the Louvain algorithm to achieve efficient graph clustering, multiple non-spatially continuous but dynamically heterogeneous land parcel units with common risk characteristics can be mined, improving the local sensitivity and overall generalization ability of risk identification.

[0023] The dynamic heterogeneous land parcel unit serves as an intermediate structural unit in this invention, used to establish stable structural constraints between land parcel-level risk assessment and meteorological time series modeling. By heterogeneously clustering land parcels based on historical disaster response characteristics and spatial connectivity, land parcels with similar disaster response patterns but spatial discontinuities are grouped into the same dynamic heterogeneous land parcel unit. This makes the modeling process no longer limited to geographical proximity but based on the consistency of disaster response behavior, thereby improving the rationality of land parcel grouping.

[0024] In the causal modeling stage, the dynamic heterogeneous plot unit limits the statistical range of the weather-driven causal triggering path encoder, so that the calculation of the intensity of meteorological variable changes, time lag effects, and disaster triggering relationships are all carried out within the same response mode hypothesis space, thereby improving the stability and consistency of causal feature encoding.

[0025] In the time series modeling and risk evolution analysis stage, dynamic heterogeneous land parcel units serve as mesoscale containers for sharing temporal structures and risk propagation constraints among land parcels. This enables the improved Chronos time series model and risk migration perception mechanism to learn and reason within structurally constrained land parcel sets, reducing model complexity and suppressing ineffective propagation between cross-mechanism land parcels, thereby enhancing the reliability and interpretability of dynamic risk assessment results.

[0026] In this embodiment, the Louvain algorithm is used to cluster nodes in the land parcel relationship graph, marking land parcels in each cluster subgraph as belonging to the same class, thereby generating a dynamic heterogeneous land parcel unit set. Specifically: Each plot node in the plot relationship graph is initialized as an independent community, and each community initially contains only one plot node. The total weight of all edges in the entire plot relationship graph is recorded. For each node, iterate through all the different communities to which all adjacent nodes belong, and attempt to move the current node to the community of an adjacent node in turn. Calculate the modularity increment during each move. The calculation of the modularity increment includes the following steps: Obtain the sum of edge weights between the current node and all nodes in the target community, denoted as the local connection weight; Get the sum of the weights of all adjacent edges of the current node, denoted as the total edge weight of the node; Obtain the sum of edge weights of all nodes in the target community, and denote it as the target community edge weight; Add the target community edge weight to the total edge weight of the node, square it, and divide it by four times the total weight of all edges in the graph to obtain the structural contribution value after the move. Divide the square of the edge weight of the target community by four times the total weight of all edges in the graph to obtain the structural contribution value before the move; Subtract the structural contribution value before the move from the structural contribution value after the move to obtain the structural contribution difference; The modularity increment is obtained by dividing the local connection weight by twice the total weight of all edges in the graph and subtracting the structural contribution difference from the result. If the modularity increment is greater than zero, the current node will be moved to the target community that brings the largest modularity increment; if the modularity increment is less than or equal to zero, the current node will remain in its original community and will not be moved; if multiple target communities have the same modularity increment, one of them will be randomly selected for merging. Repeat the node movement process until the community division of all nodes no longer changes after one round of traversal, and obtain the first stage community division result; Based on the results of the first stage, each community is regarded as a new node. If there are multiple connecting edges between two communities in the original graph, the weights of all edges are added together as the edge weight between the two new nodes, and a new simplified plot relationship graph is constructed. The process of node movement and modularity increment calculation is repeated in the new simplified plot relationship graph. Community aggregation is performed iteratively until the community division results of all nodes in two consecutive rounds of traversal are consistent and the overall graph structure no longer changes, thus generating the final community division result. The land parcel nodes contained in each community in the final community division result are marked as the same category, resulting in multiple dynamic heterogeneous land parcel units, which constitute a dynamic heterogeneous land parcel unit set. The Louvain algorithm-based land parcel relationship graph clustering method, by introducing refined calculation of modularity increment, achieves efficient community division of parcel nodes while preserving spatial proximity and disaster response similarity among parcels. By initializing each node as an independent community and performing modularity increment calculation, it can effectively determine whether node migration contributes to improving the compactness and stability of the overall graph structure. The judgment mechanism ensures that each community division evolves towards a better structure, thereby improving the clustering quality. Furthermore, by constructing a simplified land parcel relationship graph and performing iterative optimization, it can gradually merge structurally similar and functionally similar parcel units, achieving an approximation from local optimum to global optimum clustering results. The resulting multiple dynamic heterogeneous parcel units can more realistically reflect parcel groups with common meteorological driving causal relationships in disaster propagation paths, providing deep learning models with spatial-temporal modeling units with high cohesion and low coupling, improving the accuracy of parcel-level disaster risk prediction and model generalization ability.

[0027] In this embodiment, step three specifically includes: Taking each dynamic heterogeneous plot unit as the processing object, the meteorological time series data sets of each plot within the same dynamic heterogeneous plot unit are aggregated, and the precipitation, temperature, humidity and wind speed data are time-aligned according to a unified time axis to obtain a multi-meteorological variable time series. For each meteorological variable in the multi-meteorological variable time series, the values ​​of the current time and several consecutive times before it are selected and arranged in chronological order to form a time lag sequence of the meteorological variable. For each time lag sequence of a meteorological variable, the numerical change between adjacent moments is calculated step by step, and the numerical change is accumulated within the time lag sequence to obtain the change intensity value of the meteorological variable under the corresponding lag condition. The change intensity value is aligned step by step with the disaster state time series of the corresponding plot within the dynamic heterogeneous plot unit. The time point when the disaster state changes from 0 to 1 is marked, and the change intensity value is extracted at the corresponding time point. The change intensity values ​​at all marked time points are accumulated to obtain the cumulative value of the impact of meteorological variables on the change of plot disaster state under the corresponding lag conditions. A weather-driven causal triggering path encoder is constructed, and a set of triggering path nodes is established. The set of triggering path nodes includes meteorological variable nodes, lag condition nodes, and plot nodes. The meteorological variable nodes correspond to precipitation, temperature, humidity, and wind speed, respectively. The lag condition nodes correspond to different lag step sizes, and the plot nodes correspond to different plots within a dynamic heterogeneous plot unit. The change intensity values ​​of the same meteorological variable obtained under different lag conditions are normalized, and the normalized change intensity values ​​are used as the first weight; the cumulative impact values ​​of the same meteorological variable obtained under different lag conditions are normalized, and the normalized cumulative impact values ​​are used as the second weight. Construct a first directed connection from meteorological variable nodes to lag condition nodes, and assign the corresponding first weight to the first directed connection weight; construct a second directed connection from lag condition nodes to land parcel nodes, and assign the corresponding second weight to the second directed connection weight; the first directed connection and the second directed connection are connected in series to form a trigger path structure between meteorological variables and land parcels. For all trigger path structures corresponding to the same plot node, calculate the path weight of each trigger path. The path weight is the product of the first directed connection weight and the second directed connection weight in the trigger path. Arrange the path weights corresponding to precipitation, temperature, humidity and wind speed in order of increasing time lag step size to obtain the causal guided feature sequence and output it. In this invention, the weather-driven causal triggering path encoder adopts a structured path modeling approach to transform meteorological time series data into causal guided feature sequences that can be used for deep learning modeling. The encoder internally sets a triggering path node set, which consists of three types of nodes: meteorological variable nodes, lag condition nodes, and land parcel nodes, thereby forming a clear multi-layer node structure.

[0028] Regarding node connection relationships, the encoder adopts a hierarchical directed connection method to construct the trigger path structure. First, a first directed connection is established between the meteorological variable node and the lag condition node to describe the change behavior of the meteorological variable under a specific time lag condition. Then, a second directed connection is established between the lag condition node and the land parcel node to describe the correspondence between meteorological changes and changes in the disaster state of the land parcel under the lag condition. The two types of directed connections are connected in sequence to form a complete trigger path from the meteorological variable to the land parcel through time lag, structurally depicting the time transmission process of meteorological-driven disaster response.

[0029] Regarding weighting rules, the encoder assigns weight values ​​from different sources to connections at different levels. The first directed connection weight is obtained by normalizing the change intensity value of meteorological variables under corresponding lag conditions, and is used to characterize the relative strength of the changes in meteorological variables themselves. The second directed connection weight is obtained by normalizing the cumulative impact value of meteorological variables on the changes in the disaster state of the land parcel under corresponding lag conditions, and is used to characterize the actual triggering degree of meteorological changes on the disaster response of the land parcel. The path weight of the triggering path is obtained by multiplying the corresponding first directed connection weight and the second directed connection weight, and is output in order of time lag step size to form a causal guided feature sequence. The causal guided feature sequence comprehensively characterizes the causal triggering relationship of multiple meteorological variables on the changes in the disaster state of the land parcel under different time lag conditions, thereby realizing the structured encoding of the meteorological-time-land parcel causal relationship.

[0030] In this embodiment, step four specifically includes: Obtain the geographic attribute data corresponding to the land parcel, including the land parcel's elevation, slope, and impermeable surface coverage ratio. Then, perform vectorization processing on the geographic attribute data to obtain the geographic attribute encoding result. The meteorological time series data corresponding to the plots are encoded according to time steps to obtain meteorological time series coding results; The summation and average of all elements in the causal guidance feature sequence are used to obtain the causal guidance value. The causal guidance value is then input into the Sigmoid mapping function to obtain a gating coefficient with a value range of 0 to 1. The gating coefficient is multiplied one by one with each element in the geographic attribute coding result to obtain the geographic attribute enhancement feature; the gating coefficient is multiplied one by one with each element corresponding to each time step in the meteorological time series coding result to obtain the meteorological time series enhancement feature. The geographic attribute enhancement features and the meteorological time series enhancement features are concatenated along the feature dimension to form a plot-level cross-enhanced time series feature representation; This invention introduces causal guided feature sequences to achieve unified modulation and fusion modeling of geographic attribute information and meteorological time series information. By encoding geographic attributes such as elevation, slope, and impervious surface cover ratio, and participating in cross-gating processing together with meteorological time series encoding results, the static spatial characteristics and dynamic meteorological change characteristics of the land parcel are synergistically expressed in the same feature space. Using the gating coefficients generated by the causal guided feature sequences, the two types of encoding results are uniformly modulated in intensity, and the land parcel-level cross-enhanced time series feature representation formed by splicing feature dimensions provides a clearly structured and highly integrated input for the time series risk prediction model, which is conducive to improving the stability and temporal consistency of land parcel risk assessment results.

[0031] In this embodiment, the improved Chronos time series model includes a time embedding module, a multi-scale time window construction module, a causal constraint low-rank decomposition module, a time series state update module, and a result output module. The plot-level cross-enhanced temporal feature representation is input into the temporal embedding module, and the feature vector of each time step is linearly mapped to obtain the temporal embedding sequence. Specifically, for the feature vector corresponding to each time step, the original feature space is converted into a fixed-dimensional temporal embedding representation by using a linear mapping method with the same parameters, so that the features at different times have a consistent scale and structure in the same representation space. Through the temporal embedding process, it is beneficial for the temporal model to continuously model the features between time steps and improve the expression stability of the risk evolution process. In the multi-scale time window construction module, the time-embedded sequence is slidably segmented based on the set short-term, medium-term and long-term window lengths to obtain multiple time window sub-sequences with different time lengths. In this invention, the multi-scale time window construction module characterizes the risk evolution features at different time scales by sliding segmenting the time-embedded sequence. Specifically, the short-term window length is set to 7 time steps, the medium-term window length to 14 time steps, and the long-term window length to 30 time steps, and a single time step is used as the sliding step size to progressively segment the time-embedded sequence, resulting in multiple time window subsequences with different time lengths. By constructing short-term, medium-term, and long-term time windows simultaneously, the model can capture the local fluctuation features and long-term trends of risk in different time spans, thereby enhancing the comprehensive expressive ability of complex time-series risk patterns. Each time window subsequence is input into the causal constraint low-rank decomposition module. Based on the causal guided feature sequence, the time lag step length corresponding to each time step in the time window subsequence is determined, and the path weights corresponding to each meteorological variable under the time lag step length are extracted. The path weights are summed and averaged to obtain the causal weight value corresponding to each time step. Features are extracted and low-rank decomposition is performed on time steps where the causal weight value is greater than the set weight threshold to obtain a causal-constrained low-rank principal component representation; The causal constraint low-rank principal component representations obtained under different time window scales are aligned and merged in the time dimension to form a multi-scale principal risk time series representation. The multi-scale master risk time series representation is input into the time series state update module. The time series state update module uses a recurrent neural network structure for recursive processing. Specifically, at the prediction start time, the land parcel risk state is initialized. At each time step, the multi-scale master risk time series representation corresponding to the current time step and the land parcel risk state of the previous time step are used as inputs. The land parcel risk state of the current time step is calculated by the state update unit. The land parcel risk state of the current time step is used as the previous land parcel risk state of the next time step. The state update of all time steps is completed in chronological order. The land parcel risk states obtained in each time step are arranged in chronological order to obtain the land parcel risk prediction sequence. In the causal constraint low-rank decomposition module, the original time window subsequence is subtracted element-wise from the corresponding causal constraint low-rank principal component representation to obtain the risk residual representation. The risk residuals obtained at each time window scale are aligned and merged in the time dimension to obtain the risk residual sequence corresponding to the land parcel risk prediction sequence. In this invention, the risk residual representations obtained under different time window scales all originate from the same historical time axis. During generation, they all retain the corresponding time step index information. To achieve residual alignment, firstly, the time step number is used as a unified index to rearrange the risk residual representations under each time window scale, ensuring that the residuals corresponding to the same time step are located at the same time position. Subsequently, a merging operation is performed on the multi-scale risk residual representations at the same time step position. This merging operation uses an averaging method to obtain the comprehensive risk residual value corresponding to the time step. Through the above alignment and merging processes, a continuous risk residual sequence in the time dimension is formed. The risk residuals corresponding to the most recent historical time steps before the prediction start time are selected as the residual reference sequence, and the residual reference sequence is mapped sequentially to each prediction time step on the prediction time axis in chronological order. This shifts the historical residuals as a whole to the prediction time range in terms of time position. Through this time-shifting mapping method, a risk residual sequence corresponding to the land parcel risk prediction sequence in terms of the number and order of time steps is formed, thereby ensuring the consistency between the risk residuals and the prediction results in the time dimension. The result output module outputs the land parcel risk prediction sequence and the risk residual sequence, respectively. In this invention, the improved Chronos time series model inherits the basic technical framework of the existing Chronos time series model in terms of overall structure, which includes time embedding, time window modeling, state recursion, and result output. However, for the complex scenario of dynamic risk parcel assessment, the key processing mechanisms inside the model have been improved and enhanced.

[0032] A time embedding module is introduced on the input side to perform a unified linear mapping on the temporal feature representation of the plot-level cross-enhancement, so that features at different time steps have a consistent scale and representation structure in the same embedding space, ensuring the stability of multi-scale processing. This is a standardized extension of the time feature processing method of the existing Chronos time series model. By constructing a multi-scale time window construction module, short-term, medium-term and long-term time window subsequences are built simultaneously, enabling the model to characterize the evolution of risk at different time spans in parallel, retaining the advantages of multi-scale modeling of the Chronos model.

[0033] The core difference between this invention and existing Chronos time series models lies in the introduction of a causal constraint low-rank decomposition module in the multi-scale time window processing stage. This module does not directly perform low-rank decomposition on the complete time window, but rather filters time steps based on causal-guided feature sequences, extracting principal components only for key time steps with causal weights higher than a threshold. This explicitly introduces causal prior constraints between meteorological variables and land parcel risks during the low-rank modeling process, effectively suppressing noise information with low correlation to risk evolution. Through this causal constraint mechanism, the obtained multi-scale principal risk time series representation more centrally reflects the dominant factors of risk changes.

[0034] Furthermore, while retaining low-rank principal components for risk prediction, this invention simultaneously constructs a risk residual sequence and independently outputs the fluctuation information not explained by the principal components. This provides both risk trend prediction results and uncertainty characterization information at the result level. Combined with a time-series state update module in the form of a recurrent neural network, it achieves continuous recursive modeling of risk states. These improvements enable the model to have stronger interpretability, stability, and the ability to characterize abnormal fluctuations in complex time-series environments, significantly improving the reliability and practical value of dynamic risk land parcel assessment results.

[0035] In this embodiment, step six specifically includes: Using a dynamic heterogeneous land parcel set as the object of risk migration analysis, for each parcel within each dynamic heterogeneous land parcel unit, the land parcel risk prediction sequence and the corresponding risk residual sequence are obtained respectively; Within the same dynamic heterogeneous land parcel unit, the land parcel risk prediction sequence of each parcel is time-aligned according to the prediction time step. For any two different land parcels, when the risk prediction values ​​of the first land parcel and the second land parcel both show an upward change in adjacent prediction time steps, and the difference between the risk prediction value of the second land parcel in the next prediction time step and the risk prediction value of the first land parcel in the current prediction time step is greater than the set change threshold, it is recorded as a valid risk change association. Obtain the residual value in the risk residual sequence of the second plot at the corresponding time step. When the residual value is less than the set residual threshold, the effective risk change is determined to be associated with a stable risk change. When the residual value is greater than or equal to the set residual threshold, the effective risk change is determined to be associated with an unstable risk change and is removed. When the first plot and the second plot both meet the effective risk change association conditions and are both determined to be stable risk changes within a continuous period of no less than the set prediction time steps, it is determined that there is a risk migration relationship between the first plot and the second plot, and the first plot is marked as the risk source plot and the second plot is marked as the risk-bearing plot, thus forming a risk migration path from the risk source plot to the risk-bearing plot. Multiple risk migration paths formed within the same dynamic heterogeneous land parcel unit within the prediction time range are summarized and linked together according to the time sequence of risk migration and the connection relationship between land parcels to obtain a land parcel risk chain migration sequence that reflects the continuous risk transmission process between land parcels. This sequence is used to characterize the risk chain reaction process within the dynamic heterogeneous land parcel unit. In this invention, the land risk migration perception mechanism uses dynamic heterogeneous land parcel units as the basic spatial unit for risk transmission analysis, avoiding misjudgments between unrelated land parcels across units. For each dynamic heterogeneous land parcel unit, the land risk prediction sequence and corresponding risk residual sequence output by each parcel in the prediction stage are obtained. The risk prediction value is the land risk status value output by the time-series state update module at each prediction time step. To characterize the risk transmission process between land parcels, the risk prediction sequences of each parcel within the same dynamic heterogeneous land parcel unit are time-aligned according to the prediction time step, and adjacent prediction time steps are used as the basic time scale for risk transmission. When the risk prediction values ​​of the first and second land parcels both increase in adjacent prediction time steps, and the increment of the risk prediction value of the second land parcel in the next prediction time step relative to the risk prediction value of the first land parcel in the current prediction time step exceeds the change... When the threshold is 0.15, a valid risk change association is identified. Further, to suppress misjudgments caused by short-term noise, a risk residual stability constraint is introduced. The valid risk change association is only judged as a stable risk change when the risk residual value of the second plot in the corresponding time step is less than the residual threshold of 0.10. When the first plot and the second plot both meet the valid risk change association conditions and are both judged as stable risk changes for no less than 3 consecutive prediction time steps, a risk migration relationship is determined from the first plot to the second plot, and a risk migration path from the risk source plot to the risk-bearing plot is formed. By connecting multiple risk migration paths formed within the prediction time range in chronological order, a plot risk chain migration sequence reflecting the continuous transmission and amplification characteristics of risk within dynamic heterogeneous plot units is finally obtained, thereby effectively characterizing the risk chain reaction process between plots.

[0036] In this embodiment, step seven specifically includes: For each land parcel, at each prediction time step, the risk prediction value of the land parcel at the corresponding prediction time step in the land parcel risk prediction sequence is extracted and used as the benchmark risk value at the prediction time step. The risk migration path recorded in the risk chain migration sequence of the land parcel is traversed in the prediction time step. When a land parcel is marked as a risk-bearing land parcel in the prediction time step, the risk prediction values ​​of all risk source land parcels corresponding to the risk-bearing land parcel are extracted in the prediction time step. The predicted risk values ​​of all risk source plots are summed to obtain the migration risk value at the predicted time step. The migration risk value is then added to the baseline risk value to obtain the comprehensive risk value at the predicted time step. When a land parcel is not marked as a risk-bearing land parcel at the prediction time step, the benchmark risk value is used as the comprehensive risk value at the prediction time step. The comprehensive risk values ​​obtained at each prediction time step are arranged in chronological order to form a dynamic risk change curve corresponding to time and comprehensive risk value, which serves as the dynamic risk assessment result of the land parcel at a continuous time scale. In this invention, by introducing a land parcel risk chain migration sequence, the risk prediction results of a single land parcel are integrated in a time series, thereby elevating the independently predicted risk to a comprehensive risk that considers the impact of risk transmission. Specifically, for each land parcel, the risk prediction value for the corresponding prediction time step in the land parcel risk prediction sequence is first extracted at each prediction time step, serving as the baseline risk value for that prediction time step. This baseline risk value reflects the land parcel's own risk status without considering external influences. Subsequently, using the risk migration path recorded in the land parcel risk chain migration sequence, it is identified whether the land parcel appears as a risk-bearing land parcel within the corresponding prediction time step. When When a risk-bearing relationship exists, the predicted risk values ​​from multiple risk source plots are accumulated to form a migration risk value, which is then superimposed with the baseline risk value to characterize the comprehensive impact of risk transmission and superposition between plots. When a plot is not marked as a risk-bearing plot at a prediction time step, the baseline risk value is directly used as the comprehensive risk value at the prediction time step. By sorting the comprehensive risk values ​​at each prediction time step over time, a dynamic risk change curve is finally formed, enabling the assessment results to simultaneously reflect the risk evolution of the plot itself and its carrying effect in the risk chain reaction, thereby improving the continuity and authenticity of the dynamic risk plot assessment results.

[0037] refer to Figure 2 A dynamic risk assessment system for land parcels based on deep learning includes the following modules: The data acquisition module is used to acquire historical disaster data and real-time meteorological data, and to establish a corresponding historical disaster data set and meteorological time series data set for each plot of land. The land parcel relationship construction and clustering module is used to construct a land parcel disturbance response tensor based on the historical disaster data set and meteorological time series data set, construct a land parcel relationship graph by combining the spatial connectivity relationship of land parcels and the similarity relationship of disaster response, and perform heterogeneous clustering of land parcels to generate a dynamic heterogeneous land parcel unit set; The weather-driven causal triggering path encoding module is used to construct a weather-driven causal triggering path encoder with dynamic heterogeneous plot units as the processing object, generate the triggering path structure between meteorological variables and plots, and output the causal guidance feature sequence. The cross-enhanced feature generation module is used to encode the geographic attribute data and meteorological time series data of the land parcels, and calculate the gating coefficient based on the causal guided feature sequence to generate a land parcel-level cross-enhanced time series feature representation. The risk time series prediction module is used to input the plot-level cross-enhanced time series feature representation into the improved Chronos time series model, and output the plot risk prediction sequence and risk residual sequence by constructing a multi-scale time window and performing causal constraint low-rank decomposition. The risk migration perception module is used to identify the risk chain reaction process between land parcels based on the land parcel risk prediction sequence and risk residual sequence, and generate a land parcel risk chain migration sequence through the land parcel risk migration perception mechanism. The risk result output module is used to calculate the comprehensive risk value of each plot at each prediction time step based on the plot risk chain migration sequence, and generate a dynamic risk change curve.

[0038] Example 1: To verify the feasibility of this invention in practice, it was applied to a dynamic risk assessment task in a typical landslide-prone mountainous area. This area has complex terrain, densely distributed land parcels, and is significantly affected by meteorological factors such as rainfall, resulting in frequent small to medium-sized landslides throughout the year. Traditional land parcel risk assessment methods mostly rely on static modeling based on geographical proximity, ignoring the role of meteorological factors in the causal triggering relationships between different land parcels. This fails to accurately reflect the chain-like diffusion path of risks, leading to problems such as large blind spots in early warning, poor timeliness, and high false alarm rates.

[0039] In this embodiment, historical landslide disaster data for the target area over the past five years, along with corresponding hourly precipitation, temperature, humidity, and wind speed data from multiple sources, were first collected and processed. The data acquisition module of this invention was used to establish a plot-level historical disaster data set and a meteorological time series data set. Next, the perturbation response tensor was calculated using the plot relationship construction and clustering module. By fusing the spatial connectivity relationships between plots with disaster response similarity relationships, a plot relationship map was constructed and divided into multiple dynamic heterogeneous plot units using the Louvain algorithm. This successfully identified spatially dispersed plot groups with similar disaster response patterns.

[0040] In the weather-driven causal triggering path encoding module, the intensity values ​​and cumulative impact values ​​of precipitation, temperature, humidity, and wind speed under multiple time lag conditions are extracted to construct the triggering path structure and obtain the causal guided feature sequence. Through the cross-enhanced feature generation module, the causal guided feature sequence is used as a gating coefficient to cross-modulate the geographic attribute encoding results and meteorological time series encoding results, and outputs the plot-level cross-enhanced time series feature representation. In the risk time series prediction module, this invention uses an improved Chronos time series model to process the plot-level cross-enhanced time series feature representation. The improved Chronos time series model constructs multi-scale time windows and performs causal constraint low-rank decomposition, and simultaneously outputs the risk prediction sequence and the risk residual sequence, which effectively improves the robustness of the model to abnormal meteorological disturbances.

[0041] Table 1. Comparison of Risk Prediction Performance of Different Models As can be seen from the comparative data in Table 1 above, the improved Chronos time series model used in this invention shows a significant performance advantage over traditional methods in the dynamic risk parcel assessment task. Specifically, the improved Chronos time series model has a root mean square error of 0.152, lower than the original Chronos time series model's 0.174 and the traditional LSTM model's 0.201, indicating higher overall prediction accuracy. In terms of mean absolute error, the improved Chronos time series model has a mean absolute error of only 0.098, compared to the original Chronos time series model's 0.116 and the traditional LSTM model's 0.129, further validating its stronger error control capability. On the F1 score, a comprehensive evaluation indicator, the improved Chronos time series model reaches 0.871, significantly higher than the original Chronos time series model's 0.794 and the traditional LSTM model's 0.752, indicating better performance in correctly identifying high-risk plots and reducing the false negative rate. The data results demonstrate that the improved Chronos time series model not only optimizes the time series modeling structure but also effectively suppresses redundant interference factors by introducing causal constraint low-rank decomposition, greatly improving the model's accuracy and stability in actual risk assessment.

[0042] This embodiment comprehensively improves the accuracy and continuity of plot-level dynamic risk assessment by introducing a weather-driven causal triggering path encoder and an improved Chronos time series model. Addressing the challenge that traditional methods struggle to fully capture the multi-scale and lag-dependent impacts of meteorological variables on disaster risk evolution, this invention effectively uncovers key meteorological driving mechanisms through causal feature guidance and causal constraint low-rank decomposition, enhancing the model's interpretability. Simultaneously, it identifies risk chain reactions between plots through a plot risk migration perception mechanism, generating plot risk chain migration sequences. This ensures that the assessment results not only reflect the independent risk level of individual plots but also characterize the risk transmission and coupling effects between plots, thereby improving the model's adaptability and generalization ability in complex disaster evolution scenarios.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing dynamic risky land parcels based on deep learning, characterized in that, Includes the following steps: Step 1: Acquire historical disaster data and real-time meteorological data, and establish corresponding historical disaster data sets and meteorological time series data sets for each plot of land; Step 2: Construct a land parcel disturbance response tensor based on the historical disaster data set and meteorological time series data set, and construct a land parcel relationship graph by combining the spatial connectivity relationship of land parcels and the similarity relationship of disaster response. Perform heterogeneous clustering on the land parcels to generate a dynamic heterogeneous land parcel unit set. Step 3: Taking dynamic heterogeneous land parcel units as the processing object, construct a weather-driven causal triggering path encoder based on the corresponding meteorological time series data set, generate the triggering path structure between meteorological variables and land parcels, and output the causal guidance feature sequence; Step 4: Encode the geographic attribute data and meteorological time series data of the land parcels, and calculate the gating coefficient based on the causal guided feature sequence to generate a land parcel-level cross-enhanced time series feature representation; Step 5: Input the plot-level cross-enhanced time series feature representation into the improved Chronos time series model, and output the plot risk prediction sequence and risk residual sequence by constructing a multi-scale time window and performing causal constraint low-rank decomposition. Step Six: Based on the land parcel risk prediction sequence and risk residual sequence, identify the risk chain reaction process between land parcels through the land parcel risk migration perception mechanism, and generate a land parcel risk chain migration sequence; Step 7: Based on the risk chain migration sequence of the land parcels, calculate the comprehensive risk value of each land parcel at each prediction time step, and generate a dynamic risk change curve.

2. The method for dynamic risk assessment of land parcels based on deep learning according to claim 1, characterized in that, The historical disaster data specifically includes the time of disaster occurrence and the type of disaster. The real-time meteorological data specifically includes precipitation, temperature, humidity and wind speed data collected in chronological order, and is organized into a meteorological time series data set corresponding to each land parcel according to a uniform time interval.

3. The method for dynamic risk assessment of land parcels based on deep learning according to claim 1, characterized in that, Step two specifically involves: The historical disaster data of each plot are arranged in chronological order to establish a disaster status time series corresponding to the plot. Each time point in the disaster status time series corresponds to a binary identifier. The identifier is 1, which indicates that a disaster occurred at the corresponding time point, and the identifier is 0, which indicates that no disaster occurred at the corresponding time point. Based on the meteorological time series data set of each plot, taking the time point of each disaster as a reference, a time window of a set length is extracted forward, and the precipitation, temperature, humidity and wind speed data within the time window are extracted and arranged in chronological order to form a set of disturbed meteorological segments. The disturbance meteorological segments corresponding to all disaster events in the same plot are stacked on the event dimension to construct a disturbance response tensor. The disturbance response tensor includes three dimensions: the first dimension represents the number of disaster events, the second dimension represents the number of time steps, and the third dimension represents the meteorological variable dimension. The Euclidean distance between all plots is calculated based on their two-dimensional coordinates on the map. A spatial connectivity threshold is set. When the Euclidean distance between any two plots is less than the spatial connectivity threshold, the corresponding position in the spatial connectivity matrix is ​​assigned a value of 1; otherwise, it is assigned a value of 0. This results in the spatial connectivity matrix. For any two plots of land, the disturbance response tensors are standardized along the time dimension for each meteorological variable. The absolute difference of each variable and each time step is used as the basic difference value. The basic difference value is accumulated at all time steps and along the meteorological variable dimension to obtain the disturbance response difference degree value. The disturbance response difference degree value is input into a monotonically decreasing kernel function to convert the disturbance response difference degree value into a similarity value with a numerical range of 0 to 1, and a disaster response similarity matrix is ​​constructed. The spatial connectivity matrix is ​​added to the disaster response similarity matrix to obtain the joint weight matrix; Construct a land parcel relationship graph, which is an undirected weighted graph, where each node corresponds to a land parcel, and the weight of the edge is the corresponding element value in the joint weight matrix; The Louvain algorithm is used to cluster the nodes in the land parcel relationship graph, and the land parcels in each cluster subgraph are marked as the same class, generating a dynamic heterogeneous land parcel unit set.

4. The method for dynamic risk assessment of land parcels based on deep learning according to claim 3, characterized in that, The Louvain algorithm is used to cluster nodes in the land parcel relationship graph, marking land parcels in each cluster subgraph as belonging to the same class, generating a dynamic heterogeneous land parcel unit set, specifically as follows: Each plot node in the plot relationship graph is initialized as an independent community, and each community initially contains only one plot node. The total weight of all edges in the entire plot relationship graph is recorded. For each node, iterate through all the different communities to which all adjacent nodes belong, and attempt to move the current node to the community of an adjacent node in turn. Calculate the modularity increment during each move. The calculation of the modularity increment includes the following steps: Obtain the sum of edge weights between the current node and all nodes in the target community, denoted as the local connection weight; Get the sum of the weights of all adjacent edges of the current node, denoted as the total edge weight of the node; Obtain the sum of edge weights of all nodes in the target community, and denote it as the target community edge weight; Add the target community edge weight to the total edge weight of the node, square it, and divide it by four times the total weight of all edges in the graph to obtain the structural contribution value after the move. Divide the square of the edge weight of the target community by four times the total weight of all edges in the graph to obtain the structural contribution value before the move; Subtract the structural contribution value before the move from the structural contribution value after the move to obtain the structural contribution difference; The modularity increment is obtained by dividing the local connection weight by twice the total weight of all edges in the graph and subtracting the structural contribution difference from the result. If the modularity increment is greater than zero, the current node will be moved to the target community that brings the largest modularity increment; if the modularity increment is less than or equal to zero, the current node will remain in its original community and will not be moved; if multiple target communities have the same modularity increment, one of them will be randomly selected for merging. Repeat the node movement process until the community division of all nodes no longer changes after one round of traversal, and obtain the first stage community division result; Based on the results of the first stage, each community is regarded as a new node. If there are multiple connecting edges between two communities in the original graph, the weights of all edges are added together as the edge weight between the two new nodes, and a new simplified plot relationship graph is constructed. The process of node movement and modularity increment calculation is repeated in the new simplified plot relationship graph. Community aggregation is performed iteratively until the community division results of all nodes in two consecutive rounds of traversal are consistent and the overall graph structure no longer changes, thus generating the final community division result. The land parcel nodes contained in each community in the final community division result are marked as the same category, resulting in multiple dynamic heterogeneous land parcel units, which constitute a dynamic heterogeneous land parcel unit set.

5. The method for dynamic risk assessment of land parcels based on deep learning according to claim 1, characterized in that, Step three specifically involves: Taking each dynamic heterogeneous plot unit as the processing object, the meteorological time series data sets of each plot within the same dynamic heterogeneous plot unit are aggregated, and the precipitation, temperature, humidity and wind speed data are time-aligned according to a unified time axis to obtain a multi-meteorological variable time series. For each meteorological variable in the multi-meteorological variable time series, the values ​​of the current time and several consecutive times before it are selected and arranged in chronological order to form a time lag sequence of the meteorological variable. For each time lag sequence of a meteorological variable, the numerical change between adjacent moments is calculated step by step, and the numerical change is accumulated within the time lag sequence to obtain the change intensity value of the meteorological variable under the corresponding lag condition. The change intensity value is aligned step by step with the disaster state time series of the corresponding plot within the dynamic heterogeneous plot unit. The time point when the disaster state changes from 0 to 1 is marked, and the change intensity value is extracted at the corresponding time point. The change intensity values ​​at all marked time points are accumulated to obtain the cumulative value of the impact of meteorological variables on the change of plot disaster state under the corresponding lag conditions. A weather-driven causal triggering path encoder is constructed, and a set of triggering path nodes is established. The set of triggering path nodes includes meteorological variable nodes, lag condition nodes, and plot nodes. The meteorological variable nodes correspond to precipitation, temperature, humidity, and wind speed, respectively. The lag condition nodes correspond to different lag step sizes, and the plot nodes correspond to different plots within a dynamic heterogeneous plot unit. The change intensity values ​​of the same meteorological variable obtained under different lag conditions are normalized, and the normalized change intensity values ​​are used as the first weight. The cumulative impact values ​​of the same meteorological variable obtained under different lag conditions are normalized, and the normalized cumulative impact value is used as the second weight. Construct a first directed connection from meteorological variable nodes to lag condition nodes, and assign the corresponding first weight to the first directed connection weight; construct a second directed connection from lag condition nodes to land parcel nodes, and assign the corresponding second weight to the second directed connection weight; the first directed connection and the second directed connection are connected in series to form a trigger path structure between meteorological variables and land parcels. For all trigger path structures corresponding to the same plot node, calculate the path weight of each trigger path. The path weight is the product of the first directed connection weight and the second directed connection weight in the trigger path. The path weights corresponding to precipitation, temperature, humidity and wind speed are arranged in ascending order of time lag step size to obtain the causal guided feature sequence and output it.

6. The method for dynamic risk assessment of land parcels based on deep learning according to claim 1, characterized in that, Step four specifically involves: Obtain the geographic attribute data corresponding to the land parcel, including the land parcel's elevation, slope, and impermeable surface coverage ratio. Then, perform vectorization processing on the geographic attribute data to obtain the geographic attribute encoding result. The meteorological time series data corresponding to the plots are encoded according to time steps to obtain the meteorological time series coding results; The summation and average of all elements in the causal guidance feature sequence are used to obtain the causal guidance value. The causal guidance value is then input into the Sigmoid mapping function to obtain a gating coefficient with a value range of 0 to 1. The gating coefficient is multiplied one by one with each element in the geographic attribute coding result to obtain the geographic attribute enhancement feature; The gating coefficient is multiplied one by one with each element corresponding to each time step in the meteorological time series coding result to obtain the meteorological time series enhancement feature; The geographic attribute enhancement features and the meteorological time series enhancement features are concatenated along the feature dimension to form a plot-level cross-enhanced time series feature representation.

7. The method for dynamic risk assessment of land parcels based on deep learning according to claim 1, characterized in that, The improved Chronos time series model includes a time embedding module, a multi-scale time window construction module, a causal constraint low-rank decomposition module, a time series state update module, and a result output module. The plot-level cross-enhanced temporal feature representation is input into the temporal embedding module, and the feature vector of each time step is linearly mapped to obtain the temporal embedding sequence. In the multi-scale time window construction module, the time-embedded sequence is slidably segmented based on the set short-term, medium-term and long-term window lengths to obtain multiple time window sub-sequences with different time lengths. Each time window subsequence is input into the causal constraint low-rank decomposition module. Based on the causal guided feature sequence, the time lag step length corresponding to each time step in the time window subsequence is determined, and the path weights corresponding to each meteorological variable under the time lag step length are extracted. The path weights are summed and averaged to obtain the causal weight value corresponding to each time step. Features are extracted and low-rank decomposition is performed on time steps where the causal weight value is greater than the set weight threshold to obtain a causal-constrained low-rank principal component representation; The causal constraint low-rank principal component representations obtained under different time window scales are aligned and merged in the time dimension to form a multi-scale principal risk time series representation. The multi-scale master risk time series representation is input into the time series state update module. The time series state update module uses a recurrent neural network structure for recursive processing. Specifically, at the prediction start time, the land parcel risk state is initialized. At each time step, the multi-scale master risk time series representation corresponding to the current time step and the land parcel risk state of the previous time step are used as inputs. The land parcel risk state of the current time step is calculated by the state update unit. The land parcel risk state of the current time step is used as the previous land parcel risk state of the next time step. The state update of all time steps is completed in chronological order. The land parcel risk states obtained in each time step are arranged in chronological order to obtain the land parcel risk prediction sequence. In the causal constraint low-rank decomposition module, the original time window subsequence is subtracted element-wise from the corresponding causal constraint low-rank principal component representation to obtain the risk residual representation. The risk residuals obtained at each time window scale are aligned and merged in the time dimension to obtain the risk residual sequence corresponding to the land parcel risk prediction sequence. The result output module outputs the land parcel risk prediction sequence and the risk residual sequence, respectively.

8. The method for dynamic risk assessment of land parcels based on deep learning according to claim 1, characterized in that, Step six specifically involves: Using a dynamic heterogeneous land parcel set as the object of risk migration analysis, for each parcel within each dynamic heterogeneous land parcel unit, the land parcel risk prediction sequence and the corresponding risk residual sequence are obtained respectively; Within the same dynamic heterogeneous land parcel unit, the land parcel risk prediction sequence of each parcel is time-aligned according to the prediction time step. For any two different land parcels, when the risk prediction values ​​of the first land parcel and the second land parcel both show an upward change in adjacent prediction time steps, and the difference between the risk prediction value of the second land parcel in the next prediction time step and the risk prediction value of the first land parcel in the current prediction time step is greater than the set change threshold, it is recorded as a valid risk change association. Obtain the residual value in the risk residual sequence of the second plot at the corresponding time step. When the residual value is less than the set residual threshold, determine that the effective risk change is associated with the stable risk change. When the residual value is greater than or equal to the set residual threshold, the effective risk change is determined to be associated with an unstable risk change and is removed. When the first plot and the second plot both meet the effective risk change association conditions and are both determined to be stable risk changes within a continuous period of no less than the set prediction time steps, it is determined that there is a risk migration relationship between the first plot and the second plot, and the first plot is marked as the risk source plot and the second plot is marked as the risk-bearing plot, thus forming a risk migration path from the risk source plot to the risk-bearing plot. Multiple risk migration paths formed within the same dynamic heterogeneous land parcel unit within the prediction time range are summarized and linked together according to the time sequence of risk migration and the connection relationship between land parcels to obtain a land parcel risk chain migration sequence that reflects the continuous risk transmission process between land parcels. This sequence is used to characterize the risk chain reaction process within the dynamic heterogeneous land parcel unit.

9. The method for dynamic risk assessment of land parcels based on deep learning according to claim 1, characterized in that, Step seven specifically involves: For each land parcel, at each prediction time step, the risk prediction value of the land parcel at the corresponding prediction time step in the land parcel risk prediction sequence is extracted and used as the benchmark risk value at the prediction time step. The risk migration path recorded in the risk chain migration sequence of the land parcel is traversed in the prediction time step. When a land parcel is marked as a risk-bearing land parcel in the prediction time step, the risk prediction values ​​of all risk source land parcels corresponding to the risk-bearing land parcel are extracted in the prediction time step. The predicted risk values ​​of all risk source plots are summed to obtain the migration risk value at the predicted time step. The migration risk value is then added to the baseline risk value to obtain the comprehensive risk value at the predicted time step. When a land parcel is not marked as a risk-bearing land parcel at the prediction time step, the benchmark risk value is used as the comprehensive risk value at the prediction time step. The comprehensive risk values ​​obtained at each prediction time step are arranged in chronological order to form a dynamic risk change curve corresponding to the time and the comprehensive risk value.

10. A deep learning-based dynamic risk land parcel assessment system, implementing the deep learning-based dynamic risk land parcel assessment method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The data acquisition module is used to acquire historical disaster data and real-time meteorological data, and to establish a corresponding historical disaster data set and meteorological time series data set for each plot of land. The land parcel relationship construction and clustering module is used to construct a land parcel disturbance response tensor based on the historical disaster data set and meteorological time series data set, construct a land parcel relationship graph by combining the spatial connectivity relationship of land parcels and the similarity relationship of disaster response, and perform heterogeneous clustering of land parcels to generate a dynamic heterogeneous land parcel unit set; The weather-driven causal triggering path encoding module is used to construct a weather-driven causal triggering path encoder with dynamic heterogeneous plot units as the processing object, generate the triggering path structure between meteorological variables and plots, and output the causal guidance feature sequence. The cross-enhanced feature generation module is used to encode the geographic attribute data and meteorological time series data of the land parcels, and calculate the gating coefficient based on the causal guided feature sequence to generate a land parcel-level cross-enhanced time series feature representation. The risk time series prediction module is used to input the plot-level cross-enhanced time series feature representation into the improved Chronos time series model, and output the plot risk prediction sequence and risk residual sequence by constructing a multi-scale time window and performing causal constraint low-rank decomposition. The risk migration perception module is used to identify the risk chain reaction process between land parcels based on the land parcel risk prediction sequence and risk residual sequence, and generate a land parcel risk chain migration sequence through the land parcel risk migration perception mechanism. The risk result output module is used to calculate the comprehensive risk value of each plot at each prediction time step based on the plot risk chain migration sequence, and generate a dynamic risk change curve.

Images