Intelligent identification method for sudden geological disaster hidden points based on deep learning

Abstract

This invention relates to the field of image feature processing technology, specifically to an intelligent identification method for sudden geological disaster hazard points based on deep learning. The method extracts significant deformation locations, surface vegetation coverage, and local slope matrices, and determines the direction of water flow and its endpoint valley location based on the local slope distribution. Safety weights are calculated based on the local slope and vegetation coverage at each location. A topology graph is constructed using grid nodes and directed edges. A maximum flow minimum cut algorithm is performed on the topology graph to obtain the initial hazard boundary line. Subsequently, the initial hazard boundary line is truncated and verified based on the elevation distribution of significant deformation locations within each sub-region, removing plain subsidence artifacts to obtain the final hazard boundary line. After regional expansion, target identification is performed, and the hazard point identification result is output. This invention achieves adaptive scientific demarcation of the hazard impact range, significantly reducing the false alarm rate in complex terrain scenarios.

Description

Technical Field

[0001] This invention relates to the field of image feature processing technology, specifically to a method for intelligent identification of potential geological disaster sites based on deep learning. Background Technology

[0002] In the field of monitoring and investigating widespread geological hazards (such as landslides and ground subsidence), the integrated use of synthetic aperture radar interferometry (InSAR) to obtain surface deformation rates, inversion of surface vegetation cover from multispectral optical satellite imagery, and generation of high-precision digital elevation models (DEMs) and local slope matrices from lidar point clouds are currently standard techniques for conducting large-scale geological hazard surveys. The aforementioned multi-source data provides rich fundamental information for assessing surface stability.

[0003] In the practice of using the aforementioned multi-source data for hazard identification and spatial location, a common approach is to define the search or analysis area by using a circular or square buffer zone with a fixed radius centered on a significant deformation point (a subsidence / slippage point identified by radar). However, the Earth's surface is highly heterogeneous and spatially anisotropic. A fixed-radius buffer zone completely ignores the actual physical barriers of the surface (such as the anchoring effect of dense forests on the soil or the obstruction of crack propagation by steep rock walls). This leads to situations in complex terrain where either the buffer zone radius is too small, missing cracks that actually extend along ridges or fracture zones (underreporting); or the radius is too large, including large-scale, slow-deformation, non-hazardous subsidence points in plains areas caused by pumping or farmland tillage in the analysis area, resulting in numerous invalid calculations and false alarms.

[0004] Simply filtering the identification area based on topographic elevation differences, relying solely on local, absolute thresholds for the difference between start and end points, has inherent flaws in flat alluvial plains or gently sloping hilly areas. Because water flow paths are extremely long in plains, even a slight slope (e.g., one-thousandth) can accumulate over long distances (several kilometers) to result in an absolute elevation difference of several meters. Traditional thresholding methods cannot distinguish between "short-distance steep drops" (genuine landslides) and "long-distance gentle drops" (plain artifacts), leading the system to frequently misclassify large areas of slightly sloping plains as potential landslide hazards. Summary of the Invention

[0005] To address the technical problem of existing technologies failing to effectively determine the target inspection area during the intelligent identification of potential geological disaster sites, the present invention aims to provide a deep learning-based intelligent identification method for potential geological disaster sites. The specific technical solution adopted is as follows: This invention proposes a deep learning-based intelligent identification method for potential sudden geological disaster sites, the method comprising: Based on the synthetic aperture radar deformation rate at various locations in the target area, locations with significant deformation are selected; the local slope matrix corresponding to the target area is obtained from the lidar point cloud; the water flow direction at each location is determined based on the distribution of local slope in the local slope matrix, and the endpoint of the water flow direction is taken as the location of the gully; the surface vegetation cover at each location is obtained using multispectral optical satellite imagery. For each location, a safety weight is obtained based on the local slope and surface vegetation cover. Each location is considered as a node, and there are directed edges between adjacent nodes. The edge weight is the safety weight pointing to the node, thus obtaining a topology map of the target area. The maximum flow minimum cut algorithm is executed on the topology map to obtain the initial hazard boundary line. In the maximum flow minimum cut algorithm, the edge weight between the virtual water flow source node and the significantly deformed location is set to a preset maximum value, and the edge weight between the virtual water flow sink node and the valley location is set to a preset maximum value. The initial hazard boundary line divides the target area into multiple sub-regions. Based on the elevation distribution of significant deformation locations in the sub-regions, the initial hazard boundary line is intercepted to obtain the final hazard boundary line. The final hazard boundary line is expanded to obtain the hazard area; targets are identified within the hazard area to obtain the hazard point identification results.

[0006] Furthermore, the method for constructing the local slope matrix includes: For any target location, obtain the neighborhood locations within a preset neighborhood range centered on the target location. The neighborhood locations with elevation values ​​lower than the target location's elevation value are designated as low-elevation locations, and the low-elevation location with the minimum elevation value is designated as the comparison location. The local slope of the target location is obtained by using the elevation difference between the target location and the comparison location as the numerator and the horizontal distance between the target location and the comparison location as the denominator. If no corresponding low-elevation location exists for the target location, the local slope is set to 0. The local slope matrix is ​​obtained by statistically analyzing the local slopes at all locations.

[0007] Furthermore, the method for determining the location of the valley includes: The direction from the target location to the comparison location is taken as the water flow direction at the target location, and the water flow direction field of the target area is obtained. In the water flow direction field, start from any position and trace along the water flow direction until the end of the path has no water flow direction. The end of the path is then taken as the location of the gully.

[0008] Furthermore, the method for obtaining the security weight includes: For any given location, the result of mapping and normalizing the negative correlation of the local slope is weighted and summed with the normalized result of the surface vegetation cover according to a preset weight to obtain the initial safety weight; the sum of the initial safety weight and the minimum value of the preset weight is taken as the safety weight.

[0009] Furthermore, the method for obtaining the final hazard boundary line includes: For any significant deformation location within a sub-region, the elevation value corresponding to the significant deformation location is the starting elevation value; the water flow path starting from the significant deformation location is obtained according to the water flow direction, the elevation value corresponding to the end of the path is taken as the ending elevation value, the difference between the starting elevation value and the ending elevation value is taken as the numerator, and the horizontal distance between the significant deformation location and the end of the path is taken as the denominator to obtain the average slope gradient of the significant deformation location; If the average gradient of all significant deformation locations within a sub-region is less than a preset gradient threshold, then the initial hazard boundary line corresponding to the sub-region boundary is removed; traverse all sub-regions to obtain the final hazard boundary line.

[0010] Furthermore, the method for obtaining the potential hazard area includes: For each location on the final hazard boundary line, the product of the local slope at the corresponding location and the preset conversion coefficient is used as the radius adjustment amount, and the sum of the radius adjustment amount and the initial expansion radius is used as the final expansion radius; a circular area is constructed with the corresponding location as the center and the final expansion radius as the radius; the union of the circular areas formed by each location on the final hazard boundary line is used as the hazard area.

[0011] Furthermore, the target identification of the potential hazard area includes: Extract the three-channel color base map matrix of the corresponding location of the potential hazard area in the multispectral optical satellite image; during the extraction process, a smooth dimensionality reduction method is used at the edge of the potential hazard area; The target detection network is used to identify targets in the three-channel color base map matrix.

[0012] Furthermore, obtaining the hazard identification results includes: If the target detection network generates multiple target bounding boxes, the cross-union ratio (CUI) between the target bounding boxes is obtained. Target bounding boxes with a CUI greater than a preset CUI threshold are merged. All target bounding boxes obtained after merging are used as the hazard point identification results.

[0013] Furthermore, the screening of locations with significant deformation includes: If the absolute value of the synthetic aperture radar deformation rate at a certain location is greater than a preset comparison threshold, then the corresponding location is designated as the significant deformation location.

[0014] Furthermore, the method for obtaining the surface vegetation coverage includes: Obtain the normalized vegetation index for each location, and plot a histogram of the normalized vegetation index. Select the baseline values ​​for bare surface distribution and vegetation cover distribution on the histogram. Use the baseline values ​​for bare surface distribution and vegetation cover distribution as the data basis to linearly normalize the normalized vegetation index to obtain the surface vegetation cover.

[0015] The present invention has the following beneficial effects: This invention introduces a local slope matrix and surface vegetation cover to construct safety weights that characterize surface heterogeneity. A topological graph is then built using grid nodes and directed edges, and a maximum flow minimum cut algorithm is executed. This replaces the crude approach of extending a fixed-radius buffer zone centered on significant deformation points in existing technologies. Because the safety weights comprehensively reflect the surface inclination and the anchoring capacity of vegetation roots, the edge weights objectively characterize the physical barriers or conduction characteristics of deformation transmission at different spatial locations. This allows the initial hazard boundary line calculated by the maximum flow minimum cut algorithm to adaptively extend along the surface's weak mechanical zones, rather than mechanically expanding outwards at equal intervals. This mechanism effectively overcomes the problem of excessive extension or omission of areas caused by the fixed-radius buffer failing to reflect surface heterogeneity, achieving objective and dynamic spatial demarcation of the hazard's impact range.

[0016] After obtaining the initial hazard boundary line, the boundary line is further processed by extracting sections based on the elevation distribution of significant deformation locations within each sub-region delineated by the initial boundary line. This step, by analyzing the elevation distribution characteristics of significant deformation points within the sub-region, can effectively distinguish whether the driving force of the landslide is a true landslide characterized by a sudden change in elevation over a short distance, or a plain surface disturbance caused by the accumulation of small slopes over a long distance. Compared to the existing approach that relies solely on local absolute elevation thresholds for discrimination, this invention takes a holistic approach from the elevation distribution of the sub-region, filtering out plain subsidence artifacts that lack actual gravity-driven conditions simply because the elevation values ​​meet the standards, thus significantly reducing the interference of non-hazardous surface deformation on hazard identification.

[0017] In summary, this invention achieves accurate delineation of areas affected by sudden geological disaster risks through graph theory dynamic delimitation empowered by security weights and regional interception verification based on the elevation distribution of significant deformation points, effectively improving the accuracy and reliability of hazard identification in complex terrain scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The flowchart illustrates a method for intelligent identification of potential geological disaster hazard points based on deep learning, as provided in one embodiment of the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a deep learning-based intelligent identification method for sudden geological disaster hazard points proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] The following description, in conjunction with the accompanying drawings, details a specific scheme for an intelligent identification method for sudden geological disaster hazard points based on deep learning, provided by this invention.

[0023] Please see Figure 1 The diagram illustrates a flowchart of a deep learning-based intelligent identification method for sudden geological disaster hazard points according to an embodiment of the present invention. The method includes: Step S1: Based on the synthetic aperture radar deformation rate at various locations in the target area, select locations with significant deformation; obtain the surface vegetation coverage at each location using multispectral optical satellite imagery; obtain the local slope matrix corresponding to the target area based on the lidar point cloud; determine the water flow direction at each location based on the distribution of local slope in the local slope matrix, and take the endpoint of the water flow direction as the valley location.

[0024] In this embodiment of the invention, the aim is to fuse features from three types of data: synthetic aperture radar deformation rate, multispectral optical satellite imagery, and lidar point cloud data, thereby delineating effective hazard areas in the target region for target identification. Therefore, it is first necessary to determine the data in the three dimensions for the current target region. Since the three dimensions of data have different time spans—for example, synthetic aperture radar deformation rate represents the speed of surface movement along the radar line of sight over a period of time, while multispectral optical satellite imagery and lidar point cloud data are typically instantaneous data from a specific day or moment—it is also necessary to calibrate the three types of data before proceeding with this embodiment of the invention. Specifically, this includes: (1) Set the time period between the start date and the end date of the synthetic aperture radar deformation measurement, where the end date is usually the date of implementation of this embodiment. Use this time period as the target time period to perform synthetic aperture radar deformation rate statistics to obtain the rate matrix of the current target area. This rate matrix is ​​essentially a two-dimensional map. In this embodiment, the data of the three dimensions can be represented by grid data. Each location under the target area is a grid on the grid network, and the rate matrix is ​​the data representation result of the deformation rate of each grid on the grid network.

[0025] (2) To prevent large areas of invalid holes caused by cloud cover in the extracted optical feature matrix, this embodiment of the invention further introduces cloud cover screening logic before acquiring multispectral optical satellite imagery and lidar point cloud data. An empirical threshold for optical cloud coverage can be set, which can be set to 5%. Within the target time period, a single date with cloud coverage lower than the 5% empirical threshold is selected and set as the snapshot date. The multispectral optical satellite imagery and lidar point cloud data under the snapshot date are the optical feature matrix data targeted by this embodiment of the invention. Similar to the rate matrix, these two dimensions of data can also be mapped to the grid network of the target area.

[0026] (3) Finally, in the grid network of the target area, each location is divided into individual grids, each containing synthetic aperture radar deformation rate, multispectral optical satellite identification results, and lidar point cloud data. If there is missing data, the data in the three dimensions can be resampled separately based on the spatial resampling algorithm and stored uniformly in the grid network to ensure that the elevation, vegetation, or deformation data of each location share an absolutely consistent spatial boundary range. The resampling algorithm can be either bilinear interpolation or nearest neighbor interpolation, which will not be elaborated or limited here.

[0027] Synthetic Aperture Radar (SAR) deformation rate characterizes the speed of surface movement along the radar line of sight, enabling the assessment of whether a location experiences uplift or subsidence within a target timeframe, and the magnitude of such uplift or subsidence. To effectively delineate potential hazard areas, this invention focuses on the potential for landform changes within the target region. A larger absolute value of the SAR deformation rate at a given location indicates significant topographic change, making that location more likely to be a source of geological hazards. Therefore, significant deformation locations can be identified based on the SAR deformation rates at various points within the target region. These significant deformation locations can then be used as the analysis boundary in subsequent maximum flow minimum cut algorithms.

[0028] LiDAR point clouds are a type of three-dimensional data that can intuitively represent the elevation data of a target area in two dimensions. Significant environmental changes caused by geological disasters typically occur at locations with steeper slopes; that is, the steeper the slope, the faster fluid hazards such as debris flows move at that location, and the more significant the disaster. Therefore, this embodiment of the invention first obtains the local slope matrix corresponding to the target area based on the elevation data in the LiDAR point cloud. Then, based on the distribution of local slopes in the local slope matrix, the water flow direction at each location is determined, meaning the water flow should move from the current location to a location with a steeper slope. Based on this water flow direction, a water flow path simulation can be performed in the local slope matrix. Finally, the endpoint of the simulated water flow direction can be used as the location of a gully. This gully location serves the opposite purpose to the location of significant deformation; it needs to be used as a reference for the water flow sink node (endpoint) in the subsequent maximum flow minimum cut algorithm.

[0029] Multispectral optical satellite imagery can obtain vegetation cover at various locations in a target area based on spectral data, such as the Normalized Difference Vegetation Index (NDVI) feature. The greater the surface vegetation cover at a location, the richer the vegetation is at that location. The more complex the root system of the plants, the stronger the geological stability, indicating that geological disasters may have a smaller impact at that location.

[0030] It should be noted that since a location cannot be both a disaster source and a disaster sink, after obtaining the set of significant deformation locations and the set of valley locations, it is necessary to take the intersection of the two sets and remove the significant deformation locations from the set of significant deformation locations.

[0031] Step S2: For each location, obtain a safety weight based on the local slope and surface vegetation coverage; treat each location as a node, with directed edges between adjacent nodes, and the edge weight is the safety weight pointing to the node, to obtain the topology map of the target area; execute the maximum flow minimum cut algorithm on the topology map to obtain the initial hazard boundary line; in the maximum flow minimum cut algorithm, set the edge weight between the virtual water flow source node and the significantly deformed location to a preset maximum value, and set the edge weight between the virtual water flow sink node and the valley location to a preset maximum value.

[0032] As described in step S1 above, the higher the surface vegetation coverage at a given location, the more stable the corresponding geology, the less impact disasters have on it, and the lower the probability of it being a potential hazard point. A steeper local slope indicates a more rugged location, making it more susceptible to significant geological changes under the influence of disasters. Therefore, this embodiment of the invention quantifies the safety weight for each location; a higher safety weight indicates a safer location, while a lower weight indicates a higher likelihood of it being a potential hazard point.

[0033] This invention further employs graph theory for global analysis. First, each location is considered a node in a graph structure. Directed edges exist between nodes and their neighbors, with the edge weight representing the safety weight pointing to the node. A directed edge represents a local path for water flow from one node to another, and the edge weight represents the outcome along this local path; a larger edge weight indicates a more stable geology and less potential for hazard. It should be noted that in this invention, the neighboring node refers to the other eight neighboring nodes within an eight-neighborhood (there may be fewer than eight at boundary locations).

[0034] After obtaining the topology map of the target region, the maximum flow minimum cut algorithm can be performed on it. In a specific implementation of this invention, the Boykov-Kolmogorov algorithm can be used. The goal of this algorithm is to find the edge with the smallest sum of edge weights. This edge is considered to be able to "cut off" the strong connections in the entire topology map. That is, this path is considered to be a path with weak security, which is prone to large geological changes when subjected to water flow disasters, thus creating hidden dangers. The basic idea of ​​this algorithm includes: 1. Water injection (path augmentation): Injecting an infinite flow from the virtual source point to all deformation points, allowing the "water" to spread outwards along the resistance (i.e., edge weights; the larger the edge weight, the greater the resistance to water flow); 2. Bottleneck finding: When the "water" attempts to flow to the sink point through any path, the algorithm records the narrowest and most congested section (i.e., the section with the smallest cumulative resistance) along the way; 3. Saturation cutoff: As the water injection pressure is continuously increased until the total flow can no longer be increased regardless of path adjustments, the edges that are completely filled with "water" and just blocking the path have reached a saturation state; 4. Output boundary: Extracting the coordinates of the target nodes connected by these saturated edges yields the boundary values ​​for the target nodes. Figure 1 The two closed loops or blocking lines are the initial hazard boundary lines.

[0035] During the execution of the maximum flow minimum cut algorithm, a set of virtual points is set up, namely virtual water flow source nodes and virtual water flow sink nodes. These virtual points do not correspond to specific locations in the target region; they serve only as logical placeholders at the algorithm's architecture level. Because the maximum flow minimum cut algorithm is a standard binary labeling mathematical tool, it mandates that the graph must have a definite starting point and a definite ending point. However, these definite starting and ending points cannot be directly selected from significant deformation locations and valley locations, as selecting any pair would ignore other source-sink results. Therefore, the algorithm treats virtual water flow source nodes as collective representatives of all significant deformation locations and virtual water flow sink nodes as collective representatives of all valley locations. Through this set of virtual nodes, the entire scene is transformed into a path result from one set of source nodes to another set of sink nodes. Since virtual points are virtual and do not correspond to any specific location, the maximum flow minimum cut algorithm sets the edge weights between virtual water flow source nodes and significantly deformed locations to a preset maximum value, and sets the edge weights between virtual water flow sink nodes and valley locations to a preset maximum value, thereby constructing physical boundaries and preventing the algorithm from accidentally outputting values ​​on these virtual connecting lines during judgment. In a specific implementation of this invention, the preset maximum value can be directly set to infinity.

[0036] Step S3: The initial hazard boundary line divides the target area into multiple sub-regions. Based on the elevation distribution of significant deformation locations in the sub-regions, the initial hazard boundary line is intercepted to obtain the final hazard boundary line.

[0037] Because graph theory algorithms alone can only find the shortest dividing line based on surface resistance, they cannot distinguish whether the enclosed area is mountain deformation or farmland tillage deformation in a long, gentle slope of a plain. Therefore, the initial hazard boundary line may contain artifacts of non-hazard deformation in the plain caused by pumping or farmland activities, and these artifacts should be removed. The obvious characteristic of these artifacts is that they generate a small elevation gradient within a region, and the entire region is relatively flat. This means that although there is internal loosening, the flat terrain prevents further downward movement and makes landslides unlikely. Conversely, if the sub-region is a significant hazard area, the water flow path generated by the significant deformation location within the sub-region will have a high elevation gradient, indicating a large terrain drop. Gravity is sufficient to overcome the soil's shear strength, making it a genuine hazard area. Therefore, this embodiment of the invention truncates the initial hazard boundary line based on the elevation distribution of significant deformation locations within the sub-region, removing the sub-region boundaries that are artifacts to obtain the final hazard boundary line.

[0038] Step S4: Expand the final hazard boundary line to obtain the hazard area; identify targets in the hazard area to obtain the hazard point identification results.

[0039] Because the final hazard boundary line obtained is only a set of locations in the target area, it cannot represent a region. Therefore, it is necessary to expand the region with the final hazard boundary line as the center, and finally obtain the hazard area and perform target identification to obtain the hazard point identification result.

[0040] In summary, this invention acquires synthetic aperture radar deformation rate, multispectral optical satellite imagery, and lidar point cloud data of the target area to extract significant deformation locations, surface vegetation coverage, and local slope matrices. Based on the local slope distribution, the direction of water flow and the location of its endpoint valley are determined. Furthermore, safety weights are calculated based on the local slope and vegetation coverage at each location. A topology graph is constructed using grid nodes and directed edges, where edge weights are the safety weights pointing to nodes. A maximum flow minimum cut algorithm is applied to the topology graph to obtain initial hazard boundaries. Subsequently, the initial hazard boundaries are truncated and verified based on the elevation distribution of significant deformation locations within each sub-region, eliminating plain subsidence artifacts to obtain the final hazard boundaries. After regional expansion, target identification is performed, and the hazard point identification results are output. This invention achieves adaptive scientific demarcation of the hazard impact range, significantly reducing the false alarm rate in complex terrain scenarios.

[0041] Preferably, in some implementations of the present invention, considering that radar equipment inevitably introduces system thermal noise speckle when measuring millimeter-level minute deformations, directly using the location generating the synthetic aperture radar deformation rate as the significant deformation location without intervention would misinterpret the instrument noise as actual ground subsidence. Therefore, this implementation limits the selection of significant deformation locations by setting a threshold. If the absolute value of the synthetic aperture radar deformation rate at a certain location is greater than a preset comparison threshold, then the corresponding location is designated as the significant deformation location. This comparison threshold is the equipment noise reference value calibrated at the factory of the synthetic aperture radar equipment; it is a priori value that does not need to be obtained and can be directly retrieved.

[0042] Preferably, in some implementations of the embodiments of the present invention, considering that vegetation indices directly calculated from optical images often contain distorted extreme values ​​caused by high water absorption rates, affecting subsequent data standardization processing, the embodiments of the present invention further utilize histogram statistical information to truncate and remove extreme values, and normalize the distribution of vegetation growth status, specifically including: The normalized vegetation index (NDI) is obtained for each location, and a histogram of the NDI is plotted. Baseline values ​​for bare ground distribution and vegetation cover distribution are selected from the histogram. In one specific implementation of this invention, vegetation index values ​​corresponding to locations with a cumulative frequency of 5% are extracted from the histogram and set as the baseline value for bare ground distribution; vegetation index values ​​corresponding to locations with a cumulative frequency of 95% are extracted and set as the baseline value for vegetation cover distribution.

[0043] The surface vegetation cover is obtained by linearly normalizing the normalized vegetation index (NZDI) using the baseline values ​​for bare surface distribution and vegetation cover distribution as the data basis. Specifically, if the NZDI is less than the bare surface distribution baseline value, it is set to 0; if the NZDI is greater than the vegetation cover baseline value, it is set to 1; for values ​​between the bare surface distribution baseline value and the vegetation cover baseline value, a range standardization formula is used to normalize the data, taking the bare surface distribution baseline value as the minimum and the vegetation cover baseline value as the maximum, to obtain the surface vegetation cover.

[0044] It should be noted that the normalized vegetation index is a well-known technical method in the field of vegetation, which can be calculated based on the near-infrared band reflectance and the red band reflectance through the normalized difference formula, and will not be elaborated here.

[0045] Preferably, in some implementations of the embodiments of the present invention, the method for constructing the local slope matrix includes: For any target location, neighborhood locations within a preset neighborhood range are obtained, centered on the target location. Neighboring locations with elevation values ​​lower than the target location's elevation are designated as low-elevation locations, and the low-elevation location with the smallest elevation value is designated as the comparison location; that is, the elevation difference between the target location and the comparison location is the largest. The local slope of the target location is obtained by using the elevation difference between the target location and the comparison location as the numerator and the horizontal distance between them as the denominator. It should be noted that, in this embodiment of the invention, the neighborhood locations refer to the target location within an eight-neighborhood range.

[0046] If no corresponding low-elevation location exists for the target location, the local slope is set to 0.

[0047] The local slope matrix is ​​obtained by statistically analyzing the local slopes at all locations.

[0048] Furthermore, in one implementation of this invention, considering that the original point cloud data includes non-ground reflection points such as artificial buildings and tree canopies, and that equipment measurement errors can cause false grid depressions on the ground surface, which can disrupt continuous calculations of water flow paths, it is necessary to filter the original point cloud data to separate and remove non-ground points, retaining only pure ground point cloud data. Therefore, this embodiment of the invention performs spatial interpolation calculations on the pure ground point cloud data to generate a preliminary digital elevation model matrix. Then, it performs depression-filling calculations on this preliminary digital elevation model matrix to modify the elevation values ​​of pseudo-low points surrounded by local minima. The system outputs a smoothed digital elevation model matrix after filling. This filtering operation is a well-known cloth simulation filtering algorithm in the art, and is a known technology; the specific principles and algorithms will not be elaborated further.

[0049] Furthermore, in some implementations of this invention, since the comparison location is the point with the largest elevation difference within the neighborhood of the target location, water flow at the target location will significantly flow towards the comparison location. Therefore, the method for determining the valley location includes: The direction from the target location to the comparison location is taken as the water flow direction at the target location, and the water flow direction field of the target area is obtained. In the water flow direction field, start from any position and trace along the water flow direction until the end of the path has no water flow direction. The end of the path is then taken as the location of the gully.

[0050] It should be noted that the absence of a water flow direction means that there are no low-lying locations in the neighborhood of the target location, and therefore no further flow direction will be generated. Alternatively, if a valley is located at the boundary of the current detection field of view, it can also be directly regarded as a sink node.

[0051] Preferably, in this embodiment of the invention, the method for obtaining the security weight includes: For any given location, the negative correlation mapping and normalization result of the local slope is weighted and summed with the normalized result of the surface vegetation cover according to preset weights to obtain the initial safety weight.

[0052] In one specific implementation of this invention, the initial security weight is expressed by the formula:

[0053] in, Let (x, y) be the initial safety weights. As the first weight, The second weight is given by e, where e is the natural constant. This is the slope attenuation coefficient. Let be the local slope at position (x, y). Let be the vegetation cover at location (x, y).

[0054] in The negative correlation mapping result for local slope is achieved and normalized using an exponential function with a decay coefficient. This slope decay coefficient can be set to 0.05. Through this exponential decay operation, when the local slope is too large (i.e., the ground surface is steep), the calculated result of this term will drastically approach zero, thus significantly reducing the total weight value of the grid. Both the first and second weights can be set to 0.5, meaning that the feature confidence levels of the two dimensions are the same. These two weights serve as balancing empirical values ​​to ensure that the contribution of the slope influence exponential term and the vegetation influence linear term to the final weight output result is on the same order of magnitude, preventing a single feature from dominating the calculation result.

[0055] To further avoid the error of the safety weight being calculated as 0 during the optimization process in the subsequent maximum flow minimum cut algorithm, the sum of the initial safety weight and the preset minimum weight value is used as the safety weight. In this embodiment of the invention, the minimum weight value can be set to 0.001.

[0056] Preferably, in some implementations of the embodiments of the present invention, the method for obtaining the final hazard boundary line includes: For any significant deformation location within a sub-region, the elevation value corresponding to the significant deformation location is the starting elevation value. Based on the water flow direction, a water flow path is obtained starting from the significant deformation location. The elevation value corresponding to the end point of the path is taken as the ending elevation value. The difference between the starting and ending elevation values ​​is used as the numerator, and the horizontal distance between the significant deformation location and the end point of the path is used as the denominator to obtain the average slope gradient at the significant deformation location. That is, the greater the change in elevation difference along a shorter path, the greater the slope gradient, which represents a real change in disaster risk.

[0057] It should be noted that the horizontal distance can be obtained by multiplying the length of the water flow path (the number of grids it contains) by a preset conversion factor for the side length of the reference grid (for example, if the grid resolution is 2 meters, the factor is 2). Since the valley location is the endpoint of the water flow direction, the so-called endpoint of the water flow path obtained above is actually the valley location.

[0058] If the average gradient of all significant deformation locations within a sub-region is less than a preset gradient threshold, then the initial hazard boundary line corresponding to the sub-region boundary is removed; all sub-regions are traversed to obtain the final hazard boundary line. In one implementation of this invention, the gradient threshold can be pre-calibrated based on experience, typically as the upper limit of the gradient of a plain or gentle slope.

[0059] It should be noted that the above process is only performed on sub-regions with significant deformation locations. Sub-regions without significant deformation locations do not need to be analyzed, as the boundaries of these sub-regions may be the boundaries of the sub-regions that need to be analyzed, thus requiring no additional analysis. Finally, through the removal operation, false regions with terrain undulation features are significantly filtered out.

[0060] Preferably, in some implementations of the embodiments of the present invention, the method for obtaining the hidden danger area includes: For each location on the final hazard boundary line, the product of the local slope at that location and the preset conversion coefficient is used as the radius adjustment amount. That is, the larger the local slope at a location, the more vulnerable the corresponding location is. In order to obtain hazard points more accurately, a larger area is needed to avoid missed detection, so the calculated radius adjustment amount is larger.

[0061] In one specific implementation of this invention, the conversion coefficient is used to convert the dimensionless local slope to a side length unit. The final product result may be a floating-point number with a decimal point, so the radius adjustment amount can be obtained by rounding up.

[0062] The sum of the radius adjustment and the initial expansion radius is taken as the final expansion radius; a circular region is constructed with the corresponding position as the center and the final expansion radius as the radius; the union of the circular regions formed at each position on the final hazard boundary line is taken as the hazard region. The initial expansion radius is a necessary, artificially reserved environmental texture context space around the surface features to prevent the feature extraction module from failing to capture edge shape changes within an extremely narrow single-pixel range. In one specific implementation of this invention, it is set to 50 pixel units.

[0063] Preferably, in some implementations of the present invention, considering that existing visual masking methods typically force the background pixels outside the mask to be pure black, this one-size-fits-all approach artificially creates a large number of sharp, high-frequency black-and-white boundary edges on the image. When the convolutional kernel of the target detection network extracts these areas, it may mistake the artificially cut edges for physical textures of the ground surface, and due to the lack of continuous ground background, the network loses the contextual receptive field necessary for determining target attributes. Therefore, this implementation performs target recognition in hazard areas, including: Extract the three-channel color base map matrix corresponding to the location of the potential hazard area in the multispectral optical satellite imagery. Specifically, within the image field of view of the target area, the potential hazard area is set to 1, and other areas are set to 0, thus obtaining a mask. Extract the red, green, and blue band reflectance matrices from the multispectral optical satellite imagery. Combine these three single-band matrices in the depth dimension to obtain a global three-channel color base map. The three-channel color base map matrix can then be obtained using the mask.

[0064] During the extraction process, a smooth dimensionality reduction method is used at the edges of the potential hazard areas. The smooth dimensionality reduction method can select Gaussian blur weakening processing to reduce the sharpness of the boundaries, thereby achieving the goal of blocking artificial textures (such as distant highways and farmland boundaries) in large irrelevant areas from being input into the network, while ensuring that the pixels of the geological deformation areas and their surrounding environment within the retained blocks present a smooth and continuous spatial transition.

[0065] Target recognition is performed on the three-channel color background map matrix using a target detection network. In this embodiment, the target detection network can employ a standard recognition model within the YOLO architecture. This network is pre-trained. Because the three-channel color background map matrix retains continuous contextual pixels surrounding the target object, the network's convolutional kernels can effectively combine and calculate related environmental features such as soil faults and vegetation damage levels around the potential hazard point, rather than searching for edges in isolation against a stark black background. Finally, a target bounding box is output, containing the box's location coordinates and a confidence score indicating the probability of the presence of actual geological cracks or collapses within the box.

[0066] Preferably, considering that a long, real geological fault zone may trigger multiple nested or closely spaced repeated prediction boxes in image processing, it is necessary to determine whether each pair of the obtained bounding boxes should belong to the same recognition result. Specifically, this includes: If the target detection network generates multiple bounding boxes, the cross-union ratio (CUI) between these bounding boxes is obtained. Bounding boxes with a CUI greater than a preset threshold are merged. In other words, two bounding boxes with a CUI greater than the preset threshold are considered to belong to the same geological hazard feature, resulting in redundancy.

[0067] The merging process is complete when the intersection-union ratio (IU) of all the target bounding boxes is less than or equal to the IU threshold. All the merged target bounding boxes are then used as the hazard point identification results. Based on the mapping relationship between the image pixel coordinate system and the real geographic grid, these target bounding boxes can be converted into latitude and longitude coordinate boundaries in the real surface space. The final calculated spatial coordinate boundary boxes of the geological hazard points are output and used to feed back to the monitoring terminal or generate investigation and early warning work orders.

[0068] It should be noted that the intersection-union ratio is a technical feature well known to those skilled in the art, and the specific method for obtaining it will not be elaborated here. The threshold for the intersection-union ratio can be set to 0.5.

[0069] It should be noted that the above merging operation can also be replaced by a deletion operation, that is, comparing the confidence scores of the two target bounding boxes and comparing the scores. The system retains the target bounding box with the higher confidence score and deletes the redundant result with the lower confidence score.

Claims

1. A deep learning-based intelligent identification method for sudden geological disaster hidden points, characterized in that, The method includes: Based on the synthetic aperture radar deformation rate at various locations in the target area, locations with significant deformation are selected; the local slope matrix corresponding to the target area is obtained from the lidar point cloud; the water flow direction at each location is determined based on the distribution of local slope in the local slope matrix, and the endpoint of the water flow direction is taken as the location of the gully; the surface vegetation cover at each location is obtained using multispectral optical satellite imagery. For each location, a safety weight is obtained based on the local slope and surface vegetation cover. Each location is considered as a node, and there are directed edges between adjacent nodes. The edge weight is the safety weight pointing to the node, thus obtaining a topology map of the target area. The maximum flow minimum cut algorithm is executed on the topology map to obtain the initial hazard boundary line. In the maximum flow minimum cut algorithm, the edge weight between the virtual water flow source node and the significantly deformed location is set to a preset maximum value, and the edge weight between the virtual water flow sink node and the valley location is set to a preset maximum value. The initial hazard boundary line divides the target area into multiple sub-regions. Based on the elevation distribution of significant deformation locations in the sub-regions, the initial hazard boundary line is intercepted to obtain the final hazard boundary line. The final hazard boundary line is expanded to obtain the hazard area; targets are identified within the hazard area to obtain the hazard point identification results.

2. The method according to claim 1, wherein, The method for constructing the local slope matrix includes: For any target location, obtain the neighborhood locations within a preset neighborhood range centered on the target location. The neighborhood locations with elevation values ​​lower than the target location's elevation value are designated as low-elevation locations, and the low-elevation location with the minimum elevation value is designated as the comparison location. The local slope of the target location is obtained by using the elevation difference between the target location and the comparison location as the numerator and the horizontal distance between the target location and the comparison location as the denominator. If no corresponding low-elevation location exists for the target location, the local slope is set to 0. The local slope matrix is ​​obtained by statistically analyzing the local slopes at all locations.

3. The method according to claim 2, wherein, The method for determining the location of the valley includes: The direction from the target location to the comparison location is taken as the water flow direction at the target location, and the water flow direction field of the target area is obtained. In the water flow direction field, start from any position and trace along the water flow direction until the end of the path has no water flow direction. The end of the path is then taken as the location of the gully.

4. The method according to claim 1, wherein, The method for obtaining the security weight includes: For any given location, the result of mapping and normalizing the negative correlation of the local slope is weighted and summed with the normalized result of the surface vegetation cover according to a preset weight to obtain the initial safety weight; the sum of the initial safety weight and the minimum value of the preset weight is taken as the safety weight.

5. The deep learning-based intelligent identification method for sudden geological disaster hidden points according to claim 1, characterized in that, The method for obtaining the final hazard boundary line includes: For any significant deformation location within a sub-region, the elevation value corresponding to the significant deformation location is the starting elevation value; the water flow path starting from the significant deformation location is obtained according to the water flow direction, the elevation value corresponding to the end of the path is taken as the ending elevation value, the difference between the starting elevation value and the ending elevation value is taken as the numerator, and the horizontal distance between the significant deformation location and the end of the path is taken as the denominator to obtain the average slope gradient of the significant deformation location; If the average gradient of all significant deformation locations within a sub-region is less than a preset gradient threshold, then the initial hazard boundary line corresponding to the sub-region boundary is removed; traverse all sub-regions to obtain the final hazard boundary line.

6. The deep learning-based intelligent identification method for sudden geological disaster hidden points according to claim 1, characterized in that, The method for obtaining the potential hazard area includes: For each location on the final hazard boundary line, the product of the local slope at the corresponding location and the preset conversion coefficient is used as the radius adjustment amount, and the sum of the radius adjustment amount and the initial expansion radius is used as the final expansion radius; a circular area is constructed with the corresponding location as the center and the final expansion radius as the radius; the union of the circular areas formed by each location on the final hazard boundary line is used as the hazard area.

7. The deep learning-based intelligent identification method for sudden geological disaster hidden points according to claim 1, characterized in that, The target identification of the potential hazard area includes: Extract the three-channel color base map matrix of the corresponding location of the potential hazard area in the multispectral optical satellite image; during the extraction process, a smooth dimensionality reduction method is used at the edge of the potential hazard area; The target detection network is used to identify targets in the three-channel color base map matrix.

8. The method according to claim 7, wherein, The results of obtaining the hazard identification include: If the target detection network generates multiple target bounding boxes, the cross-union ratio (CUI) between the target bounding boxes is obtained. Target bounding boxes with a CUI greater than a preset CUI threshold are merged. All target bounding boxes obtained after merging are used as the hazard point identification results.

9. The deep learning-based intelligent identification method for sudden geological disaster hidden points according to claim 1, characterized in that, The selection of locations with significant deformation includes: If the absolute value of the synthetic aperture radar deformation rate at a certain location is greater than a preset comparison threshold, then the corresponding location is designated as the significant deformation location.

10. The intelligent identification method for sudden geological disaster hazard points based on deep learning according to claim 1, characterized in that, The method for obtaining the surface vegetation coverage includes: Obtain the normalized vegetation index for each location, and plot a histogram of the normalized vegetation index. Select the baseline values ​​for bare surface distribution and vegetation cover distribution on the histogram. Use the baseline values ​​for bare surface distribution and vegetation cover distribution as the data basis to linearly normalize the normalized vegetation index to obtain the surface vegetation cover.

Images