Mountain torrent disaster danger area intelligent division method and system based on deep learning

Abstract

The invention relates to the technical field of deep learning and image recognition, in particular to a mountain torrent disaster danger area intelligent division method and system based on deep learning, and the method comprises the steps: carrying out the calculation from multi-source space image data, and generating a multi-modal landform feature map layer; based on the spatial adjacency relation and the elevation flow direction relation of the structure nodes, constructing a scene structure chart containing heterogeneous edge types; constructing a graph feature fusion network, generating a context feature vector for each structure node through an attention mechanism, and outputting the classification confidence of each structure node; generating a feature saliency map and a model confidence map by adopting a gradient attribution analysis method; and after processing, fusing the classification confidence, the feature saliency map and the model confidence map, generating a semantic segmentation map of the target area, and carrying out dangerous area division. According to the method, the scene structure diagram is constructed, and the diagram feature fusion network, the attention mechanism and the gradient attribution analysis method are adopted, so that intelligent division of the mountain torrent disaster danger area is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and image recognition, in particular to a mountain flood disaster dangerous area intelligent division method and system based on deep learning. BACKGROUND

[0002] Mountain flood disaster is sudden and destructive, so it is crucial to accurately and efficiently identify and divide the dangerous area.

[0003] Currently, the methods for delineating mountain flood disaster dangerous areas mainly include two types: physical model-based and data-driven. However, in the data-driven method, the existing technology fails to fully utilize the ability of deep learning models to directly perform end-to-end feature learning and pattern recognition from raw image data.

[0004] One of the mainstream methods follows a two-stage process of "manual feature extraction - traditional model classification". This process relies on pre-defined rules to extract discrete geographic indicators, which hinders deep neural networks from directly acting on raw remote sensing images. Since it fails to utilize deep learning models to automatically learn and discover hierarchical features from low-level textures to high-level semantics in images, this method not only loses key visual information such as topography and spatial context, but also cannot perform pixel-level dense prediction tasks, making it difficult to generate high-resolution dangerous area semantic segmentation maps.

[0005] Another type of technology has applied deep learning, but its network architecture and training objectives are designed for different visual tasks. For example, they mainly solve local target detection or classification problems, such as identifying specific objects or defects in images. However, the identification of mountain flood disaster areas is essentially a complex large-scale geographic scene semantic segmentation task that requires the model to understand the spatial combination and correlation patterns of topography, water systems, vegetation, and other geographic elements. Therefore, the structure and learning ability of deep learning models designed for simple object recognition tasks are not suitable for the analysis of this macro scene.

[0006] In summary, there is a clear gap in the existing technology: the lack of a deep learning solution specifically designed for mountainous terrain scene understanding. Traditional methods cannot apply the advantages of deep learning due to their non-end-to-end processing flow; while existing deep learning applications cannot be directly transferred due to task mismatch. Therefore, there is a need for a new technical solution that can build a deep learning model that can directly process multi-source remote sensing images to complete the complete task from feature representation to dangerous area semantic segmentation in an end-to-end manner.

[0007] To this end, a mountain flood disaster dangerous area intelligent division method and system based on deep learning is proposed. SUMMARY

[0008] The application aims to provide a deep learning-based intelligent mountain torrent disaster hazard area division method and system.

[0009] To achieve the above-mentioned purpose, the application provides the following technical scheme.

[0010] A deep learning-based intelligent mountain torrent disaster hazard area division method comprises the following steps:

[0011] Receiving multi-source spatial image data; using a geomorphological feature extractor, calculating and generating multi-modal geomorphological feature layers from the multi-source spatial image data;

[0012] Spatial gridding is performed on the multi-modal geomorphological feature layers, and the region to be analyzed is divided into elementary grid cells; the elementary grid cells are defined as structure nodes; based on the spatial adjacency relationship and elevation flow direction relationship of the structure nodes, a scene structure graph containing heterogeneous edge types is constructed;

[0013] A graph feature fusion network is constructed, and the scene structure graph is taken as input; the graph feature fusion network aggregates neighborhood information through an attention mechanism, and generates context feature vectors for each structure node; the feature vectors are input into a pattern classifier, and the classification confidence of each structure node is output;

[0014] A gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multi-modal geomorphological feature layers, and feature saliency maps and model confidence maps are generated based on the gradient; the classification confidence is subjected to spatial interpolation and smoothing processing, and the feature saliency maps and model confidence maps are fused to generate a semantic segmentation map of the target region and perform hazard area division.

[0015] Preferably, the multi-source spatial image data specifically comprises digital elevation model data, geological maps, soil maps, and remote sensing image data; the multi-modal geomorphological feature layers specifically comprise geometric feature layers, physical feature layers, and coverage feature layers; the geometric feature layers contain slope, slope direction, and planar curvature calculated by processing the digital elevation model data; the physical feature layers contain surface roughness and soil infiltration rate obtained by querying the geological map and soil map data; and the coverage feature layers are normalized vegetation indices calculated by processing the remote sensing image data.

[0016] Preferably, the process of constructing the scene structure graph with heterogeneous edge types is specifically: defining the primitive grid as the structure node of the scene structure graph; establishing a set of heterogeneous edge types, wherein: between any two structure nodes sharing a spatial boundary, a spatial adjacency edge is established; according to the calculation from the digital elevation model data, a structure node from which a surface runoff flows out to a structure node into which the surface runoff flows, an elevation flow edge is established; the spatial adjacency edge and the elevation flow edge together constitute the set of heterogeneous edge types of the scene structure graph; obtaining the grid values of all multi-modal geomorphic feature layers in the primitive grid corresponding to the structure node, calculating and splicing the mean and standard deviation of the multi-modal geomorphic feature layer grid values to form the initial attribute vector of the structure node.

[0017] Preferably, the specific implementation of the attention mechanism is: for any two structure nodes connected by an edge, one of the structure nodes is regarded as a center node and the other structure node is regarded as a neighborhood node when calculating the attention weight between them; according to the type of the edge connecting the center node and the neighborhood node, a learnable transformation matrix exclusive to the edge type is used to transform the feature vector of the neighborhood node; the original feature vector of the center node and the transformed feature vector of the neighborhood node are input into an attention function to calculate the attention weight.

[0018] Preferably, the process of generating a context feature vector for each structure node includes: for any structure node to be processed, defining it as a center node and all structure nodes connected to the center node by edges as neighborhood nodes; for each neighborhood node, according to the type of the edge connecting the neighborhood node and the center node, selecting a corresponding edge type transformation matrix from a plurality of preset transformation matrices and performing linear transformation on the feature vector of the neighborhood node; inputting the feature vector of the center node and the feature vectors of all transformed neighborhood nodes into a shared attention function to calculate the normalized attention weight for each neighborhood node; using the normalized attention weight to perform weighted summation on the feature vectors of the transformed neighborhood nodes to generate the context feature vector of the center node.

[0019] Preferably, the gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multi-modal terrain feature layer, and a feature saliency map and a model confidence map are generated based on the gradient, and the process is specifically as follows: for each base element grid, when the classification confidence for a predefined mode category meets a preset judgment condition, the partial derivatives of the classification confidence with respect to each feature dimension in the initial attribute vector of the structural node corresponding to the base element grid are calculated through the back propagation mechanism of the graph feature fusion network and the mode classifier; the absolute values of the partial derivatives are normalized and visualized as a heat map to generate the feature saliency map; the Monte Carlo dropout technique is introduced into the graph feature fusion network and the mode classifier, and the specific process is as follows: in the model inference stage, the forward propagation process with random inactivation is performed multiple times to obtain the classification confidence output for each base element grid; the variance of the classification confidence output is calculated, and the variance is visualized to generate the model confidence map.

[0020] Preferably, the process of spatial interpolation and smoothing processing of the classification confidence is specifically as follows: the classification confidence of each base element grid is assigned to the geometric center of the base element grid to form a confidence point set; inverse distance weighted interpolation is performed on the confidence point set to generate a continuous classification confidence surface; Gaussian filtering is applied to the classification confidence surface for smoothing processing to obtain a smoothed classification confidence surface.

[0021] A deep learning-based intelligent mountain torrent disaster hazard zone division system, comprising: an image feature extraction module: receiving multi-source spatial image data; using a terrain feature extractor to calculate and generate multi-modal terrain feature layers from the multi-source spatial image data; a spatial graph construction module: spatially gridding the multi-modal terrain feature layers to divide the region to be analyzed into base element grids; defining the base element grids as structural nodes; constructing a scene structure graph containing heterogeneous edge types based on the spatial adjacency relationship and the elevation flow direction relationship of the structural nodes; a graph network inference and recognition module: constructing a graph feature fusion network with the scene structure graph as input; the graph feature fusion network aggregates neighborhood information through an attention mechanism to generate context feature vectors for each structural node; inputting the feature vectors into a mode classifier to output the classification confidence of each structural node; a fusion and segmentation graph module: using a gradient attribution analysis method to calculate the gradient of the classification confidence with respect to the input multi-modal terrain feature layer, and generating a feature saliency map and a model confidence map based on the gradient; performing spatial interpolation and smoothing processing on the classification confidence, and fusing the feature saliency map and the model confidence map to generate a semantic segmentation map of the target region and perform hazard zone division.

[0022] Compared with the prior art, the beneficial effects of the present application are:

[0023] 1. The present application solves the major technical problem that traditional grid methods cannot effectively model non-Euclidean relationships in geographic space by abstracting geographic space into a scenario structure graph containing two heterogeneous edge types of "spatial adjacency" and "elevation flow direction". Traditional methods can only handle local and regular neighborhood relationships, but cannot capture long-distance physical associations along irregular paths such as surface runoff. The heterogeneous graph construction method of the present application can accurately express the proximity of geographic topology and the causality of hydrological processes at the same time, making the data model more consistent with the real geographic environment and providing a scientific and stable data structure foundation for the accuracy of subsequent analysis.

[0024] 2. The present application uses a graph feature fusion network and introduces an attention mechanism to realize dynamic and intelligent reasoning of complex spatial relationships, overcoming the limitations of traditional models in feature fusion. Traditional models often assign the same or fixed weight to all spatial influencing factors, and cannot identify the main disaster-causing factors in a specific geographic context. The attention mechanism of the present application can automatically learn and assign different weights according to the node's own characteristics and spatial connection mode, so that the generated context feature vector can better reflect the comprehensive influence of key geographic elements. This greatly improves the accuracy and robustness of pattern recognition, making the division results of the danger zone more consistent with the real physical disaster-causing law.

[0025] 3. The present application solves the black box problem that exists in existing deep learning models in disaster assessment applications by introducing gradient attribution analysis and fusing generated feature saliency maps and model confidence maps. Traditional methods can only output a prediction result, but cannot explain the cause and reliability, making it difficult for decision-makers to fully trust and apply. The present application not only gives the semantic segmentation map of where the danger zone is, but also clearly reveals the dominant cause of danger through the feature saliency map, and quantifies the reliability of the prediction result through the "model confidence map". This comprehensive output of "prediction-explanation-reliability evaluation" provides more comprehensive and reliable decision support for disaster prevention. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 A flowchart of a deep learning-based mountain flood disaster danger zone intelligent division method of the present application;

[0027] Figure 2 A timing diagram of a deep learning-based mountain flood disaster danger zone intelligent division method and system of the present application;

[0028] Figure 3 A structural diagram of a deep learning-based mountain flood disaster danger zone intelligent division system of the present application. DETAILED DESCRIPTION

[0029] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0030] Please refer to Figures 1 to 3 The present application provides a deep learning-based intelligent mountain torrent disaster hazard zone division method and system, and the technical solutions are as follows:

[0031] A deep learning-based intelligent mountain torrent disaster hazard zone division method comprises the following steps:

[0032] Receiving multi-source spatial image data; using a topographic feature extractor, calculating and generating multi-modal topographic feature layers from the multi-source spatial image data;

[0033] Spatial gridding is performed on the multi-modal topographic feature layers, and the region to be analyzed is divided into elementary grid cells; the elementary grid cells are defined as structure nodes; based on the spatial adjacency relationship and the elevation flow direction relationship of the structure nodes, a scene structure graph containing heterogeneous edge types is constructed;

[0034] A graph feature fusion network is constructed, and the scene structure graph is taken as input; the graph feature fusion network aggregates neighborhood information through an attention mechanism to generate context feature vectors for each structure node; the feature vectors are input into a pattern classifier to output the classification confidence of each structure node;

[0035] Using a gradient attribution analysis method, the gradient of the classification confidence with respect to the input multi-modal topographic feature layers is calculated, and a feature saliency map and a model confidence map are generated based on the gradient; the classification confidence is spatially interpolated and smoothed, and the feature saliency map and the model confidence map are fused to generate a semantic segmentation map of the target region and perform hazard zone division.

[0036] Embodiment one

[0037] This embodiment provides a specific application of a deep learning-based intelligent mountain torrent disaster hazard zone division method, taking the division of a mountain torrent disaster hazard zone in a certain mountainous river basin as an example to illustrate the specific implementation of the present application. Please refer to Figure 1 、 Figure 2 This embodiment processes according to the flow of a deep learning-based intelligent mountain torrent disaster hazard zone division method provided by the present application, and shows the data flow, and the specific implementation process is as follows:

[0038] Further, multi-source spatial image data is received; a multi-modal geomorphic feature layer is calculated and generated from the multi-source spatial image data by using a geomorphic feature extractor, and the specific process includes:

[0039] A certain mountainous basin is selected as the region to be analyzed, and the area of the region is about 200 square kilometers. First, multi-source spatial image data is received, which specifically includes 30-meter resolution digital elevation model data, 10-meter resolution remote sensing image data, and 1:50000 scale geological map and soil map. Then, the geomorphic feature extractor is used to calculate the above data to generate a multi-modal geomorphic feature layer. Among them, the slope, slope direction and plane curvature are calculated by processing the digital elevation model data to form a geometric feature layer; the surface roughness and soil infiltration rate are obtained by querying the geological map and soil map data to form a physical feature layer; the normalized vegetation index is calculated by processing the remote sensing image data to form a coverage feature layer.

[0040] By explicitly defining the multi-modal geomorphic feature layer as three types of geometry, physics and coverage, and specifying key features such as slope, surface roughness and normalized vegetation index, the model is provided with comprehensive and closely related input information to the mechanism of mountain flood disasters. This ensures that the basic data for model analysis is sufficient and has clear physical meaning, improving the accuracy and physical authenticity of the final division results.

[0041] Further, the multi-modal geomorphic feature layer is spatially gridded, and the region to be analyzed is divided into elementary grids; the elementary grids are defined as structure nodes; based on the spatial adjacency relationship and elevation flow direction relationship of the structure nodes, a scene structure graph containing heterogeneous edge types is constructed, and the specific process includes:

[0042] All multi-modal geomorphic feature layers are spatially gridded at 30m x 30m, and the region to be analyzed is divided into about 220000 elementary grids. Each elementary grid is defined as a structure node. Based on these structure nodes, a scene structure graph containing heterogeneous edge types is constructed: between any two structure nodes corresponding to elementary grids that share a spatial boundary, a spatial adjacency edge is established; according to the digital elevation model data, a structure node corresponding to an elementary grid from which surface runoff flows is established to a structure node corresponding to an elementary grid into which it flows, and a elevation flow direction edge is established. These two types of edges together form the heterogeneous edge set of the scene structure graph.

[0043] The process of establishing a elevation flow direction edge is as follows: for any elementary grid, among its eight adjacent elementary grids, determine the unique target elementary grid with the maximum elevation difference; establish a directed edge from the structure node corresponding to the elementary grid to the structure node corresponding to the target elementary grid as the elevation flow direction edge.

[0044] By adopting the manner of determining the target element grid with the largest elevation difference in the eight adjacent element grids, it is ensured that the surface runoff direction of each non-converging point element grid is unique, the path ambiguity that may occur when constructing the graph topology is avoided, and the finally generated scene structure graph has reproducibility and structural uniqueness.

[0045] For each structure node in the graph, the mean and standard deviation of all multi-modal terrain feature layer grid values in the corresponding element grid are extracted, and these values are spliced to form an initial attribute vector of 12 dimensions. Through the above steps, the scene structure graph of the region to be analyzed is obtained.

[0046] By respectively establishing spatial adjacency edges and elevation flow direction edges, specific technical implementation is provided for the construction of the scene structure graph. This way enables the graph topology to represent both the proximity relationship in geographic space and the surface runoff directionality determined by elevation, which are two different physical associations. By calculating and splicing the mean and standard deviation of multi-modal terrain feature layer grid values to form the initial attribute vector of the node, the initial state of the node not only reflects the central tendency of the regional features, but also reflects the internal variation degree of the features, thereby providing more representative input information for the subsequent graph feature fusion network.

[0047] Further, a graph feature fusion network is constructed, taking the scene structure graph as input; the graph feature fusion network aggregates neighborhood information through an attention mechanism to generate a context feature vector for each structure node; the feature vector is input into a mode classifier to output the classification confidence of each structure node, and the specific process includes:

[0048] A graph feature fusion network is constructed, and the scene structure graph generated in the previous step is taken as input.

[0049] The graph feature fusion network includes a plurality of stacked graph feature fusion layers; each graph feature fusion layer receives the output feature vector of the previous layer as its input feature vector; the graph feature fusion layer aggregates neighborhood information through the attention mechanism to generate an intermediate feature vector, and element-wise adds the intermediate feature vector and the input feature vector to obtain the output feature vector of the layer; the output feature vector of the last layer is the context feature vector.

[0050] In this embodiment, the graph feature fusion network is preferably a deep graph network comprising three stacked graph feature fusion layers. The input layer of the network accepts the 12-dimensional initial attribute vector of each structure node in the scene structure graph. The first graph feature fusion layer receives the 12-dimensional initial attribute vector as input. This layer aggregates neighborhood information through the attention mechanism and the edge type transformation matrix, and outputs a 32-dimensional intermediate feature vector. The second graph feature fusion layer receives the 32-dimensional intermediate feature vector output by the previous layer as input, aggregates neighborhood information using the same attention mechanism, and outputs a 64-dimensional intermediate feature vector. The third graph feature fusion layer receives the 64-dimensional intermediate feature vector output by the previous layer as input, again performs neighborhood information aggregation, and finally outputs a 64-dimensional context feature vector. This vector is the final output of the entire graph feature fusion network. Within each graph feature fusion layer, after linear transformation of the feature vector, a rectified linear unit is used as the activation function to enhance the learning effect of the model. In each layer, the intermediate feature vector generated by aggregation is added element-wise to the input feature vector of the layer to form a residual connection, ensuring that the gradient is effectively propagated in the deep network.

[0051] By introducing the operation of adding the intermediate feature vector and the input feature vector element-wise between the stacked graph feature fusion layers, a residual connection is constructed, which helps to alleviate the gradient vanishing problem that may occur during the training of the deep graph network, and enables the network to pass the original feature information from the shallow nodes to the deep layer, thereby supporting the construction of a deeper graph feature fusion network.

[0052] When generating the context feature vector for any center node, for each of its neighborhood nodes, a corresponding edge type transformation matrix is selected from a plurality of preset transformation matrices according to the type of the connecting edge, and the feature vector of the neighborhood node is linearly transformed and uniformly transformed into 64 dimensions. Then, the feature vector of the center node and the feature vectors of all the transformed neighborhood nodes are input into the shared attention function to calculate the normalized attention weight for each of the neighborhood nodes.

[0053] By setting a dedicated learnable transformation matrix for different types of edges, the attention mechanism can differentially process information from different relationship neighborhoods. The model can therefore learn different influence patterns on the center node, thereby more finely modeling the nonlinear spatial coupling relationships between different landform features and improving the effectiveness and accuracy of feature fusion.

[0054] Finally, the normalized attention weights are used to weight and sum the feature vectors of the transformed neighbor nodes to generate a 64-dimensional context feature vector for the center node. The context feature vector is input to a pattern classifier, which is a multilayer perceptron. The number of output layer nodes of the multilayer perceptron corresponds to the number of predefined pattern categories. The pattern classifier uses a Softmax activation function to process the outputs of the output layer nodes, thereby generating a classification confidence for each structure node for each of the predefined pattern categories.

[0055] By calculating and weighting and summing the normalized attention weights of all nodes in the entire neighborhood, the application ensures that the final feature representation of each structure node dynamically and adaptively fuses the information of all important neighbors, thereby enhancing the comprehensive perception and discrimination ability of the model for local complex environments.

[0056] Further, a gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multi-modal topographic feature map layer, and a feature saliency map and a model confidence map are generated based on the gradient. The classification confidence is spatially interpolated and smoothed, and the feature saliency map and the model confidence map are fused to generate a semantic segmentation map of the target region and to divide the dangerous area. The specific process includes:

[0057] After obtaining the classification confidence of each structure node, a gradient attribution analysis method is used. For each primitive grid, when its classification confidence for a predefined pattern category (in this example, the "high risk" pattern) satisfies a preset judgment condition (for example, the confidence is greater than 0.8), the partial derivatives of the classification confidence with respect to each feature dimension in the initial attribute vector of the structure node corresponding to the primitive grid are calculated through the backpropagation mechanism of the graph feature fusion network and the pattern classifier. The absolute values of the partial derivatives are normalized and visualized as a heat map to generate the feature saliency map.

[0058] The application generates a feature saliency map by calculating partial derivatives through backpropagation, realizes visualization of model decision-making, and can clearly reveal the key topographic features that cause a certain region to be judged as a dangerous area, thereby providing an intuitive scientific basis for disaster prevention. By introducing Monte Carlo dropout technology to quantify the uncertainty of model prediction, the application can avoid overconfident errors in division and improve the safety and reliability of the entire division method.

[0059] Meanwhile, the Monte Carlo dropout technique is introduced in the graph feature fusion network and the pattern classifier, and the forward propagation process with a dropout rate of 0.2 is performed for 50 times; the model confidence map is generated by calculating the variance of the output results of 50 times. Finally, the classification confidence of each base element grid is assigned to the center point of the grid, the classification confidence surface is generated by using the inverse distance weighted interpolation method, and the classification confidence surface is smoothed by applying a Gaussian filter with a window size of 3*3 to obtain the smoothed classification confidence surface. The smoothed classification confidence surface is used as the basic risk level layer of the semantic segmentation map; the model confidence map is used as the confidence overlay layer of the semantic segmentation map; and the feature saliency map is used as the attribution explanation layer of the semantic segmentation map; and the basic risk level layer, the confidence overlay layer and the attribution explanation layer jointly constitute the semantic segmentation map of the target area.

[0060] By specifically embodying the final semantic segmentation map into a multi-layer data structure including a basic risk level, a confidence overlay and an attribution explanation, the predicted risk level, the predicted reliability evaluation and the key feature basis for making the prediction are simultaneously presented. The multi-dimensional output structure provides the user with more comprehensive decision reference information than a single risk layer, and improves the comprehensive information carrying capacity and practical application value of the division result.

[0061] By using the inverse distance weighted interpolation and the Gaussian filter smoothing processing, the grid effect and noise that may be generated in the calculation process are eliminated, so that the boundary of the final generated flash flood disaster risk zoning map is more natural, and the result is more consistent with the actual geographical distribution, and the practicability and visualization effect of the result are improved.

[0062] The application proposes an end-to-end intelligent division framework. By abstracting discrete geographic spatial data into a scene structure graph containing heterogeneous edge types, and analyzing by using a graph feature fusion network, the complex spatial adjacency relationship between landforms and the elevation flow relationship in hydrology can be effectively captured, and the problem that the traditional method is difficult to model the complex spatial dependence is overcome; meanwhile, the gradient attribution analysis method is introduced, not only the division result is generated, but also the feature saliency map and the model confidence map of the model decision basis are provided, and the explainability and reliability of the division result are improved.

[0063] Embodiment two

[0064] This embodiment aims to explain the specific workflow of the image feature extraction module and the spatial graph construction module in the application. In the scenario of this embodiment, a flash flood disaster risk area in a high mountain and valley region is identified. It is expected that through the method of the application, fine data modeling and spatial relationship construction are performed on the region to provide high-quality input for subsequent risk identification.

[0065] ReferenceFigure 3 The system structure of the present application is shown in the figure.

[0066] Firstly, the image feature extraction process is executed. Multi-source spatial image data is received, specifically including 5-meter resolution digital elevation model data, 1:25000 scale geological map and 2-meter resolution remote sensing image data. The above data is processed by using a topographic feature extractor to generate multi-modal topographic feature layers: by processing the digital elevation model data, the slope, slope direction and planar curvature are calculated to form a geometric feature layer; by querying the geological map, the surface roughness corresponding to the surface lithology is obtained to form a physical feature layer; by processing the remote sensing image data, the normalized vegetation index is calculated to form a cover feature layer.

[0067] Next, the spatial graph construction process is executed. All multi-modal topographic feature layers are spatially gridded at 10m x 10m, and the approximately 50 square kilometers of the area to be analyzed is divided into nearly 500,000 elementary grid cells. The process of constructing a scene structure graph containing heterogeneous edge types is as follows: the elementary grid cells are defined as structure nodes of the scene structure graph; a set of heterogeneous edge types is established, in which: between any two structure nodes that share a spatial boundary, a spatial adjacency edge is established; according to the digital elevation model data, from a structure node where a surface runoff flows out, to a structure node where it flows in, an elevation flow edge is established. The spatial adjacency edge and the elevation flow edge together constitute the set of heterogeneous edge types of the scene structure graph. Finally, the grid values of all multi-modal topographic feature layers in the elementary grid cells corresponding to the structure nodes are obtained, the mean and standard deviation of the multi-modal topographic feature layer grid values are calculated and spliced to form a 10-dimensional initial attribute vector of the structure node. Through the above steps, the scene structure graph of the area to be analyzed is obtained.

[0068] Example Three

[0069] This example aims to detail the specific workflow of the graph network reasoning identification module and the fusion segmentation graph module in the present application. In the scenario of this example, the identification of mountain flood disaster hazard zones is carried out in a certain plain river network area. It is expected that through the method of the present application, intelligent reasoning and analysis are carried out on the scene structure graph that has been constructed, and hazard zone division results with both explainability and reliability are generated.

[0070] Reference Figure 3 The system structure of the present application is shown in the figure.

[0071] The embodiment starts from a completed scene structure graph, and directly enters the graph network inference identification process. The structure node of the scene structure graph represents a catchment unit, and the initial attribute vector of the structure node contains statistical values of slope, slope direction, plane curvature, soil infiltration rate and other characteristics. The scene structure graph is input into a trained graph feature fusion network. The graph feature fusion network aggregates neighborhood information through an attention mechanism. When generating a context feature vector for any center node, for each neighborhood node of the center node, a learnable transformation matrix specific to the type of the connection edge is used for processing according to the type of the connection edge. The original feature vector of the center node and the feature vector of the neighborhood node after transformation are input into an attention function to calculate an attention weight, and finally a 128-dimensional context feature vector is generated. The context feature vector is then input into a pattern classifier to output the classification confidence of each structure node.

[0072] Finally, the fusion segmentation graph process is performed. Using gradient attribution analysis, for each primitive grid identified as a specific target pattern, the partial derivative of the classification confidence with respect to the input multi-modal topographic feature layer is calculated through a back propagation mechanism to generate a feature saliency map. Meanwhile, in the model inference stage, the Monte Carlo dropout technique is introduced into the graph feature fusion network and the pattern classifier, including: performing a forward propagation process with a dropout rate of 0.2 for 50 times; generating a model confidence map by calculating the variance of the 50 output results. After obtaining the classification confidence of each node, the pattern classification confidence of each primitive grid is assigned to the center point of the grid, and a reverse distance weighted interpolation method is used to generate a classification confidence surface. The confidence surface is smoothed by applying a Gaussian filter with a window size of 5x5. Finally, the smoothed result is visually fused with the feature saliency map and the model confidence map to generate a semantic segmentation map of the target area in the region.

[0073] The present application provides a set of function modularization and process automation intelligent division system. The system integrates all functions from data reception, feature extraction, graph construction, network inference to result generation, and realizes automatic and efficient processing of mountain flood disaster risk zoning tasks.

[0074] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent division method for mountain flood disaster hazard areas, characterized in that, The application relates to a method for semantic segmentation of a target region, comprising the following steps: receiving multi-source spatial image data; using a landform feature extractor to calculate and generate multi-modal landform feature layers from the multi-source spatial image data; spatial griding the multi-modal landform feature layers to divide the region to be analyzed into elementary grids; defining the elementary grids as structure nodes; constructing a scene structure graph containing heterogeneous edge types based on the spatial adjacency relationship and the elevation flow direction relationship of the structure nodes; constructing a graph feature fusion network with the scene structure graph as input; the graph feature fusion network aggregates neighborhood information through an attention mechanism to generate context feature vectors for each structure node; inputting the feature vectors into a pattern classifier to output the classification confidence of each structure node; using a gradient attribution analysis method to calculate the gradient of the classification confidence with respect to the input multi-modal landform feature layers, and generating a feature saliency map and a model confidence map based on the gradient; performing spatial interpolation and smoothing processing on the classification confidence, and fusing the feature saliency map and the model confidence map to generate a semantic segmentation map of the target region and perform hazard zoning. 2.The deep learning-based intelligent mountain flood disaster hazard zone partitioning method of claim 1, wherein, The multi-source spatial image data specifically comprises digital elevation model data, a geological map, a soil map and remote sensing image data; the multi-modal landform feature layers specifically comprise a geometric feature layer, a physical feature layer and a coverage feature layer; the geometric feature layer contains slope, slope direction and plane curvature calculated by processing the digital elevation model data; the physical feature layer contains surface roughness and soil infiltration rate obtained by querying the geological map and soil map data; and the coverage feature layer is a normalized vegetation index calculated by processing the remote sensing image data. 3.The deep learning-based intelligent mountain flood disaster hazard zone partitioning method of claim 2, wherein, The process of constructing the scene structure graph containing heterogeneous edge types specifically comprises the following steps: defining the elementary grids as structure nodes of the scene structure graph; establishing a set of heterogeneous edge types, wherein: a spatial adjacency edge is established between any two structure nodes sharing a spatial boundary; an elevation flow edge is established from a structure node from which a surface runoff flows to a structure node into which the surface runoff flows according to the digital elevation model data; the spatial adjacency edge and the elevation flow edge jointly constitute the set of heterogeneous edge types of the scene structure graph; obtaining the grid values of all multi-modal landform feature layers in the elementary grid corresponding to the structure node, calculating and splicing the mean and standard deviation of the multi-modal landform feature layer grid values to form the initial attribute vector of the structure node.

4. The method according to claim 1, wherein, The specific implementation mode of the attention mechanism is as follows: for any two structure nodes connected by an edge, one of the structure nodes is regarded as a center node and the other structure node is regarded as a neighborhood node when calculating the attention weight between the two structure nodes; according to the type of the edge connecting the center node and the neighborhood node, a learnable transformation matrix exclusive to the edge type is used to transform the feature vector of the neighborhood node; the original feature vector of the center node and the transformed feature vector of the neighborhood node are input into an attention function to calculate the attention weight. 5.The deep learning-based intelligent mountain flood disaster hazard zone partitioning method of claim 1, wherein, The process of generating the context feature vector for each structure node includes: for any to-be-processed structure node, defining the structure node as a center node and defining all structure nodes connected to the center node by edges as neighborhood nodes; for each neighborhood node, selecting a corresponding edge type transformation matrix from a plurality of preset transformation matrices according to the type of the edge connecting the neighborhood node and the center node, and performing linear transformation on the feature vector of the neighborhood node; inputting the feature vector of the center node and the feature vectors of all transformed neighborhood nodes into a shared attention function to calculate normalized attention weights for each neighborhood node; and performing weighted summation on the feature vectors of the transformed neighborhood nodes by using the normalized attention weights to generate the context feature vector of the center node. 6.The method of claim 1, wherein, The gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multi-modal topographic feature map layer, and a feature saliency map and a model confidence map are generated based on the gradient. The process is as follows: for each base element grid, when the classification confidence for a predefined mode category meets a preset determination condition, the partial derivative of the classification confidence with respect to each feature dimension in the initial attribute vector of the structure node corresponding to the base element grid is calculated through the back propagation mechanism of the graph feature fusion network and the mode classifier; the absolute values of the partial derivatives are normalized and visualized as a heat map to generate the feature saliency map; and the Monte Carlo dropout technique is introduced into the graph feature fusion network and the mode classifier. The specific process is as follows: in the model inference stage, the forward propagation process with random inactivation is performed multiple times to obtain the classification confidence output for each base element grid; The variance of the classification confidence output is calculated and visualized to generate the model confidence map. 7.The method of claim 1, wherein, The process of spatial interpolation and smoothing of the classification confidence is as follows: the classification confidence of each base element grid is assigned to the geometric center of the base element grid to form a confidence point set; inverse distance weighted interpolation is performed on the confidence point set to generate a continuous classification confidence surface; and Gaussian filtering is applied to the classification confidence surface for smoothing to obtain a smoothed classification confidence surface.

8. A deep learning-based intelligent division system for mountain flood disaster hazard areas, characterized in that, The process of spatial interpolation and smoothing of the classification confidence is as follows: the classification confidence of each base element grid is assigned to the geometric center of the base element grid to form a confidence point set; inverse distance weighted interpolation is performed on the confidence point set to generate a continuous classification confidence surface; and Gaussian filtering is applied to the classification confidence surface for smoothing to obtain a smoothed classification confidence surface. The image feature extraction module receives multi-source spatial image data; and a topographic feature extractor is used to calculate and generate a multi-modal topographic feature map layer from the multi-source spatial image data; The space graph construction module: the multi-modal topographic feature layer is spatially gridded, and the region to be analyzed is divided into a basic element grid; the basic element grid is defined as a structure node; based on the spatial adjacency relationship and the elevation flow direction relationship of the structure node, a scene structure graph containing heterogeneous edge types is constructed; the graph network inference recognition module: a graph feature fusion network is constructed, and the scene structure graph is taken as input; the graph feature fusion network aggregates neighborhood information through an attention mechanism to generate a context feature vector for each structure node; the feature vector is input into a pattern classifier to output the classification confidence of each structure node; the fusion segmentation graph module: a gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multi-modal topographic feature layer, and a feature saliency map and a model confidence map are generated based on the gradient; the classification confidence is spatially interpolated and smoothed, and the feature saliency map and the model confidence map are fused to generate a semantic segmentation map of the target region and perform hazard zoning.

Images