A climate risk zoning method based on complex network
By constructing a heterogeneous weighted complex network and a spatiotemporal convolutional graph neural network, the contradiction between high accuracy and low cost in traditional climate risk zoning methods is resolved, realizing dynamic and lightweight climate risk zoning and improving the spatiotemporal accuracy and adaptability of the zoning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA INST OF WATER RESOURCES & HYDROPOWER RES
- Filing Date
- 2025-10-16
- Publication Date
- 2026-06-26
AI Technical Summary
Existing climate risk zoning methods cannot achieve dynamic updates with high accuracy and low cost, and suffer from problems such as spatiotemporal misalignment and high computational costs.
A complex network-based approach is adopted, which constructs a heterogeneous weighted complex network, combines temporal variation and adaptive threshold freezing mechanism for lightweight updates, and utilizes cross-scale hierarchical community detection algorithm and spatiotemporal convolutional graph neural network for transfer learning to output the final climate risk zoning results.
It achieves high-resolution, interpretable, and robust climate risk zoning, improving spatiotemporal accuracy and boundary accuracy while reducing computational costs and timeliness requirements.
Smart Images
Figure CN121303840B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of meteorological disaster risk assessment technology, and in particular relates to a climate risk zoning method based on complex networks. Background Technology
[0002] Against the backdrop of global warming and frequent extreme events, there is an urgent need for high-resolution, dynamic, and sustainable zoning of regional climate risks.
[0003] However, existing solutions generally follow the approach of "single-hazard statistics + manual overlay": first, administrative boundaries or regular grids are used as fixed units, and then climate model outputs, remote sensing images, and socio-economic data from different sources are simply overlaid. Finally, risk levels are given through expert scoring or static weighting. Due to significant differences in coordinate systems, temporal resolution, and dimensions of various raw data, spatiotemporal misalignments frequently occur on the "same map"; regular grids ignore real terrain and population and economic distribution, resulting in huge heterogeneity within units; more challenging is that when new observations or disaster records appear, the traditional process often has to re-run the entire modeling chain, resulting in high computational and human costs, and the results are released after a delay compared to actual needs.
[0004] Therefore, how to break down the barriers between heterogeneous data, achieve multi-dimensional coupled characterization at the raster scale, and complete lightweight dynamic updates and reliable version management under the premise of controllable accuracy has become a key bottleneck that urgently needs to be overcome in the field of climate risk zoning. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides a climate risk zoning method based on complex networks.
[0006] This invention proposes a climate risk zoning method based on complex networks, comprising the following steps:
[0007] Obtain raw spatiotemporal data of climate-causing factors, exposure of disaster-bearing bodies, and vulnerability indicators within the area to be studied;
[0008] The original spatiotemporal data is subjected to homogenization preprocessing to generate a standardized raster dataset;
[0009] Using grid cells as nodes, a heterogeneous weighted complex network is constructed by comprehensively considering the spatial correlation of climate disaster-causing factors, the intensity of interaction between disaster-bearing bodies, and the uncertainty of node attributes.
[0010] Based on the temporal changes in the edge weights between nodes, an adaptive threshold-freeze mechanism is executed to update the heterogeneous weighted complex network.
[0011] Based on the updated heterogeneous weighted complex network, a cross-scale hierarchical community detection algorithm is run to obtain the initial climate risk zoning.
[0012] The initial climate risk zoning is input into a spatiotemporal convolutional graph neural network, and transfer learning and fine-tuning are performed using a zoning-loss coupling loss function to output the final climate risk zoning results.
[0013] Optionally, the process of performing homogenization preprocessing on the original spatiotemporal data to generate a standardized raster dataset includes:
[0014] The original spatiotemporal data is time-stamp aligned and spatial reference frame transformed; each layer is resampled using bilinear interpolation based on a 1km×1km square grid; missing values are marked on the resampled grid values and filled with spatiotemporal kriging; then the filled grid values are standardized dimensionlessly using the Z-score method, and finally a standardized grid dataset is output.
[0015] Optionally, using grid cells as nodes, and comprehensively considering the spatial correlation of climate-causing factors, the intensity of interactions between disaster-bearing bodies, and the uncertainty of node attributes, the process of constructing a heterogeneous weighted complex network includes:
[0016] Based on a standardized raster dataset, each raster cell is defined as a network node;
[0017] The spatial correlation of climate-causing factors among nodes is calculated to obtain the first weighted component.
[0018] Calculate the interaction strength between disaster-bearing bodies between nodes to obtain the second weight component;
[0019] The uncertainty of each node's attributes is evaluated to obtain the third weight component;
[0020] The first weight component, the second weight component, and the third weight component are weighted and summed according to a preset weight coefficient to generate the edge weights between nodes.
[0021] Construct a heterogeneous weighted complex network based on the node and edge weights.
[0022] Optionally, the process of updating the heterogeneous weighted complex network by implementing an adaptive threshold-freeze mechanism based on the temporal changes in the edge weights between nodes includes:
[0023] Within a continuous time window, calculate the relative rate of change of the current weight of each edge compared to the weight of the previous time window;
[0024] The relative rate of change is compared with a preset adaptive threshold. If it is lower than the threshold, the edge is marked as static and its weight is frozen. If it is not lower than the threshold, it is retained as a dynamic edge.
[0025] Only the weights of dynamic edges are recalculated and the network topology is updated, while the weights and topology of static edges remain unchanged, thus completing the update of heterogeneous weighted complex networks.
[0026] Optionally, the process of obtaining the initial climate risk zoning by running a cross-scale hierarchical community detection algorithm based on the updated heterogeneous weighted complex network includes:
[0027] Using the updated heterogeneous weighted complex network as input, we perform modularity maximization community detection at the finest level to obtain the primary community division;
[0028] The primary communities are aggregated into coarse-grained super nodes, while retaining the weights of the edges between communities, to build the upper-level network;
[0029] In the upper-level network, modularity maximization is performed again, and iterative cohesion is used to form a multi-level community structure.
[0030] For each level of community structure, the historical disaster loss variance is calculated synchronously, and the optimal level is selected by minimizing the loss variance as a constraint.
[0031] Map the communities corresponding to the optimal level back to the original raster cells to output the initial climate risk zoning.
[0032] Optionally, the process of inputting the initial climate risk zoning into a spatiotemporal convolutional graph neural network, performing transfer learning and fine-tuning using a zoning-loss coupling loss function, and outputting the final climate risk zoning result includes:
[0033] Using the initial climate risk zoning as graph node labels, the updated heterogeneous weighted complex network is transformed into graph structure data, and historical climate sequences of raster units are superimposed to construct a spatiotemporal feature matrix.
[0034] A pre-trained spatiotemporal convolutional graph neural network is used as the backbone to perform forward propagation on graph structure data and extract spatiotemporally coupled node representations.
[0035] The node representation is input into the partition-loss coupled loss function, where the partition term measures the consistency between the node representation and the center of its partition, the loss term measures the regression error of the node representation on historical disaster losses, and the regularization term constrains the model complexity.
[0036] The parameters of the spatiotemporal convolutional graph neural network are iteratively updated until convergence by simultaneously optimizing the weighted sum of the partitioning term, loss term, and regularization term through backpropagation.
[0037] After convergence, the optimized spatiotemporal convolutional graph neural network is used to infer all nodes and output the final climate risk level at the raster scale, forming the final climate risk zoning result.
[0038] This invention also proposes a climate risk zoning system based on complex networks for implementing the method, comprising:
[0039] The data acquisition module is used to acquire raw spatiotemporal data of climate-causing factors, exposure of disaster-bearing bodies, and vulnerability indicators within the area to be studied.
[0040] The data processing module is used to perform homogenization preprocessing on the original spatiotemporal data to generate a standardized raster dataset;
[0041] The model building module is used to construct a heterogeneous weighted complex network with grid cells as nodes, taking into account the spatial correlation of climate disaster-causing factors, the intensity of interaction between disaster-bearing bodies and the uncertainty of node attributes.
[0042] The model optimization module is used to update the heterogeneous weighted complex network by executing an adaptive threshold-freeze mechanism based on the temporal changes in the weights of the edges between nodes.
[0043] The risk partitioning module is used to run a cross-scale hierarchical community detection algorithm based on the updated heterogeneous weighted complex network to obtain the initial climate risk partitioning.
[0044] The result optimization module is used to input the initial climate risk zoning into a spatiotemporal convolutional graph neural network, use the zoning-loss coupling loss function for transfer learning and fine-tuning, and output the final climate risk zoning result.
[0045] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.
[0046] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.
[0047] The present invention also proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the method.
[0048] Compared with the prior art, the present invention has the following advantages and technical effects:
[0049] This application first aligns data on a unified spatiotemporal benchmark and grid granularity. Then, it constructs a heterogeneous weighted complex network based on three dimensions: climate correlation, disaster-bearing body interaction, and node uncertainty. This allows nodes and edges to dynamically reflect the coupling changes between disaster causation and vulnerability. Through an adaptive threshold-freeze mechanism, only edges with significant weight fluctuations are locally recalculated, achieving lightweight incremental updates to the network topology and avoiding computational redundancy caused by full reconstruction. Subsequently, a cross-scale hierarchical community discovery algorithm aggregates nodes from the bottom up under the dual constraints of modularity and disaster loss, forming a refined and spatially continuous initial climate risk zoning. This zoning label and historical disaster loss jointly drive the transfer learning of the spatiotemporal convolutional graph neural network, continuously fine-tuning the network parameters with a zoning-loss coupling loss function. This allows the model to accurately capture the spatiotemporal evolution of disaster risk while maintaining spatial consistency, ultimately obtaining the updated zoning results.
[0050] This invention solves the core problem that traditional climate risk zoning cannot simultaneously achieve high precision, high timeliness, and low cost. It significantly improves the spatiotemporal accuracy, boundary accuracy, and regional adaptability of climate risk zoning, and realizes high-resolution, interpretable, and robust integrated fine-grained climate risk zoning. Attached Figure Description
[0051] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0052] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;
[0053] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation
[0054] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0055] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0056] Example 1
[0057] like Figure 1 As shown, this embodiment provides a climate risk zoning method based on complex networks, including the following steps:
[0058] Obtain raw spatiotemporal data of climate-causing factors, exposure of disaster-bearing bodies, and vulnerability indicators within the area to be studied;
[0059] The original spatiotemporal data is subjected to homogenization preprocessing to generate a standardized raster dataset;
[0060] Using grid cells as nodes, a heterogeneous weighted complex network is constructed by comprehensively considering the spatial correlation of climate disaster-causing factors, the intensity of interaction between disaster-bearing bodies, and the uncertainty of node attributes.
[0061] Based on the temporal changes in the edge weights between nodes, an adaptive threshold-freeze mechanism is executed to update the heterogeneous weighted complex network.
[0062] Based on the updated heterogeneous weighted complex network, a cross-scale hierarchical community detection algorithm is run to obtain the initial climate risk zoning.
[0063] The initial climate risk zoning is input into a spatiotemporal convolutional graph neural network, and transfer learning and fine-tuning are performed using a zoning-loss coupling loss function to output the final climate risk zoning results.
[0064] The feasible process of performing homogenization preprocessing on the original spatiotemporal data to generate a standardized raster dataset includes:
[0065] The original spatiotemporal data is time-stamp aligned and spatial reference frame transformed; each layer is resampled using bilinear interpolation based on a 1km×1km square grid; missing values are marked on the resampled grid values and filled with spatiotemporal kriging; then the filled grid values are standardized dimensionlessly using the Z-score method, and finally a standardized grid dataset is output.
[0066] As an optional implementation, on a Linux server with 32 vCPUs and 128 GB RAM, four types of raw raster data—CMIP 6-day precipitation, ERA5 10 m wind speed, MODIS NDVI, and WorldPop population density—were first retrieved in parallel using xarray and rasterio. Then, the gdalwarp command of GDAL was used to uniformly transform all data to the WGS-84 / UTM Zone50N coordinate system, and the time axis was aligned to 00:00 UTC daily, completing the unification of the spatiotemporal reference. Next, using a 1km × 1km square grid as the reference, bilinear interpolation was used for one-time resampling, ensuring precise overlap of data with different spatial resolutions on the same grid. After resampling, the script automatically scanned for missing pixels and marked them as NaN; for missing regions, a spatiotemporal kriging module with a 5×5 pixel window was invoked to interpolate with a spatial radius of 20km and a ±7d time window, ensuring that the null pixels were consistent with the surrounding spatiotemporal background. Finally, Z-score standardization is performed on each layer using scikit-learn's StandardScaler, outputting a dimensionless standardized raster dataset with a mean of 0 and a variance of 1. This dataset is then stored in object storage in Cloud-Optimized GeoTIFF format for direct use in subsequent network modeling.
[0067] The feasible process of constructing a heterogeneous weighted complex network, using grid cells as nodes and comprehensively considering the spatial correlation of climate-causing factors, the intensity of interactions between disaster-bearing bodies, and the uncertainty of node attributes, includes:
[0068] Based on a standardized raster dataset, each raster cell is defined as a network node; the spatial correlation of climate-causing factors between nodes is calculated to obtain the first weight component; the interaction strength of disaster-bearing bodies between nodes is calculated to obtain the second weight component; the uncertainty of each node attribute is evaluated to obtain the third weight component; the first weight component, the second weight component, and the third weight component are weighted and summed according to preset weight coefficients to generate the edge weights between nodes; a heterogeneous weighted complex network is constructed based on the node and edge weights.
[0069] As an optional implementation, each grid cell is treated as a network node, generating a total of N nodes; then, three types of weight components are calculated in parallel:
[0070] 1) Spatial correlation of climate-causing factors:
[0071] For the two disaster-causing factors, precipitation and wind speed, calculate the Pearson correlation coefficient within a 5×5 sliding window for each node pair (i,j):
[0072] ;
[0073] in, , Let i and j be the hazard factors at time t. Its time series mean is used; the correlation coefficient is mapped to [0,1] as the first weighted component.
[0074] 2) Interaction intensity of disaster-bearing bodies:
[0075] Using population density (Pop) and total economic output (GDP) as the sources, a gravity model is introduced:
[0076] ;
[0077] Where Pop represents population density and GDP represents total economic output. Let ε be the Euclidean distance between nodes, and ε be a small constant to prevent zero; after normalizing the result, the second weight component is obtained.
[0078] 3) Uncertainty in node attributes:
[0079] First, calculate the standard deviation of the raster value for each node. , Then use the Gaussian kernel:
[0080] ;
[0081] in, , Let h be the standard deviation of the raster values for nodes i and j, and h be the bandwidth parameter. h is taken as 0.5 times the standard deviation of the entire image, thus converting the uncertainty difference into a similarity between 0 and 1, which is used as the third weight component.
[0082] Overall edge weights:
[0083] The three components are then weighted and summed using preset coefficients α=0.4, β=0.4, and γ=0.2:
[0084] ;
[0085] Only keep Edges with a value greater than 0.05 can be used to obtain a heterogeneous weighted complex network, and the adjacency matrix can be output for the next step of the "adaptive threshold-freezing" mechanism.
[0086] The feasible process of updating the heterogeneous weighted complex network by implementing an adaptive threshold-freeze mechanism based on the temporal changes in the edge weights between nodes includes:
[0087] Within a continuous time window, the relative rate of change of the current weight of each edge with the weight of the previous time window is calculated. The relative rate of change is compared with a preset adaptive threshold. If it is lower than the threshold, the edge is marked as static and its weight is frozen. If it is not lower than the threshold, it is retained as a dynamic edge. Only the weights of dynamic edges are recalculated and the network topology is updated, while the weights and topology of static edges remain unchanged, thus completing the update of the heterogeneous weighted complex network.
[0088] As an optional implementation, based on the constructed heterogeneous weighted complex network, the system automatically triggers an "adaptive threshold-freeze" update process once daily at midnight. First, the backend uses a 7-day sliding window to read the edge weights between the current and previous windows one by one, calculating the percentage difference between the two weights relative to the previous weight, as the relative rate of change for that edge. Then, this rate of change is compared with a preset threshold (the 5th percentile of the rate of change for all edges over the past 30 days): if it is lower than the threshold, the system immediately marks the edge as "static," freezing its weight and adjacency relationship, and it will not participate in any further recalculation; if it is equal to or higher than the threshold, it is marked as "dynamic," and only the three-factor weighting formula is reapplied for dynamic edges to obtain the latest weights and update the adjacency matrix in real time. The weights and topology of static edges remain unchanged, thus reducing redundant computation by approximately 80%. After the update is complete, the newly generated adjacency matrix is written to the graph database in real time in COO format, and a "network refreshed" signal is pushed to the front-end message queue, ensuring that subsequent cross-scale community detection algorithms always run on the lightest and most dynamic network structure.
[0089] A feasible process for obtaining initial climate risk zoning by running a cross-scale hierarchical community detection algorithm based on an updated heterogeneous weighted complex network includes:
[0090] Using the updated heterogeneous weighted complex network as input, modularity maximization is performed at the finest level to discover communities and obtain primary community divisions. These primary communities are then aggregated into coarse-grained supernodes, retaining the weights of the edges between communities, to construct an upper-level network. Modularity maximization is performed again in the upper-level network, iteratively agglomerating to form a multi-level community structure. Historical disaster loss variance is calculated synchronously for each level of community structure, and the optimal level is selected by minimizing the loss variance. The communities corresponding to the optimal level are mapped back to the original raster cells, outputting the initial climate risk zoning.
[0091] As an optional implementation, once the updated heterogeneous weighted complex network is written into the graph database, the system immediately sends it to the cross-scale hierarchical community discovery process. First, the Louvain algorithm is run at the finest-grained level, retaining all original grid nodes, to complete the first round of partitioning with the single objective of maximizing modularity, resulting in a large number of primary communities. Subsequently, each primary community is treated as a "super node," with its internal attributes calculated using a weighted average, and the edge weights between communities inherited according to the total weights between communities in the original network, thus forming a higher-level, sparser, coarse-grained network. Louvain is called again on this higher-level network, repeating the "agglomeration-repartitioning" steps until the modularity gain is below 0.001 or the number of communities drops to a preset lower limit, thereby generating a bottom-up multi-level tree structure.
[0092] After each level of partitioning is completed, the system synchronously calls historical disaster loss raster data to calculate the variance of disaster loss within each community. It then uses "variance minimization" as a constraint to filter levels: if the variance within a community at a certain level is significantly lower than that of its parent level and the difference between communities is significantly higher than that of its child levels, then that level is marked as optimal. Finally, the system losslessly maps the community numbers of the optimal level back to the original 1km×1km raster cells, outputting an initial climate risk zoning file that corresponds one-to-one with the geographic grid, for direct reading by the subsequent spatiotemporal convolutional graph neural network.
[0093] The feasible process of inputting the initial climate risk zoning into a spatiotemporal convolutional graph neural network, using a zoning-loss coupling loss function for transfer learning and fine-tuning, and outputting the final climate risk zoning result includes:
[0094] Using the initial climate risk zoning as graph node labels, the updated heterogeneous weighted complex network is transformed into graph-structured data, and historical climate sequences of raster units are superimposed to construct a spatiotemporal feature matrix. A pre-trained spatiotemporal convolutional graph neural network is used as the backbone to perform forward propagation on the graph-structured data to extract spatiotemporally coupled node representations. The node representations are input into a zoning-loss coupling loss function, where the zoning term measures the consistency between the node representation and its zoning center, the loss term measures the regression error of the node representation to historical disaster losses, and the regularization term constrains the model complexity. The weighted sum of the zoning term, loss term, and regularization term is simultaneously optimized through backpropagation, and the parameters of the spatiotemporal convolutional graph neural network are iteratively updated until convergence. After convergence, the optimized spatiotemporal convolutional graph neural network is used to infer all nodes and output the final climate risk level at the raster scale, forming the final climate risk zoning result.
[0095] As an optional implementation, once the cross-scale community discovery process outputs the initial climate risk zoning, the system immediately injects it into the fine-tuning pipeline of the Spatiotemporal Convolutional Graph Neural Network (ST-GCN). First, using the initial zoning results as graph node labels, the updated heterogeneous weighted complex network is directly converted into a graph data structure: nodes are raster cells, and edge weights are climate-socioeconomic coupling weights. Subsequently, historical climate sequences such as precipitation, wind speed, and temperature from the past thirty years are stacked according to time steps to form a spatiotemporal feature matrix corresponding to each node, achieving a complete expression of "one graph, one time axis".
[0096] Next, the ST-GCN model, pre-trained on tens of millions of meteorological and disaster samples, is used as the backbone network. This network consists of three layers of spatiotemporal graph convolutional blocks. Each layer first performs graph convolution on the graph space to capture neighborhood dependencies, and then performs gated convolution along the time dimension to extract temporal patterns. A single forward propagation is sufficient to obtain a high-dimensional spatiotemporal coupled node representation.
[0097] Subsequently, the node representations are fed into a partition-loss coupled loss function: the first term, "Partition Consistency," ensures the spatial continuity of the partitions by narrowing the Euclidean distance between the node representation and the center of its respective partition; the second term, "Loss Regression," uses a fully connected layer to map the representations to historical disaster loss values and measures regression accuracy using mean squared error; the third term, "Model Regularization," uses L2 weight decay to suppress overfitting. These three terms are weighted and summed in a 1:1:0.01 ratio, and then backpropagated through the AdamW optimizer, with the learning rate smoothly decreasing from 0.001 to 0.00001 using a cosine annealing strategy. Evaluation is performed on the validation set every 20 epochs; if the loss decreases by less than 0.1% for three consecutive validation rounds, training is stopped early.
[0098] After training, the optimal parameters are frozen, and inference is performed on all nodes within the study area in one go. The network outputs the final climate risk level (level 1–5) for each grid cell and automatically writes it into the cloud-optimized GeoTIFF. After overlaying the original coordinate system, the final climate risk zoning results can be directly published.
[0099] Furthermore, to upgrade static zoning into a "traceable and retrospective" dynamic product, after outputting the final climate risk zoning GeoTIFF for the current period, it is immediately written to the "zoning version library" with a timestamp, and stacked period by period to form a three-dimensional risk spatiotemporal cube R(x,y,t). Whenever a new version is added to the library, the "evolution feature extractor" is automatically triggered: first, a 5×5 level risk transition matrix T(t→t+Δt) is established for each raster to quantify the probability of level jumps; then, the risk contour expansion / contraction rate V(x,y,t) is estimated using a 3×3×3 pixel-time window, in km / year; finally, the Mann-Kendall test is performed on the time series of each raster to generate a "significant increase-stable-significant decrease" trend layer P(x,y). T, V, and P are written to COG along with the current zoning results and labeled "climate-risk-evolution" in the metadata. The front end can continuously play the evolution of risk boundaries through a time slider, or click on any grid point to pop up the risk level curve, transition probability and trend p value of its historical year to the latest year, realizing closed-loop management of "one-time zoning, long-term evolution and real-time visibility".
[0100] like Figure 2 As shown, this invention also proposes a climate risk zoning system based on complex networks for implementing the method, comprising:
[0101] The data acquisition module is used to acquire raw spatiotemporal data of climate-causing factors, exposure of disaster-bearing bodies, and vulnerability indicators within the area to be studied.
[0102] The data processing module is used to perform homogenization preprocessing on the original spatiotemporal data to generate a standardized raster dataset;
[0103] The model building module is used to construct a heterogeneous weighted complex network with grid cells as nodes, taking into account the spatial correlation of climate disaster-causing factors, the intensity of interaction between disaster-bearing bodies and the uncertainty of node attributes.
[0104] The model optimization module is used to update the heterogeneous weighted complex network by executing an adaptive threshold-freeze mechanism based on the temporal changes in the weights of the edges between nodes.
[0105] The risk partitioning module is used to run a cross-scale hierarchical community detection algorithm based on the updated heterogeneous weighted complex network to obtain the initial climate risk partitioning.
[0106] The result optimization module is used to input the initial climate risk zoning into a spatiotemporal convolutional graph neural network, use the zoning-loss coupling loss function for transfer learning and fine-tuning, and output the final climate risk zoning result.
[0107] Example 2
[0108] This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in Embodiment 1.
[0109] Example 3
[0110] This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.
[0111] Example 4
[0112] This embodiment also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 1.
[0113] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A climate risk zoning method based on complex networks, characterized in that, Includes the following steps: Obtain raw spatiotemporal data of climate-causing factors, exposure of disaster-bearing bodies, and vulnerability indicators within the area to be studied; The original spatiotemporal data is subjected to homogenization preprocessing to generate a standardized raster dataset; Using grid cells as nodes, a heterogeneous weighted complex network is constructed by comprehensively considering the spatial correlation of climate disaster-causing factors, the intensity of interaction between disaster-bearing bodies, and the uncertainty of node attributes. Based on the temporal changes in the edge weights between nodes, an adaptive threshold-freeze mechanism is executed to update the heterogeneous weighted complex network. Based on the updated heterogeneous weighted complex network, a cross-scale hierarchical community detection algorithm is run to obtain the initial climate risk zoning. The initial climate risk zoning is input into a spatiotemporal convolutional graph neural network, and transfer learning and fine-tuning are performed using the zoning-loss coupling loss function to output the final climate risk zoning result. The zoning-loss coupling loss function includes: the zoning item metric node represents the consistency with its respective zoning center, the loss item metric node represents the regression error of historical disaster losses, and the regularization term constrains the model complexity. The process of constructing a heterogeneous weighted complex network, using grid cells as nodes and comprehensively considering the spatial correlation of climate-causing factors, the intensity of interactions between disaster-bearing bodies, and the uncertainty of node attributes, includes: Based on a standardized raster dataset, each raster cell is defined as a network node; The spatial correlation of climate-causing factors among nodes is calculated to obtain the first weighted component. Calculate the interaction strength between disaster-bearing bodies between nodes to obtain the second weight component; The uncertainty of each node's attributes is evaluated to obtain the third weight component; The first weight component, the second weight component, and the third weight component are weighted and summed according to a preset weight coefficient to generate the edge weights between nodes. Construct a heterogeneous weighted complex network based on the node and edge weights.
2. The method according to claim 1, characterized in that, The process of performing homogenization preprocessing on the original spatiotemporal data to generate a standardized raster dataset includes: The original spatiotemporal data is time-stamp aligned and spatial reference frame transformed; each layer is resampled using bilinear interpolation based on a 1km×1km square grid; missing values are marked on the resampled grid values and filled with spatiotemporal kriging; then the filled grid values are standardized dimensionlessly using the Z-score method, and finally a standardized grid dataset is output.
3. The method according to claim 1, characterized in that, The process of updating the heterogeneous weighted complex network based on the temporal changes in the edge weights between nodes, and by implementing an adaptive threshold-freeze mechanism, includes: Within a continuous time window, calculate the relative rate of change of the current weight of each edge compared to the weight of the previous time window; The relative rate of change is compared with a preset adaptive threshold. If it is lower than the threshold, the edge is marked as static and its weight is frozen. If it is not lower than the threshold, it is retained as a dynamic edge. Only the weights of dynamic edges are recalculated and the network topology is updated, while the weights and topology of static edges remain unchanged, thus completing the update of heterogeneous weighted complex networks.
4. The method according to claim 1, characterized in that, The process of obtaining initial climate risk zoning based on an updated heterogeneous weighted complex network and running a cross-scale hierarchical community detection algorithm includes: Using the updated heterogeneous weighted complex network as input, we perform modularity maximization community detection at the finest level to obtain the primary community division; The primary communities are aggregated into coarse-grained super nodes, while retaining the weights of the edges between communities, to build the upper-level network; In the upper-level network, modularity maximization is performed again, and iterative cohesion is used to form a multi-level community structure. For each level of community structure, the historical disaster loss variance is calculated synchronously, and the optimal level is selected by minimizing the loss variance as a constraint. Map the communities corresponding to the optimal level back to the original raster cells to output the initial climate risk zoning.
5. The method according to claim 1, characterized in that, The process of inputting the initial climate risk zoning into a spatiotemporal convolutional graph neural network, performing transfer learning and fine-tuning using a zoning-loss coupling loss function, and outputting the final climate risk zoning result includes: Using the initial climate risk zoning as graph node labels, the updated heterogeneous weighted complex network is transformed into graph structure data, and historical climate sequences of raster units are superimposed to construct a spatiotemporal feature matrix. A pre-trained spatiotemporal convolutional graph neural network is used as the backbone to perform forward propagation on graph structure data and extract spatiotemporally coupled node representations. The node is represented by the input partition-loss coupled loss function; The parameters of the spatiotemporal convolutional graph neural network are iteratively updated until convergence by simultaneously optimizing the weighted sum of the partitioning term, loss term, and regularization term through backpropagation. After convergence, the optimized spatiotemporal convolutional graph neural network is used to infer all nodes and output the final climate risk level at the raster scale, forming the final climate risk zoning result.
6. A climate risk zoning system based on complex networks, characterized in that, For implementing the method according to any one of claims 1-5, comprising: The data acquisition module is used to acquire raw spatiotemporal data of climate-causing factors, exposure of disaster-bearing bodies, and vulnerability indicators within the area to be studied. The data processing module is used to perform homogenization preprocessing on the original spatiotemporal data to generate a standardized raster dataset; The model building module is used to construct a heterogeneous weighted complex network with grid cells as nodes, taking into account the spatial correlation of climate disaster-causing factors, the intensity of interaction between disaster-bearing bodies and the uncertainty of node attributes. The model optimization module is used to update the heterogeneous weighted complex network by executing an adaptive threshold-freeze mechanism based on the temporal changes in the weights of the edges between nodes. The risk partitioning module is used to run a cross-scale hierarchical community detection algorithm based on the updated heterogeneous weighted complex network to obtain the initial climate risk partitioning. The result optimization module is used to input the initial climate risk zoning into a spatiotemporal convolutional graph neural network, use the zoning-loss coupling loss function for transfer learning and fine-tuning, and output the final climate risk zoning result.
7. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-5.
9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-5.