A deep learning-based intelligent division method and system for mountain flood disaster hazard zones

Abstract

The application relates to the technical field of deep learning and image recognition, in particular to a mountain torrent disaster dangerous area intelligent division method and system based on deep learning, which comprises the following steps: calculating and generating a multi-modal landform feature layer from multi-source spatial image data; constructing a scene structure graph containing heterogeneous edge types based on the spatial adjacency relationship and the elevation flow direction relationship of structure nodes; constructing a graph feature fusion network, generating context feature vectors 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 using a gradient attribution analysis method; and processing the classification confidence, the feature saliency map and the model confidence map, generating a semantic segmentation map of a target area and performing dangerous area division. The application realizes intelligent division of a mountain torrent disaster dangerous area by constructing the scene structure graph and using the graph feature fusion network, the attention mechanism and the gradient attribution analysis method.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and image recognition technology, specifically to a method and system for intelligent delineation of flash flood hazard zones based on deep learning. Background Technology

[0002] Flash floods are sudden and destructive, making it crucial to accurately and efficiently identify and delineate danger zones.

[0003] Currently, methods for delineating flash flood hazard zones are mainly divided into two categories: those based on physical models and those driven by data. However, in the data-driven approach, existing technologies have failed to fully utilize the ability of deep learning models to perform end-to-end feature learning and pattern recognition directly from raw image data.

[0004] One mainstream approach follows a two-stage process: "manual feature extraction - traditional model classification." This process relies on predefined rules to extract discrete geographic indicators, which prevents deep neural networks from directly applying to raw remote sensing imagery. Because it fails to leverage deep learning models to automatically learn and discover hierarchical features in images, ranging from low-level texture to high-level semantics, this method not only loses crucial visual information such as topographical features and spatial context, but its models also cannot perform pixel-level dense prediction tasks, thus making it difficult to generate high-resolution semantic segmentation maps of danger zones.

[0005] Another type of technology, while applying deep learning, has network architectures and training objectives designed for different visual tasks. For example, they primarily address local object detection or classification problems, such as identifying specific objects or defects in images. However, identifying flash flood hazard zones is essentially a complex, large-scale geographic scene semantic segmentation task, requiring the model to understand the spatial combinations and association patterns of various geographic features such as terrain, water systems, and vegetation. Therefore, deep learning models designed for simple object recognition tasks are not suitable for analyzing this macroscopic scene due to their unsuitable structure and learning capabilities.

[0006] In summary, a clear gap exists in the existing technology: the lack of a deep learning solution specifically designed for understanding flash flood landforms. Traditional methods cannot leverage 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, the field needs a new technical solution capable of constructing a deep learning model that can directly process multi-source remote sensing imagery, completing the entire task from feature representation to semantic segmentation of danger zones in an end-to-end manner.

[0007] To address this, a method and system for intelligent delineation of flash flood hazard zones based on deep learning is proposed. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for intelligent delineation of flash flood hazard zones based on deep learning. By constructing the aforementioned scene structure graph and employing the aforementioned graph feature fusion network, attention mechanism, and gradient attribution analysis method, intelligent delineation of flash flood hazard zones is achieved.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A deep learning-based intelligent method for classifying flash flood hazard zones includes:

[0011] Receive multi-source spatial image data; use a geomorphic feature extractor to calculate and generate a multimodal geomorphic feature layer from the multi-source spatial image data;

[0012] The multimodal landform feature layer is spatially gridded, and the region to be analyzed is divided into primitive grids; the primitive grids are defined as structural nodes; based on the spatial adjacency relationship and elevation flow direction relationship of the structural nodes, a scene structure graph containing heterogeneous edge types is constructed.

[0013] 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 context feature vectors for each structure node; the feature vectors are input to a pattern classifier, which outputs the classification confidence of each structure node;

[0014] The gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multimodal landform 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 to delineate the danger zone.

[0015] Preferably, the multi-source spatial image data specifically includes digital elevation model data, geological maps, soil maps, and remote sensing image data; the multimodal geomorphic feature layer specifically includes a geometric feature layer, a physical feature layer, and a cover feature layer; the geometric feature layer includes slope, aspect, and plane curvature calculated by processing the digital elevation model data; the physical feature layer includes surface roughness and soil infiltration rate obtained by querying the geological map and soil map data; and the cover feature layer is the normalized vegetation index calculated by processing the remote sensing image data.

[0016] Preferably, the process of constructing a scene structure graph containing heterogeneous edge types specifically involves: defining the primitive mesh as a structure node of the scene structure graph; establishing a set of heterogeneous edge types, wherein: a spatial adjacency edge is established between any two structure nodes that share a boundary in space; calculating, based on the digital elevation model data, an elevation flow direction edge is established from a structure node from which surface runoff flows out to a structure node from which it flows in; the spatial adjacency edge and the elevation flow direction edge together constitute the set of heterogeneous edge types of the scene structure graph; obtaining the raster values ​​of all multimodal landform feature layers within the primitive mesh corresponding to the structure node, calculating and splicing the mean and standard deviation of the raster values ​​of the multimodal landform feature layers to form the initial attribute vector of the structure node.

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

[0018] Preferably, the process of generating context feature vectors for each structural node includes: defining any structural node to be processed as a center node, and defining all structural nodes connected to the center node by edges as neighboring nodes; for each neighboring node, selecting a corresponding edge type transformation matrix from a preset set of multiple transformation matrices according to the type of the edge connecting the neighboring node and the center node, and performing a linear transformation on the feature vector of the neighboring node; inputting the feature vector of the center node and the feature vectors of all transformed neighboring nodes into a shared attention function to calculate a normalized attention weight for each neighboring node; and using the normalized attention weight to perform a weighted summation on the feature vectors of the transformed neighboring 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 multimodal landform feature layer, and a feature saliency map and a model confidence map are generated based on the gradient. The specific process is as follows: For each primitive grid, when the classification confidence for a predefined pattern category meets the preset judgment condition, the partial derivatives of the classification confidence with respect to each feature dimension in the initial attribute vector of the corresponding structural node of 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 with a heatmap to generate the feature saliency map; Monte Carlo dropout technology is introduced into the graph feature fusion network and the pattern classifier. The specific process is as follows: During the model inference stage, the forward propagation process with random deactivation is executed multiple times to obtain the classification confidence output for each primitive grid; the variance of the classification confidence output is calculated and visualized to generate the model confidence map.

[0020] Preferably, the process of spatial interpolation and smoothing of the classification confidence score specifically involves: assigning the classification confidence score of each primitive grid to the geometric center of the primitive grid to form a confidence score point set; performing inverse distance weighted interpolation on the confidence score point set to generate a continuous classification confidence score surface; and applying Gaussian filtering to the classification confidence score surface for smoothing to obtain a smoothed classification confidence score surface.

[0021] A deep learning-based intelligent delineation system for flash flood hazard zones includes: an image feature extraction module: receiving multi-source spatial image data; using a geomorphic feature extractor to calculate and generate a multimodal geomorphic feature layer from the multi-source spatial image data; a spatial map construction module: spatially gridding the multimodal geomorphic feature layer, dividing the area to be analyzed into primitive grids; defining the primitive grids as structural nodes; and constructing a scene structure map containing heterogeneous edge types based on the spatial adjacency relationships and elevation-flow direction relationships of the structural nodes; and a graph network reasoning and recognition module: constructing a graph feature fusion network, using the scene structure map as... Input: The graph feature fusion network aggregates neighborhood information through an attention mechanism to generate context feature vectors for each structural node; the feature vectors are input to a pattern classifier, which outputs the classification confidence of each structural node; fusion segmentation and graphing module: using gradient attribution analysis, the gradient of the classification confidence with respect to the input multimodal landform feature layer is calculated, and a feature saliency map and a model confidence map are generated based on the gradient; spatial interpolation and smoothing are performed on the classification confidence, and the feature saliency map and the model confidence map are fused to generate a semantic segmentation map of the target region and perform danger zone delineation.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. This invention solves the major technical challenge of traditional raster methods failing to effectively model non-Euclidean relationships in geographic space by abstracting geographic space into a scene structure graph containing two heterogeneous edge types: "spatial adjacency" and "elevation-direction." Traditional methods can only handle local, regular neighborhood relationships, failing to capture long-distance physical connections such as surface runoff generated along irregular paths. The heterogeneous map construction method of this invention can simultaneously and accurately express the proximity of geographic topology and the causality of hydrological processes, resulting in a higher consistency between the data model and the real geographic environment, providing a scientific and robust data structure foundation for the accuracy of subsequent analysis.

[0024] 2. This invention employs a graph feature fusion network and introduces an attention mechanism to achieve dynamic and intelligent reasoning about complex spatial relationships, overcoming the limitations of traditional models in feature fusion. Traditional models often assign equal or fixed weights to all spatial influencing factors, failing to identify the primary disaster-causing factors in a specific geographical context. The attention mechanism of this invention can automatically learn and allocate different weights based on the node's own characteristics and spatial connectivity, enabling the generated contextual feature vector to better reflect the comprehensive impact of key geographical elements. This significantly improves the accuracy and robustness of pattern recognition, making the delineation of danger zones more consistent with real physical disaster-causing patterns.

[0025] 3. This invention addresses the black-box problem prevalent in existing deep learning models for disaster assessment by introducing gradient attribution analysis and fusing it with generated feature saliency maps and model confidence maps. Traditional methods only output a prediction result but cannot explain its causes and reliability, making it difficult for decision-makers to fully trust and apply them. This invention not only provides a semantic segmentation map indicating the location of the danger zone but also clearly reveals the dominant causes of the danger through feature saliency maps and quantifies the credibility of the prediction results through a "model confidence map." This integrated output of "prediction-interpretation-reliability assessment" provides more comprehensive and reliable decision support for disaster prevention and control. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a deep learning-based intelligent classification method for flash flood hazard zones according to the present invention.

[0027] Figure 2 This is a time sequence diagram of a method and system for intelligent delineation of flash flood hazard zones based on deep learning, as proposed in this invention.

[0028] Figure 3 This is a schematic diagram of the structure of a deep learning-based intelligent classification system for flash flood hazard zones according to the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see Figures 1 to 3 This invention provides a method and system for intelligent delineation of flash flood hazard zones based on deep learning, the technical solution of which is as follows:

[0031] A deep learning-based intelligent method for classifying flash flood hazard zones includes:

[0032] Receive multi-source spatial image data; use a geomorphic feature extractor to calculate and generate a multimodal geomorphic feature layer from the multi-source spatial image data;

[0033] The multimodal landform feature layer is spatially gridded, and the region to be analyzed is divided into primitive grids; the primitive grids are defined as structural nodes; based on the spatial adjacency relationship and elevation flow direction relationship of the structural nodes, a scene structure graph containing heterogeneous edge types is constructed.

[0034] 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 context feature vectors for each structure node; the feature vectors are input to a pattern classifier, which outputs the classification confidence of each structure node;

[0035] The gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multimodal landform 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 to delineate the danger zone.

[0036] Example 1

[0037] This embodiment provides a specific application of a deep learning-based intelligent method for classifying flash flood hazard zones. Taking the classification of flash flood hazard zones in a mountainous watershed as an example, the specific implementation of the invention is described. (Refer to...) Figure 1 , Figure 2 This embodiment follows the process of a deep learning-based intelligent classification method for flash flood hazard zones provided by the present invention, and demonstrates its data flow. The specific implementation process is as follows:

[0038] Furthermore, multi-source spatial image data is received; a geomorphic feature extractor is used to calculate and generate a multimodal geomorphic feature layer from the multi-source spatial image data. The specific process includes:

[0039] A mountainous watershed, covering approximately 200 square kilometers, was selected as the area to be analyzed. First, multi-source spatial imagery data was received, specifically including 30-meter resolution digital elevation model (DEM) data, 10-meter resolution remote sensing imagery data, and 1:50,000 scale geological and soil maps. Next, a geomorphic feature extractor was used to calculate the above data, generating a multimodal geomorphic feature layer. Specifically, slope, aspect, and plane curvature were calculated from the DEM data to form a geometric feature layer; surface roughness and soil infiltration rate were obtained from the geological and soil map data to form a physical feature layer; and the normalized vegetation index (NDI) was calculated from the remote sensing imagery data to form a cover feature layer.

[0040] By explicitly defining the multimodal geomorphic feature layers into three categories—geometric, physical, and cover—and specifying key features such as slope, surface roughness, and normalized difference vegetation index, the model was provided with comprehensive input information closely related to the mechanism of flash flood disasters. This ensured that the basic data for model analysis was sufficient and physically meaningful, improving the accuracy and physical authenticity of the final classification results.

[0041] Furthermore, the multimodal terrain feature layer is spatially gridded, dividing the region to be analyzed into a primitive grid; the primitive grid is defined as a structural node; based on the spatial adjacency relationship and elevation flow direction relationship of the structural nodes, a scene structure graph containing heterogeneous edge types is constructed. The specific process includes:

[0042] All multimodal geomorphic feature layers were spatially gridded at 30m × 30m, dividing the area to be analyzed into approximately 220,000 primitive grids. Each primitive grid was defined as a structural node. Based on these structural nodes, a scene structure graph containing heterogeneous edge types was constructed: a spatial adjacency edge was established between structural nodes corresponding to any two primitive grids sharing a spatial boundary; and an elevation flow direction edge was established from the structural node corresponding to the primitive grid from which surface runoff flows out, to the structural node corresponding to the primitive grid into which it flows in, calculated based on the digital elevation model data. These two types of edges together constitute the heterogeneous edge set of the scene structure graph.

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

[0044] By identifying the target primitive grid with the largest elevation difference among eight adjacent primitive grids, the surface runoff direction is ensured to be unique for each primitive grid that is not a confluence point. This avoids path ambiguity that may occur when constructing the graph topology, and makes the final generated scene structure graph reproducible and structurally unique.

[0045] For each structural node in the graph, the mean and standard deviation of the raster values ​​of all multimodal landform feature layers within its corresponding primitive grid are extracted, and these values ​​are concatenated to form a 12-dimensional initial attribute vector. Through the above steps, a scene structure map of the region to be analyzed is obtained.

[0046] By establishing spatial adjacency edges and elevation flow direction edges separately, a concrete technical implementation is provided for constructing the scene structure graph. This approach enables the graph's topology to simultaneously represent two different physical relationships: geographic spatial proximity and surface runoff directionality determined by elevation. By calculating and concatenating the mean and standard deviation of the raster values ​​from the multimodal geomorphic feature layer, the initial attribute vectors of the nodes are constructed. This ensures that the initial state of the nodes reflects not only the central tendency of regional features but also the degree of internal variation of the features, thus providing more representative input information for the subsequent graph feature fusion network.

[0047] Furthermore, 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 context feature vectors for each structure node; the feature vectors are input to a pattern classifier, which outputs the classification confidence of each structure node. The specific process includes:

[0048] Construct a graph feature fusion network and take the scene structure graph generated in the previous step as input.

[0049] The graph feature fusion network includes multiple 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 adds the intermediate feature vector to the input feature vector element by element 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 containing three stacked graph feature fusion layers. The network's input layer receives 12-dimensional initial attribute vectors from each structural node in the scene structure graph. The first graph feature fusion layer receives the 12-dimensional initial attribute vectors as input. This layer aggregates neighborhood information through the attention mechanism and 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 from the previous layer as input, uses the same attention mechanism to aggregate neighborhood information, and outputs a 64-dimensional intermediate feature vector. The third graph feature fusion layer receives the 64-dimensional intermediate feature vector output from the previous layer as input, performs neighborhood information aggregation again, 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 modified linear unit is used as the activation function to enhance the model's learning effect. In each layer, the aggregated intermediate feature vector is element-wise added to the input feature vector of that layer to form a residual connection, ensuring that the gradient propagates effectively in the deep network.

[0051] By introducing an operation that adds intermediate feature vectors to input feature vectors element by element between stacked graph feature fusion layers, residual connections are constructed. This helps to alleviate the gradient vanishing problem that may occur during the training of deep graph networks and enables the network to pass raw feature information from shallow nodes to deeper layers, thereby supporting the construction of deeper graph feature fusion networks.

[0052] When generating a context feature vector for any central node, for each of its neighboring nodes, a corresponding edge type transformation matrix is ​​selected from a set of preset transformation matrices based on the type of connecting edges. This matrix is ​​then linearly transformed into a 64-dimensional feature vector for that neighboring node. Finally, the feature vector of the central node and the transformed feature vectors of all neighboring nodes are input into a shared attention function to calculate normalized attention weights for each neighboring node.

[0053] By assigning dedicated learnable transformation matrices to different types of edges, the attention mechanism can differentiate information from neighborhoods with different relationships. The model can thus learn different influence patterns on the central node, thereby more precisely 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 perform a weighted summation of the feature vectors of the transformed neighboring nodes to generate a 64-dimensional context feature vector for the center node. This 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 output of the output layer nodes, thereby generating a classification confidence score for each predefined pattern category for each structural node.

[0055] By calculating and weighting the normalized attention weights of nodes in the entire neighborhood, this invention ensures that the final feature representation of each structural node dynamically and adaptively integrates the information of all its important neighbors, thereby enhancing the model's comprehensive perception and discrimination capabilities for complex local environments.

[0056] Furthermore, a gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multimodal landform feature layer, and a feature saliency map and a model confidence map are generated based on the gradient; spatial interpolation and smoothing are performed on the classification confidence, and the feature saliency map and model confidence map are fused to generate a semantic segmentation map of the target region and perform danger zone delineation. The specific process includes:

[0057] After obtaining the classification confidence scores of each structural node, a gradient attribution analysis method is employed. For each primitive grid cell, when its classification confidence score for a predefined pattern category ("high-risk" pattern in this example) meets a preset criterion (e.g., confidence score greater than 0.8), the partial derivatives of the classification confidence score with respect to each feature dimension in the initial attribute vector of the corresponding structural node of the primitive grid cell are calculated using the backpropagation mechanism of the graph feature fusion network and the pattern classifier. The absolute values ​​of the partial derivatives are normalized and visualized using a heatmap to generate the feature saliency map.

[0058] This invention generates a feature saliency map by calculating partial derivatives through backpropagation, thereby visualizing the model's decision-making. It can clearly reveal the key geomorphic features that lead to a certain area being judged as a danger zone, providing an intuitive scientific basis for disaster prevention and control. Furthermore, by introducing the Monte Carlo dropout technique to quantify the uncertainty of the model's predictions, this invention can avoid overconfident and erroneous classifications, thus improving the safety and reliability of the entire classification method.

[0059] Simultaneously, Monte Carlo dropout technology is introduced into the graph feature fusion network and pattern classifier, performing a forward propagation process with a dropout rate of 0.2 50 times. The variance of the 50 output results is calculated to generate the model confidence map. Finally, the classification confidence of each primitive grid is assigned to the center point of that grid, and an inverse distance weighted interpolation method is used to generate a classification confidence surface. This surface is then smoothed using a Gaussian filter with a window size of 3×3 to obtain a smoothed classification confidence surface. The smoothed classification confidence surface is used as the base hazard 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. The base hazard level layer, the confidence overlay layer, and the attribution explanation layer together constitute the semantic segmentation map of the target region.

[0060] By concretizing the final semantic segmentation map into a multi-layered data structure that includes basic hazard levels, confidence level stacking, and attribution explanations, the system simultaneously presents the predicted risk level, the reliability assessment of the prediction, and the key feature basis for making the prediction. This multi-dimensional output structure provides users with more comprehensive decision-making reference information than a single risk layer, enhancing the overall information carrying capacity and practical application value of the segmentation results.

[0061] By employing inverse distance weighted interpolation and Gaussian filtering for smoothing, the grid effect and noise that may occur during the calculation process are eliminated, resulting in a more natural boundary for the final generated flash flood hazard zoning map and a result that better matches the actual geographical distribution, thus improving the practicality and visualization effect of the results.

[0062] This invention proposes an end-to-end intelligent partitioning framework. By abstracting discrete geospatial data into a scene structure graph containing heterogeneous edge types, and using a graph feature fusion network for analysis, it can effectively capture the complex spatial adjacency relationships between geomorphic units and the elevation-flow direction relationships in hydrology, overcoming the problem that traditional methods struggle to model complex spatial dependencies. Simultaneously, by introducing gradient attribution analysis, it not only generates partitioning results but also provides feature saliency maps and model confidence maps as the basis for model decision-making, improving the interpretability and reliability of the partitioning results.

[0063] Example 2

[0064] This embodiment aims to detail the specific workflow of the image feature extraction module and the spatial map construction module in this invention. In this embodiment, the scenario involves identifying flash flood hazard zones in a high mountain canyon area. The goal is to use the method of this invention to perform refined data modeling and spatial relationship construction for this area, providing high-quality input for subsequent risk identification.

[0065] See Figure 3 The system structure of the present invention is shown in the figure.

[0066] First, the image feature extraction process is executed. Multi-source spatial image data is received, specifically including 5-meter resolution digital elevation model data, a 1:25000 scale geological map, and 2-meter resolution remote sensing image data. A geomorphic feature extractor is used to process the above data, generating multimodal geomorphic feature layers: by processing the digital elevation model data, slope, aspect, and plane 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; and by processing the remote sensing image data, the normalized vegetation index is calculated to form a cover feature layer.

[0067] Next, the spatial map construction process is executed. All multimodal geomorphic feature layers are spatially gridded at 10m x 10m, dividing the approximately 50 square kilometer area to be analyzed into nearly 500,000 primitive grids. The process of constructing a scene structure map containing heterogeneous edge types is as follows: The primitive grids are defined as structural nodes of the scene structure map; a set of heterogeneous edge types is established, wherein: a spatial adjacency edge is established between any two structural nodes that share a spatial boundary; based on the digital elevation model data, an elevation flow direction edge is established from a structural node from which surface runoff flows to the structural node into which it flows. The spatial adjacency edge and the elevation flow direction edge together constitute the set of heterogeneous edge types of the scene structure map. Finally, the raster values ​​of all multimodal geomorphic feature layers within the primitive grid corresponding to the structural node are obtained, and the mean and standard deviation of the raster values ​​of the multimodal geomorphic feature layers are calculated and concatenated to form a 10-dimensional initial attribute vector for the structural node. Through the above steps, the scene structure map of the area to be analyzed is obtained.

[0068] Example 3

[0069] This embodiment aims to detail the specific workflow of the graph network reasoning and identification module and the fusion segmentation and graph generation module in this invention. In this embodiment, the scenario involves identifying flash flood hazard zones in a plain river network area. The goal is to use the method of this invention to perform intelligent reasoning and analysis on the constructed scene structure graph, generating hazard zone delineation results that are both interpretable and reliable.

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

[0071] This embodiment starts with a pre-constructed scene structure graph and directly enters the graph network inference and recognition process. The structural nodes of this scene structure graph represent water catchment units, and their initial attribute vectors already contain statistical values ​​of features such as slope, aspect, plane curvature, and soil infiltration rate. This scene structure graph is used as input and fed into a pre-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 central node, for each of its neighboring nodes, a learnable transformation matrix specific to that edge type is used for processing. The original feature vector of the central node and the transformed feature vectors of the neighboring nodes are input together into the attention function to calculate the attention weights, and finally aggregated to generate a 128-dimensional context feature vector. This context feature vector is then input into a pattern classifier, which outputs the classification confidence of each structural node.

[0072] Finally, the fusion and segmentation mapping process is executed. Using gradient attribution analysis, for each primitive grid identified as a specific target pattern, the partial derivative of its classification confidence with respect to the input multimodal terrain feature layer is calculated via backpropagation to generate a feature saliency map. Simultaneously, during the model inference stage, Monte Carlo dropout technology is introduced into the graph feature fusion network and pattern classifier. This includes executing forward propagation 50 times with a dropout rate of 0.2; and 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 that grid, and an inverse distance weighted interpolation method is used to generate a classification confidence surface. A Gaussian filter with a window size of 5×5 is applied to smooth the confidence surface. 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 region.

[0073] This invention provides a modular and automated intelligent classification system. The system integrates all functions from data reception, feature extraction, graph construction, network inference to result generation, achieving automated and efficient processing of the task of classifying flash flood hazard zones.

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

Claims

1. A method for intelligent delineation of flash flood hazard zones based on deep learning, characterized in that, include: Receive multi-source spatial image data; use a geomorphic feature extractor to calculate and generate a multimodal geomorphic feature layer from the multi-source spatial image data; The multimodal landform feature layer is spatially gridded, and the region to be analyzed is divided into primitive grids; the primitive grids are defined as structural nodes. Based on the spatial adjacency and elevation flow relationships of structural nodes, a scene structure graph containing heterogeneous edge types is constructed. 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 context feature vectors for each structure node; the feature vectors are input to a pattern classifier, which outputs the classification confidence of each structure node; The gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multimodal landform 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 to delineate the danger zone.

2. The method for intelligent delineation of flash flood hazard zones based on deep learning according to claim 1, characterized in that, The multi-source spatial image data specifically includes digital elevation model data, geological maps, soil maps, and remote sensing image data; the multimodal geomorphic feature layer specifically includes a geometric feature layer, a physical feature layer, and a cover feature layer; the geometric feature layer contains slope, aspect, 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; the cover feature layer is the normalized vegetation index calculated by processing the remote sensing image data.

3. The method for intelligent delineation of flash flood hazard zones based on deep learning according to claim 2, characterized in that, The process of constructing a scene structure graph containing heterogeneous edge types is as follows: defining the primitive mesh as the structure node of the scene structure graph; establishing a set of heterogeneous edge types, wherein: a spatial adjacency edge is established between any two structure nodes that share a boundary in space; calculating, based on the digital elevation model data, an elevation flow direction edge is established from a structure node from which surface runoff flows out to a structure node from which it flows in; the spatial adjacency edge and the elevation flow direction edge together constitute the set of heterogeneous edge types of the scene structure graph; obtaining the raster values ​​of all multimodal landform feature layers within the primitive mesh corresponding to the structure node, calculating and splicing the mean and standard deviation of the raster values ​​of the multimodal landform feature layers to form the initial attribute vector of the structure node.

4. The method for intelligent delineation of flash flood hazard zones based on deep learning according to claim 1, characterized in that, The specific implementation of the attention mechanism is as follows: For any two structural nodes connected by an edge, when calculating the attention weight between them, one structural node is regarded as the center node and the other structural node is regarded as the neighbor node; according to the type of the edge connecting the center node and the neighbor node, a learnable transformation matrix specific to the edge type is used to transform the feature vector of the neighbor node. The original feature vector of the central node and the transformed feature vectors of the neighboring nodes are input into the attention function to calculate the attention weight.

5. The method for intelligent delineation of flash flood hazard zones based on deep learning according to claim 1, characterized in that, The process of generating context feature vectors for each structural node includes: for any structural node to be processed, it is defined as a center node, and all structural nodes connected to the center node by edges are defined as neighboring nodes; for each neighboring node, according to the type of the edge connecting the neighboring node and the center node, a corresponding edge type transformation matrix is ​​selected from a preset set of multiple transformation matrices, and a linear transformation is performed on the feature vector of the neighboring node; the feature vector of the center node and the feature vectors of all the transformed neighboring nodes are input into a shared attention function to calculate a normalized attention weight for each neighboring node; the normalized attention weights are used to perform a weighted summation on the feature vectors of the transformed neighboring nodes to generate the context feature vector of the center node.

6. The method for intelligent delineation of flash flood hazard zones based on deep learning according to claim 1, characterized in that, The gradient attribution analysis method is used to calculate the gradient of the classification confidence with respect to the input multimodal landform feature layer, and to generate a feature saliency map and a model confidence map based on the gradient. Specifically, for each primitive grid, when the classification confidence for a predefined pattern 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 corresponding structural node of 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 using a heatmap to generate the feature saliency map. Monte Carlo dropout technology is introduced into the graph feature fusion network and the pattern classifier. Specifically, during the model inference stage, a forward propagation process with random deactivation is executed multiple times to obtain the classification confidence output for each primitive grid. Calculate the variance of the classification confidence output, visualize the variance, and generate the model confidence map.

7. The method for intelligent delineation of flash flood hazard zones based on deep learning according to claim 1, characterized in that, The process of spatial interpolation and smoothing of the classification confidence is as follows: the classification confidence of each primitive grid is assigned to the geometric center of the primitive grid to form a confidence point set; inverse distance weight 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 to obtain a smoothed classification confidence surface.

8. A deep learning-based intelligent classification system for flash flood hazard zones, characterized in that, include: Image feature extraction module: Receives multi-source spatial image data; uses a geomorphic feature extractor to calculate and generate a multimodal geomorphic feature layer from the multi-source spatial image data; The module performs spatial gridding on the multimodal landform feature layer, dividing the area to be analyzed into primitive grids; defining primitive grids as structural nodes; and constructing a scene structure map containing heterogeneous edge types based on the spatial adjacency and elevation-flow direction relationships of the structural nodes. The graph network reasoning and recognition module constructs a graph feature fusion network, taking the scene structure map as input; the graph feature fusion network aggregates neighborhood information through an attention mechanism to generate contextual feature vectors for each structural node; the feature vectors are input to a pattern classifier, outputting the classification confidence of each structural node. The fusion and segmentation mapping module uses gradient attribution analysis to calculate the gradient of the classification confidence with respect to the input multimodal landform feature layer, and generates a feature saliency map and a model confidence map based on the gradient; spatial interpolation and smoothing are performed on the classification confidence, and the feature saliency map and model confidence map are fused to generate a semantic segmentation map of the target area and delineate danger zones.

Images