Method and system for intelligent identification and classification of geological disaster hidden points by unmanned aerial vehicle surveying

Abstract

The present application relates to the technical field of geological disaster monitoring, and discloses a method and system for unmanned aerial vehicle photogrammetric intelligent identification and grading of geological disaster hidden danger points, which converts pixel-level identification results into hierarchical data retaining spatial topological structure by constructing a confidence four-ary tree, realizes coupling modeling of observation conditions and confidence uncertainty based on a weighted decision grid established based on photogrammetric parameters, calculates a topological difference degree to generate a spatial difference field to quantify the deviation degree of the confidence field from a standard grading decision surface, and finally executes adaptive grading determination to output a risk level result with a confidence label; the present application solves the problem of topological mismatch between the confidence field and the decision boundary space caused by the truncation of confidence continuous measurement information in the grading stage in the traditional serial architecture, avoids the systematic underestimation of the misjudgment risk in the observation condition limited area such as high mountain and canyon area, and realizes the consistency between the actual confidence of the risk quantization result and the nominal level.

Description

Methods and systems for intelligent identification and classification of geological hazard sites using UAV aerial surveys Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and in particular to a method and system for intelligent identification and classification of geological disaster hazard points by UAV aerial survey. Background Technology

[0002] Geological hazard identification and classification technology primarily utilizes UAV aerial surveys to acquire surface image data. Deep learning models, such as convolutional neural networks, are then used to perform pixel-level or object-level identification, outputting the probability distribution of hazard areas. Risk levels are then classified based on preset threshold indicators such as deformation, slope, and area. This technology enables large-scale, rapid investigation and features automated processing and high operational efficiency, making it widely used in geological hazard surveys.

[0003] In engineering practice, technicians conducting UAV aerial surveys in high mountain and canyon areas frequently encountered persistent discrepancies between the hazard level assessment results and on-site verification. Specifically, hazard points at the bottom of the canyon were classified as low-risk in the grading results, but on-site investigations revealed clear signs of surface deformation, warranting a higher risk level. Conversely, some areas at the mountain top were classified as high-risk, yet on-site verification revealed no obvious deformation. This discrepancy between assessment results and reality repeatedly occurred in areas with limited observation conditions, leading to a mismatch between the reliability of the risk quantification results and the nominal level. The root cause of this problem lies in the current technology's serial processing architecture of "identification first, grading later." The pixel-level or object-level confidence scores generated by the identification model are only used in the binary thresholding stage. The subsequent grading model receives discretized identification results, completely truncating the continuous measurement information of the original confidence scores during the grading phase. This information chain break prevents the grading decision function from perceiving spatial reliability differences in the identification results, treating the identification outputs of high-confidence and low-confidence areas as homogeneous inputs. The root of the problem lies in the fact that the decision boundary of the grading standard is a fixed hyperplane in the feature space, while the actual identification confidence field exhibits a non-uniform distribution in space due to factors such as fluctuations in the basal-to-elevation ratio of UAV aerial surveys, uneven image overlap, and differences in terrain texture richness. The lack of an explicit mapping relationship between the two in their topological structure leads to a spatial misalignment between the high-gradient region of the confidence field and the boundary of the grading decision. In engineering practice, technicians typically attribute this error to insufficient accuracy of the identification model itself, failing to recognize that it stems from the loss of information about the spatial structure of the confidence level by the grading decision layer. This problem is particularly prominent in areas with limited observation conditions, such as high mountains and canyons, where the variance of the posterior probability distribution of hazard identification results increases significantly. However, the grading still uses a uniform standard, causing the actual credibility of the risk quantification results to differ from the nominal level. Furthermore, this mismatch intensifies non-linearly with terrain complexity, and conventional statistical calibration methods cannot correct it. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a method for intelligent identification and classification of geological hazard hazard points using UAV aerial surveying, including:

[0005] Step S1: Acquire UAV aerial survey image data, generate pixel-level hazard identification confidence field and construct confidence quadtree;

[0006] Step S2: Based on the confidence quadtree and aerial survey parameters, establish a partial order relationship and generate a weighted decision lattice;

[0007] Step S3: Based on the weighted decision lattice and the standard hierarchical decision surface, calculate the topological dissimilarity degree and generate the spatial dissimilarity field;

[0008] Step S4: Based on the spatial difference field and weighted decision lattice, perform adaptive hierarchical judgment and output the evaluation results.

[0009] Furthermore, the process of generating a pixel-level hazard identification confidence field and constructing a confidence quadtree in step S1 includes:

[0010] Acquire multi-view images from UAV aerial surveys, perform aerial triangulation and orthorectification operations, and generate standardized digital orthorectified image maps;

[0011] Based on standardized digital orthophotos, an edge-enhanced convolutional neural network is invoked to perform pixel-level hazard identification operations and generate a pixel-level hazard identification confidence field.

[0012] Based on the pixel-level hazard identification confidence field, a quadtree construction operation is performed according to the geospatial recursive partitioning rule to generate a confidence quadtree. The geospatial recursive partitioning rule refers to dividing the rectangular geographical area of ​​the current node into four equal sub-rectangular areas according to the principle of spatial four-eighths division.

[0013] Furthermore, the edge-enhancing convolutional neural network processing in step S12 includes:

[0014] The standardized digital orthophoto image is input into the first layer of the network, and basic visual features are extracted through standard convolutional kernels.

[0015] In the second layer of the network, the Laplacian edge operator is introduced and multiplied element-wise with the feature map of the first layer to enhance the feature response of the boundary region of the potential danger.

[0016] Expanding the receptive field through dilated convolution in the third layer of the network to capture information about surface deformation areas;

[0017] By using skip connections in the fourth layer of the network, the features from the third layer are concatenated with those from the first layer, preserving multi-scale contextual information.

[0018] The fifth layer of the network uses 1×1 convolution to reduce the dimensionality of the spliced ​​features. After inputting the classification head, it outputs a dual-channel probability map of the hazard category and the non-hazard category. The hazard category channel is taken as the pixel-level hazard recognition confidence field.

[0019] Furthermore, the process of establishing partial order relations and generating weighted decision lattices in step S2 includes:

[0020] Receive the confidence quadtree, extract node information and obtain UAV aerial survey technical parameters, generate node basic dataset and normalized aerial survey parameters; UAV aerial survey technical parameters include base-to-height ratio, image overlap and terrain texture complexity. Normalization processing refers to subtracting the minimum value from the original parameter value and then dividing by the difference between the maximum value and the minimum value, so that the parameter is linearly mapped to the interval of 0 to 1.

[0021] Based on the node-based dataset and normalized aerial survey parameters, priority relationships are defined and transitive closure operations are performed to generate a partial order relation matrix.

[0022] Based on the partial order relation matrix, a decision lattice is constructed and the comprehensive weight of the nodes is calculated to generate a weighted decision lattice. The comprehensive weight refers to the variance of the confidence level of the corresponding quadtree node multiplied by the depth normalization factor, and then multiplied by the geometric mean of three normalized parameters: the inverse of the base-height ratio, the image overlap complement, and the terrain texture complexity entropy.

[0023] Furthermore, the process of defining precedence relations and performing transitive closure operations in step S22 includes:

[0024] Define a priority relationship. For any two nodes, determine the spatial inclusion relationship and the observation condition weight inferiority relationship. When both the spatial inclusion relationship and the observation condition weight inferiority relationship are satisfied, establish a priority connection. The spatial inclusion relationship means that the geographic rectangle of node A completely contains the geographic rectangle of node B. The observation condition weight inferiority relationship means that the observation condition weight inferiority index is obtained by calculating the weighted sum of the inverse of the basal height ratio of the node, the image overlap complement, and the terrain texture complexity entropy. If the observation condition weight inferiority index of node A is greater than that of node B, then the observation condition weight of node A is determined to be inferior to that of node B.

[0025] Initialize the priority relation matrix, setting the matrix elements corresponding to the positions where priority connections are established to 1;

[0026] Perform transitive closure operation on the priority relation matrix and iterate until no new priority connections are generated to obtain the partial order relation matrix.

[0027] Furthermore, the process of calculating the topological dissimilarity degree and generating the spatial dissimilarity field in step S3 includes:

[0028] The quantitative indicator thresholds in the geological hazard hazard classification standard are obtained, and a standard hazard classification decision surface is constructed. The quantitative indicator thresholds include deformation threshold, hazard area threshold and slope threshold. The standard hazard classification decision surface refers to the combination of the boundary hyperplane formed by the deformation threshold, hazard area threshold and slope threshold in the three-dimensional feature space composed of deformation axis, hazard area axis and slope axis.

[0029] Based on the weighted decision lattice and the standard hazard classification decision surface, the topological difference index of each node is calculated to obtain the node topological difference set.

[0030] Based on the set of node topological dissimilarity, a bilinear interpolation operation is performed to generate a spatial dissimilarity field. The bilinear interpolation operation refers to the operation of using the dissimilarity values ​​of four adjacent grid nodes and averaging them according to distance weights to calculate the field value of non-node locations.

[0031] Furthermore, the process of calculating the topology dissimilarity index for each node in step S32 includes:

[0032] Traverse all nodes of the weighted decision lattice, extract the average confidence score and confidence score interval, and map the average confidence score to coordinate points in the three-dimensional feature space;

[0033] Calculate the vertical distance from the coordinate point to the standard hazard classification decision surface, and use it as the spatial distance component;

[0034] Extract the total weight of the node and the sum of the total weights of its child nodes as the basic weight components. Combine this with the confidence interval width to calculate the weight adjustment coefficient and obtain the final weight components.

[0035] The topological dissimilarity index is calculated by multiplying the spatial distance component by the weight component, and then by the depth normalization factor, and then integrating them to form a set of node topological dissimilarity.

[0036] Furthermore, the process of performing adaptive classification determination and outputting evaluation results in step S4 includes:

[0037] The system receives the spatial difference field, sets the difference tolerance threshold, and identifies areas that require level adjustment. The difference tolerance threshold is obtained by calculating the standard deviation of all grid cell values ​​in the spatial difference field and then multiplying it by the balance coefficient.

[0038] Based on the adjustment of regional and spatial difference fields and weighted decision grids, a level correction operation is performed to obtain a corrected risk level distribution map.

[0039] Based on the revised risk level distribution map, spatial difference field, and weighted decision lattice, the posterior confidence assessment is calculated and the final result is output.

[0040] Furthermore, the process of performing the level correction operation in step S42 includes:

[0041] For each geographic unit that needs adjustment, find the corresponding minimum coverage node, and obtain the node's comprehensive weight value, the sum of the comprehensive weights of its child nodes, and the original risk level.

[0042] Calculate the median of the combined weights of all nodes in the weighted decision lattice and the global mean of the sum of the combined weights of the child nodes;

[0043] The grade is adjusted based on the node's overall weight, the sum of the child node's overall weight, and the spatial difference field value. The adjustment rules are as follows: if the node's overall weight is greater than the median weight or the sum of the child node's overall weight is greater than the global mean, and the spatial difference field value is positive, the original grade is increased by one level; if the node's overall weight is greater than the median weight or the sum of the child node's overall weight is greater than the global mean, and the spatial difference field value is negative, the original grade is decreased by one level; otherwise, the original grade remains unchanged.

[0044] Step S424: Perform a local smoothing filter operation on the corrected risk level distribution map to obtain the corrected risk level distribution map.

[0045] A system for intelligent identification and classification of geological hazard hazard points using UAV aerial surveys, comprising the aforementioned method for intelligent identification and classification of geological hazard hazard points using UAV aerial surveys, the system including:

[0046] Image processing and confidence quadtree construction module: used to acquire UAV aerial survey image data, generate pixel-level hazard identification confidence field and construct confidence quadtree;

[0047] Partial order relation establishment and weighted decision lattice generation module: used to establish partial order relations and generate weighted decision lattices based on confidence quadtrees and aerial survey parameters;

[0048] Topological Dissimilarity Calculation and Spatial Dissimilarity Field Generation Module: Used to calculate topological dissimilarity and generate spatial dissimilarity fields based on weighted decision lattices and standard hierarchical decision surfaces;

[0049] Adaptive grading judgment and evaluation result output module: Based on the spatial difference field and weighted decision lattice, it performs adaptive grading judgment and outputs the final risk level and credibility evaluation results.

[0050] Compared to existing technologies, the advantages of this invention are as follows: By constructing a two-layer structure of a confidence quadtree and a weighted decision lattice, this invention solves the technical defect that continuous confidence measurement information is truncated during the grading stage in a serial architecture. This method transforms the pixel-level confidence field into a tree structure that preserves spatial aggregation topological relationships, and introduces parameters such as aerial survey base-height ratio, image overlap, and terrain texture complexity to establish partial order relationships between nodes. This enables the grading decision function to perceive spatial reliability differences in the identification results, achieving coupled modeling of observation conditions and confidence uncertainty. In areas with limited observation conditions, such as high mountain and canyon areas, this invention can locate regions where confidence variance abnormally increases due to insufficient image overlap. By calculating topological difference, it identifies the structural misalignment between areas with drastic changes in confidence gradients and the grading decision boundary, and then performs adaptive grade correction. This mechanism avoids the phenomenon where areas with clear deformation signs at the bottom of canyons are systematically underestimated as low-risk failures due to confidence fluctuations, ensuring that the actual credibility of the risk quantification results matches the nominal grade. By using a spatial difference field-driven hierarchical adjustment, this invention transforms the traditional fixed hyperplane decision boundary into a lattice structure that can dynamically expand and contract with observation conditions. This preserves the spatial structure information of the original confidence level at the decision-making level, fundamentally eliminating the risk of misjudgment caused by the loss of spatial structure information of the confidence level at the hierarchical decision-making level, and achieving consistency between the geological hazard hazard classification results and the actual risk conditions on site. Attached Figure Description

[0051] To more clearly illustrate the technical solutions 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.

[0052] Figure 1 is a flowchart of the method for intelligent identification and classification of geological disaster hazard points by UAV aerial survey in this invention;

[0053] Figure 2 is a schematic diagram of the confidence field for pixel-level hazard identification in an embodiment of the present invention;

[0054] Figure 3 is a schematic diagram of the confidence quadtree structure in an embodiment of the present invention;

[0055] Figure 4 is a schematic diagram of the weighted decision lattice structure and its partial order relationship in an embodiment of the present invention;

[0056] Figure 5 is a system functional module diagram of the intelligent identification and classification of geological disaster hazard points by UAV aerial survey in this invention. Detailed Implementation

[0057] 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.

[0058] Example 1:

[0059] Please refer to Figure 1. This embodiment provides a method for intelligent identification and classification of geological hazard hazard points using UAV aerial surveying, including:

[0060] Step S1: Acquire UAV aerial survey image data, generate pixel-level hazard identification confidence field and construct confidence quadtree.

[0061] This step addresses the need for spatial confidence level structuring in geological hazard identification. First, it acquires UAV aerial survey imagery data. Through a series of preprocessing steps and model inference, it generates a pixel-level hazard identification confidence field, ultimately constructing a confidence quadtree. This confidence quadtree is a hierarchical tree structure formed by recursively dividing the pixel-level hazard identification confidence field according to geographic space. Each node stores the confidence statistical features, spatial adjacency relationships, and hierarchical depth for the corresponding spatial range. This transforms the confidence search in planar space into a hierarchical traversal operation of the tree structure, enabling rapid retrieval of confidence distribution patterns and local fluctuation characteristics within any spatial range. The UAV aerial survey imagery data consists of multi-view images from UAV aerial surveys.

[0062] Specifically, the steps for generating a pixel-level hazard identification confidence field and constructing a confidence quadtree are as follows:

[0063] Step S11: Acquire multi-view images from UAV aerial surveys, perform aerial triangulation and orthorectification operations, and generate standardized digital orthophoto maps.

[0064] The aerial triangulation densification process calculates the exterior orientation elements of the image using aerial triangulation, achieving high-precision spatial calibration of multi-view images. The orthorectification process utilizes a digital elevation model to eliminate geometric distortions caused by terrain undulations, generating a digital orthorectified image with a uniform scale. Specifically, to obtain standardized, high-precision image data, the aerial triangulation densification and orthorectification operations follow these steps:

[0065] Step S111: Collect multi-view images obtained by UAV aerial survey, and extract the initial values ​​of the interior orientation elements (determined by camera calibration parameters) and exterior orientation elements (provided by UAV flight control logs) of the multi-view images.

[0066] Step S112: Perform aerial triangulation encryption, extract corresponding points between multi-view images through image matching algorithm, calculate accurate exterior orientation elements based on bundle adjustment principle, and complete spatial registration of multi-view images.

[0067] Step S113: Load the digital elevation model of the corresponding area, perform orthorectification on the registered multi-view images to eliminate geometric distortion caused by terrain undulations, and generate a standardized digital orthophoto map.

[0068] Step S12: Based on the standardized digital orthophoto image generated in step S11, call the edge-enhanced convolutional neural network to perform pixel-level hazard identification operation and generate a pixel-level hazard identification confidence field.

[0069] The edge-enhanced convolutional neural network is a network structure that introduces a Laplacian edge detection operator into the encoder part and performs point-by-point multiplication with the convolutional feature map to enhance the feature response intensity of the hazard boundary region. The pixel-level hazard identification confidence field is a set of soft maximum probability values ​​for each pixel output by the network, indicating whether it belongs to the geological hazard category. The soft maximum probability values ​​are obtained by normalization using the softmax function and range from 0 to 1. Specifically, to generate a confidence field with clear boundaries and semantic continuity, the processing flow of the edge-enhanced convolutional neural network is as follows:

[0070] Step S121: Input the standardized digital orthophoto image into the first layer of the network and extract basic visual features through standard convolutional kernels.

[0071] Step S122: In the second layer of the network, the Laplacian edge operator is introduced and multiplied element-wise with the feature map of the first layer to enhance the feature response of the boundary region of the potential danger.

[0072] Step S123: Expand the receptive field and capture information about the surface deformation area by using a dilated convolution with a dilation rate of 2 in the third layer of the network without increasing the number of parameters.

[0073] Step S124: By using the skip connection of the fourth layer of the network, the features of the third layer are concatenated with the features of the first layer, preserving multi-scale contextual information.

[0074] Step S125: The fifth layer of the network uses a 1×1 convolution to reduce the dimensionality of the stitched features. After inputting the classification head, it outputs a dual-channel probability map of the hazard category and the non-hazard category. The hazard category channel is taken as the pixel-level hazard identification confidence field. Figure 2 is a schematic diagram of the pixel-level hazard identification confidence field. As shown in Figure 2, on the standardized digital orthophoto map (the figure uses a light green base to simulate terrain texture), after processing by the edge enhancement convolutional neural network, a confidence distribution with significant spatial heterogeneity is generated. This figure intuitively shows the confidence performance of three typical hazard areas: the red ellipse area represents the high-confidence hazard area, with sharp boundaries and internal probability values ​​close to 1, corresponding to the stable observation area with clear deformation signs at the bottom of the canyon and sufficient image overlap; the orange ellipse area represents the medium-confidence hazard area, which shows a gradual diffusion distribution, corresponding to the weak observation condition area of ​​the mountainside transition zone where the image resolution is uneven due to the fluctuation of the base-to-height ratio; the yellow ellipse area represents the low-confidence hazard area, which is characterized by discrete and fragmented low-density hotspots, corresponding to the area of ​​high slope at the top of the mountain where the model uncertainty is increased due to the lack of texture information.

[0075] The advantage of this network structure lies in the fact that the edge enhancement operation in the second layer increases the feature weights of the boundary regions of potential hazards, the skip connections in the fourth layer preserve multi-scale contextual information, and the dilated convolution in the third layer expands the model's perception range of surface deformation regions without increasing the number of parameters. Compared with traditional fully convolutional networks, it can generate a confidence field with clearer boundaries and more continuous semantics.

[0076] Step S13: Based on the pixel-level hazard identification confidence field generated in step S12, perform a quadtree construction operation according to the geospatial recursive partitioning rules to generate a confidence quadtree.

[0077] The geospatial recursive partitioning rule divides the rectangular geographic area of ​​the current node into four equal sub-rectangles according to the principle of spatial quadrature. Spatial adjacency refers to the adjacent connections of pixels within a node in the four directions (up, down, left, and right) of the geographic coordinate system. The statistical features after confidence fusion refer to the fact that the average confidence of the parent node is the arithmetic mean of the average confidence of the four child nodes, and the variance of the parent node's confidence is the sum of the weighted sum of the variances of the four child nodes' confidence and the sum of the average variances of the confidence among the child nodes. Specifically, to construct a hierarchical confidence quadtree, the construction process is as follows:

[0078] Step S131: Initialize the root node to cover the geographic range of the complete normalized digital orthophoto map, and calculate the average and variance of the confidence scores of all pixels within the root node range as the root node attributes.

[0079] Step S132: Determine whether the number of pixels covered by the current node is greater than 1. If it is greater than 1, take the midpoint of the geographical range of the node along the longitude and latitude directions and divide it into four equal-area sub-rectangles, and create a child node for each sub-rectangle.

[0080] Step S133: Recursively execute the node attribute calculation in step S131 and the partitioning judgment in step S132 until the node coverage area contains only a single pixel, then mark the node as a leaf node and stop partitioning.

[0081] Step S134: During the construction process, the depth value of each node is recorded synchronously (the root node has a depth of 0, and the depth value increases by 1 with each downward division). Leaf nodes store the confidence value and geographic coordinates of a single pixel. Non-leaf nodes store the statistical features after confidence fusion, the parent-child connection relationship between nodes, and the spatial coverage. Leaf nodes also record the spatial connection relationship index with adjacent leaf nodes, forming a spatial adjacency table within the node, ultimately resulting in a confidence quadtree. Please refer to Figure 3 for a schematic diagram of the confidence quadtree structure. This figure illustrates the adaptive multi-scale partitioning process of the confidence field in space. The initial confidence of the overall region is 0.72, serving as the root node of the quadtree. The root node is first partitioned into four equal parts according to the confidence gradient threshold criterion, generating four child nodes: the upper left sub-region has a confidence of 0.85, the upper right sub-region has a confidence of 0.60, the lower left sub-region has a confidence of 0.90, and the lower right sub-region has a confidence of 0.45. Among them, the upper left sub-region (0.85) and lower left sub-region (0.90) were judged as homogeneous regions because their internal confidence distribution was uniform and their gradient changes did not exceed the preset threshold, so further subdivision was stopped and they were retained as leaf nodes; the upper right sub-region (0.60) triggered secondary subdivision because of the moderate degree of confidence gradient fluctuation within it, and was further subdivided into four sub-nodes with confidence levels of 0.70, 0.55, 0.65 and 0.50, respectively, corresponding to the confidence transition characteristics caused by the local attenuation of image overlap in this region; the right The lower subregion (0.45) has the most dramatic internal confidence gradient, triggering a second-level partitioning that generates four child nodes with confidence levels of 0.50, 0.35, 0.40, and 0.55. The child node with a confidence level of 0.35, still exhibiting significant internal heterogeneity, triggers a third-level partitioning, further subdividing it into four smallest granularity units with confidence levels of 0.30, 0.40, 0.25, and 0.45. This corresponds to the region with the highest model uncertainty due to the severe lack of texture information in the high-slope mountain area. This quadtree structure implements an adaptive spatial partitioning strategy that uses coarse-grained representation for high-confidence regions and fine-grained characterization for low-confidence regions. This avoids redundant computation in stable regions while ensuring refined analysis capabilities for regions with concentrated uncertainty.

[0082] This construction method organizes two-dimensional planar data into a tree-level structure through recursive spatial partitioning, enabling subsequent operations to locate confidence statistics at any spatial location within logarithmic time complexity, significantly improving retrieval efficiency compared to the linear complexity of pixel-by-pixel traversal.

[0083] The core advantage of this step lies in transforming the discrete pixel-level hazard identification confidence field into a confidence quadtree with hierarchical aggregation characteristics, solving the problem of lost spatial structure information of confidence in traditional methods. In areas with limited observation conditions, such as high mountains and canyons, UAV aerial surveys often suffer from insufficient image overlap in some areas due to fluctuations in the base-to-height ratio, resulting in drastic spatial fluctuations in the pixel-level confidence field. Traditional raster processing methods struggle to quickly locate low-confidence clusters. The confidence quadtree can quickly identify node levels with drastic confidence fluctuations through hierarchical variance statistics. By comparing the attributes of parent and child nodes, it can automatically detect abnormal drops in local confidence caused by differences in observation conditions, providing rapidly retrievable spatial reliability difference information for subsequent hierarchical decision-making. This structure transforms planar data processing into tree traversal operations, enabling large-scale confidence pattern analysis to be completed in milliseconds. Compared to pixel-by-pixel processing, it offers an order-of-magnitude improvement in computational efficiency. More importantly, it preserves the spatial aggregation topology of confidence, a structural advantage that traditional matrix representations cannot achieve. The generated confidence quadtree can be directly used as the input variable for step S2, providing a foundation for the subsequent construction of weighted decision lattices.

[0084] Step S2: Based on the confidence quadtree and aerial survey parameters, establish a partial order relationship and generate a weighted decision lattice.

[0085] This step addresses the structuring requirements of the confidence quadtree decision-making process. First, it receives the confidence quadtree generated in step S1. Then, combining the basal-to-elevation ratio, image overlap, and terrain texture complexity parameters from UAV aerial surveys, it establishes partial order relationships between nodes and generates a weighted decision lattice. The weighted decision lattice is a lattice structure induced by the partial order set, containing a set of lattice nodes, coverage edges between nodes, comprehensive weight attributes for each node, and a corresponding confidence interval mapping table. Each node in the lattice corresponds to a combination state of confidence interval and observation condition weights, enabling coupled modeling of the uncertainty of observation conditions and confidence.

[0086] Specifically, the steps for establishing the partial order relation and generating the weighted decision lattice are as follows:

[0087] Step S21: Receive the confidence quadtree generated in step S1, extract node information and obtain UAV aerial survey technical parameters, and generate node basic dataset and normalized aerial survey parameters.

[0088] The normalization process involves subtracting the minimum value from the original parameter value and then dividing by the difference between the maximum and minimum values, thus linearly mapping the parameter to the interval between 0 and 1. Specifically, the process for generating the node base dataset and normalized aerial survey parameters is as follows:

[0089] Step S211: Traverse all nodes of the confidence quadtree, extract the confidence statistics (average confidence, confidence variance), spatial range parameters and node depth values ​​of each node, and integrate them to form the node basic dataset.

[0090] Step S212: Obtain UAV aerial survey technical parameters recorded synchronously with the standardized digital orthophoto map, including base-to-height ratio, image overlap, and terrain texture complexity. The base-to-height ratio is the ratio of the UAV's flight altitude to the camera baseline length, calculated by dividing the flight altitude data in the UAV flight control log by the baseline length data in the camera calibration parameters. The image overlap is the proportion of the overlapping area of ​​adjacent flight strip images to the area of ​​a single image, calculated using an image matching algorithm to determine the proportion of corresponding points distributed between adjacent images. The terrain texture complexity is the richness of texture in a local area of ​​the ground surface, obtained by calculating the entropy value of the gray-level co-occurrence matrix of the standardized digital orthophoto map; a higher entropy value indicates more complex texture.

[0091] Step S213: Perform normalization processing on the three aerial survey technical parameters respectively, so that their value range is unified to the interval between 0 and 1, and obtain normalized aerial survey parameters.

[0092] Step S22: Based on the node base dataset and normalized aerial survey parameters generated in step S21, define priority relationships and perform transitive closure operations to generate a partial order relation matrix.

[0093] The priority relation is a binary relation defined on the node set, satisfying reflexivity, antisymmetry, and transitivity. The transitive closure operation is an operation that supplements all indirectly reachable node pairs with direct connections by repeatedly applying the priority relation rules, forming a complete partial order relation. Specifically, to establish a partial order relation reflecting spatial coverage and observation reliability constraints, the process is as follows:

[0094] Step S221: Define the priority relationship. For any two nodes A and B in the node base dataset, first determine the spatial inclusion relationship. If the geographic rectangle of node A completely contains the geographic rectangle of node B, then the spatial inclusion condition is satisfied. Second, calculate the observation condition weight inferiority index of nodes A and B (a weighted sum of the inverse of the base-height ratio, the image overlap complement, and the terrain texture complexity entropy, with each weight coefficient being one-third). If the inferiority index of node A is greater than that of node B, then it is determined that the observation condition weight of A is inferior to that of B. When both the spatial inclusion relationship and the weight inferiority relationship are satisfied, establish a priority connection between nodes A and B (direction from A to B).

[0095] This priority connection means that during the decision-making process, when two nodes have spatial overlap and the observation conditions of the outer node are worse, the outer node should be considered over the inner node. This is because the outer node represents a larger area and has higher observation uncertainty, thus having a more significant impact on the overall hierarchical decision. This rule ensures that the subsequently constructed partial order structure can reflect the dual constraints of spatial scale and observation reliability.

[0096] Step S222: Initialize a square matrix with the same number of nodes as the confidence quadtree as the priority relation matrix. The initial value of the matrix elements is 0. Traverse all established priority connections and set the matrix elements at the corresponding positions to 1.

[0097] Step S223: Perform transitive closure operation on the priority relation matrix and iterate until no new priority connections are generated (if there is a connection from i to j and a connection from j to k, then set the connection from i to k to 1), and finally obtain the partial order relation matrix.

[0098] The advantage of this transitive closure operation is that the computational complexity is polynomial to the number of nodes, and the iteration converges in O(logN) times (in practice, it does not exceed the maximum depth of the quadtree). For a quadtree with thousands of nodes, it can be completed in milliseconds.

[0099] Step S23: Based on the partial order relation matrix generated in step S22, construct a decision lattice and calculate the comprehensive weight of the nodes to generate a weighted decision lattice.

[0100] The decision lattice is a lattice structure induced by a partially ordered set, where any two nodes have a unique minimum upper bound and maximum lower bound; the covering edge is a parent-child node relationship edge directly connected in the lattice, indicating that there are no intermediate nodes between the two nodes; the comprehensive weight attribute is a quantitative weight value characterizing the uncertainty of node confidence and the reliability of observation conditions (calculated jointly by the node's own confidence variance, node depth value, and aerial survey parameter factors; the calculation formula uses the geometric mean of three normalized parameters: confidence variance multiplied by the depth normalization factor, then multiplied by the inverse of the basal-to-height ratio, image overlap complement, and terrain texture complexity entropy); the confidence interval mapping table records the spatial geographic range corresponding to each lattice node and the minimum and maximum confidence values ​​of its internal pixels. Specifically, the process of constructing the decision lattice and calculating the comprehensive weight of nodes is as follows:

[0101] Step S231: Identify incomparable node pairs based on the partial order relation matrix, confirm that the partial order set satisfies the lattice structure condition (any two nodes have a unique minimum upper bound and a maximum lower bound), map each node in the lattice to a node of the weighted decision lattice, and determine the covering edge between nodes (there is a priority connection and there is no intermediate node that makes the connection transitive).

[0102] Step S232: Calculate the comprehensive weight of each grid node. The comprehensive weight is equal to the confidence variance of the corresponding quadtree node confidence multiplied by the depth normalization factor (1 / (node ​​depth value + 1)), and then multiplied by the geometric mean of three normalized parameters: the inverse of the base-height ratio, the image overlap complement, and the terrain texture complexity entropy value (if any parameter is missing, the geometric mean of the other parameters is used as a substitute).

[0103] Nodes with large confidence variance and located in low overlap regions receive higher uncertainty weights, which can accurately reflect the differences in the reliability of node confidence. Please refer to Figure 4 for a schematic diagram of the weighted decision lattice structure and its partial order relation. As shown in Figure 4, the left side is the Hasse diagram representation of the partial order relation, and the right side is the corresponding weighted decision lattice node attribute structure. The two are topologically mapped through covering edges (connected by arrows).

[0104] In engineering practice in high mountains and canyons, UAV aerial surveys suffer from drastic variations in observation conditions due to terrain undulations: Node A (top green node) at the bottom of the canyon covers a large area but has a low base-to-height ratio and insufficient overlap, with a confidence interval as wide as [0.15, 0.95], indicating high uncertainty in this area; Node D (top right) in the stable observation area at the top of the mountain, although also with wide coverage, has a confidence interval as narrow as [0.45, 0.85] due to clear terrain texture, with a weight of 0.79, reflecting that its observation reliability is significantly better than that of the canyon area.

[0105] In the diagram, the color coding of nodes (green-blue-orange) corresponds to the depth level of the confidence quadtree. Nodes B / C / D (second layer, blue) are children of node A, and their weight values ​​(0.72 / 0.68 / 0.79) are coupled with the inverse of the base-height ratio and the overlap complement through the geometric mean formula, realizing the quantitative coupling between observation conditions and confidence uncertainty. If node B corresponds to the left bank slope area of ​​the canyon, its weight of 0.72 is higher than that of node C (0.68), indicating that region B has a higher grading decision influence due to slightly better image overlap. This difference will drive the spatial difference field to produce a positive deviation in region B in the subsequent topological difference calculation, triggering the grading upgrade mechanism. Conversely, nodes E / F / G / H (third layer, orange) are leaf layer nodes, and their low weights (0.45-0.61) accurately characterize the weak observation features of local micro-topographic areas.

[0106] Step S233: Establish a confidence interval mapping table for each grid node, record the spatial geographic range corresponding to the node and the minimum and maximum confidence values ​​of the internal pixels, and store the weighted decision grid in the form of an adjacency linked list (each grid node maintains a list of covering edges, records the direct child node numbers, and the node attributes include node ID, corresponding confidence quadtree node ID, comprehensive weight value, minimum confidence value, maximum confidence value, and geographic range rectangle coordinates), and finally obtain the weighted decision grid.

[0107] This storage structure supports fast parent-child node traversal and weight aggregation operations, enabling subsequent topology difference calculations to be performed efficiently on the lattice structure.

[0108] The core advantage of this step lies in transforming the hierarchical confidence quadtree into a weighted decision lattice with a weight propagation mechanism, solving the problem of fixed boundaries in traditional hierarchical decision-making. In high mountain and canyon areas, the base-to-elevation ratio of UAV aerial surveys often fluctuates dramatically due to terrain undulations, and the image overlap can differ by several times between the bottom of the canyon and the top of the mountain. Traditional fixed-threshold grading cannot adapt to this spatial heterogeneity. The weighted decision lattice explicitly encodes the spatial coverage relationship and the inferiority relationship of observation conditions through the partial order relationship between nodes, transforming the grading standard from a rigid hyperplane into a lattice structure that can scale with observation conditions. The coverage edges of the lattice structure support the upward aggregation of weights along the edges, allowing higher-level nodes to perceive the cumulative uncertainty of lower-level nodes as a whole. This weight propagation mechanism cannot be achieved in decision tables. When the confidence variance increases due to insufficient overlap at the bottom of the canyon, the comprehensive weight of the corresponding node automatically increases, thereby affecting the overall decision tendency of the lattice, realizing coupled modeling of observation conditions and confidence uncertainty. Although the confidence interval mapping table and the sum of child node weights are not directly used in the current hierarchical decision-making, they reserve an interface for subsequent multi-scale uncertainty quantification analysis, demonstrating the scalability of the architecture and providing a weight-aware lattice structure foundation for subsequent topological difference calculation. This dynamic adaptability is something that traditional decision tree or rule chain methods cannot achieve.

[0109] Step S3: Calculate the topological dissimilarity degree and generate the spatial dissimilarity field based on the weighted decision lattice and the standard hierarchical decision surface.

[0110] This step focuses on quantifying the difference between the confidence topology and the standard classification. First, it receives the weighted decision grid generated in step S2, constructs the standard hazard classification decision surface, calculates the topological difference degree between the two, and generates a spatial difference field. The spatial difference field is a continuous field distribution smoothed from discrete node topological difference indices. The field value at each geographic coordinate represents the degree of deviation between the original confidence topology and the standard hazard classification decision surface at that location. A higher field value indicates a greater mismatch between the identification reliability of that area and the standard classification system.

[0111] Specifically, the steps for calculating the topological difference degree and generating the spatial difference field are as follows:

[0112] Step S31: Obtain the threshold values ​​of quantitative indicators in the geological hazard hazard classification standard, and construct the standard hazard classification decision surface.

[0113] The standard hazard classification decision surface is a combination of boundary hyperplanes formed by the combination of deformation threshold, hazard area threshold, and slope threshold in a three-dimensional feature space (deformation axis, hazard area axis, and slope axis). The deformation threshold is the critical value at which the cumulative surface deformation reaches different risk levels (obtained through stability evaluation standards in geological exploration specifications); the hazard area threshold is the critical value corresponding to different risk levels for the hazard area area (obtained through statistical analysis of historical disaster cases); and the slope threshold is the critical value corresponding to different risk levels for the average slope of the hazard area (determined through research on the mechanism of geological disasters). Specifically, the process for constructing the standard hazard classification decision surface is as follows:

[0114] Step S311: Establish a three-dimensional rectangular coordinate system, with the horizontal axis defined as the deformation, the vertical axis defined as the area of ​​the hidden danger, and the vertical axis defined as the slope.

[0115] Step S312: Project the deformation threshold, the hidden danger area threshold, and the slope threshold vertically onto the corresponding coordinate axes to form three orthogonal planes.

[0116] Step S313: Intersect the three orthogonal planes to divide the feature space into multiple sub-regions (each sub-region corresponds to a risk level). Combine the boundary hyperplanes of all sub-regions to form a piecewise linear standard hazard classification decision surface, which is stored in the form of parametric equations (including the normal vector and intercept of each hyperplane).

[0117] Step S32: Based on the weighted decision lattice generated in step S2 and the standard hazard classification decision surface generated in step S31, calculate the topological difference index of each node to obtain the node topological difference set.

[0118] Wherein, the vertical distance is the shortest Euclidean distance from the feature space coordinate point to the standard hazard classification decision surface; the basic weight component is the sum of the node's comprehensive weight value and the comprehensive weight of all child nodes within the node's coverage area; the weight adjustment coefficient is the node's confidence interval width (maximum value minus minimum value plus 1), and the larger this width, the more drastic the fluctuation in confidence within the node; the depth normalization factor is 1 / (node ​​depth value + 1), used to enhance the difference contribution of shallow nodes (shallow nodes correspond to a larger spatial range, and their impact on the overall classification decision should be greater than that of deep, small-range nodes). Specifically, the process for calculating the topological difference index of each node is as follows:

[0119] Step S321: Traverse all nodes of the weighted decision grid, extract the average confidence score and confidence score interval within the geographical range corresponding to each node, and map the average confidence score to coordinate points in the three-dimensional feature space.

[0120] Step S322: Calculate the vertical distance from the coordinate point to the standard hazard classification decision surface, and use it as a spatial distance component to accurately quantify the geometric deviation between the coordinate point and the standard classification boundary.

[0121] Step S323: Extract the comprehensive weight value of the node and the sum of the comprehensive weights of all child nodes within the node's coverage area as the basic weight component; calculate the weight adjustment coefficient using the minimum and maximum values ​​in the confidence interval mapping table, and multiply the basic weight component by the weight adjustment coefficient to obtain the final weight component, taking into account both the node's own weight and internal confidence fluctuation characteristics.

[0122] Step S324: Calculate the topology dissimilarity index, which is the spatial distance component multiplied by the weight component, then multiplied by the depth normalization factor, and integrate the topology dissimilarity indices of all nodes to form a node topology dissimilarity set.

[0123] The topological dissimilarity index quantifies the topological differences of lattice structures through multi-dimensional component coupling, breaking through the limitations of traditional methods that only consider geometric distance.

[0124] Step S33: Based on the set of node topological differences generated in step S32, perform bilinear interpolation to generate a spatial difference field.

[0125] The bilinear interpolation method utilizes the difference values ​​of four adjacent grid nodes, weighted by distance, to calculate the field value at non-node locations. This method maintains the continuity and smoothness of the field values. The storage structure of the spatial difference field uses the same raster size as the standardized digital orthophoto map. Each raster cell stores a floating-point field value, which is normalized and mapped to the range of -1 to +1 (positive values ​​indicate that the confidence topology of the area is more complex than the standard classification, and negative values ​​indicate that it is simpler than the standard classification), facilitating subsequent retrieval by geographic coordinate index. Specifically, the process of performing bilinear interpolation to generate the spatial difference field is as follows:

[0126] Step S331: Locate each node topology difference index in the node topology difference set according to its corresponding geographic coordinates.

[0127] Step S332: For all non-node locations in the geographic coordinate space, the field value is calculated using the bilinear interpolation method: locate the four nearest grid nodes, calculate the horizontal and vertical distances between the location and the coordinates of the four nodes, multiply the topological dissimilarity of each node by the square of the inverse of the distance as the weight, and sum the four weights after normalization to obtain the field value of the location. This avoids the step-like discontinuity of the spatial difference field caused by the discrete distribution of grid nodes.

[0128] Step S333: Integrate the field values ​​of all raster units to form a spatial difference field covering the entire range of the standardized digital orthophoto map, and store it according to the corresponding storage structure.

[0129] The core advantage of this step lies in transforming the abstract structural differences of the weighted decision lattice into a spatially localizable continuous field, solving the problem that traditional methods cannot quantify the degree of misalignment between the confidence field and the hierarchical boundary space. The partial order relationships and weight distributions among nodes within the weighted decision lattice constitute a complex topological form, which traditional spatial analysis, calculating only a single distance, cannot capture. The topological dissimilarity index incorporates the hierarchical structure of the lattice by introducing a node depth normalization factor, allowing for a reasonable weighting of the differences at different spatial scales. In regions with sharp confidence gradients, both the spatial distance and weight components of the corresponding lattice nodes increase simultaneously, leading to a significant increase in topological dissimilarity, accurately indicating the degree of mismatch between this region and the standard hierarchical boundary. Bilinear interpolation smooths the discrete node differences into a continuous field, making the spatial dissimilarity field usable for subsequent spatial optimization calculations. Although the confidence interval mapping table and the sum of child node weights are not directly used in the current hierarchical decision, they reserve interfaces for subsequent multi-scale uncertainty quantification analysis, demonstrating the scalability of the architecture. This method breaks through the technical limitations of traditional buffer analysis, which only considers geometric distance, and realizes the field-like expression of lattice order structure differences. It provides a topological basis with spatial constraints for hierarchical decision-making, which is a structural sensitivity that simple superposition analysis cannot achieve.

[0130] Step S4: Based on the spatial difference field and weighted decision lattice, perform adaptive hierarchical judgment and output the evaluation results.

[0131] This step, focusing on the difference-driven hazard classification requirement, first receives the spatial difference field generated in step S3 and the weighted decision grid generated in step S2. Under the difference constraint, it performs adaptive classification determination and outputs the final risk level and credibility assessment results. The final risk level and credibility assessment results are a comprehensive result including the risk classification map and the posterior credibility label for each hazard point. The risk classification map is stored in GeoTIFF format, compatible with mainstream GIS software, facilitating subsequent visualization and analysis.

[0132] Specifically, the steps for performing adaptive grading determination and outputting evaluation results are as follows:

[0133] Step S41: Receive the spatial difference field generated in step S3, set the difference tolerance threshold, and identify the areas that need level adjustment.

[0134] The difference tolerance threshold is the maximum allowable absolute value of topological difference; exceeding this value necessitates a grade adjustment. The balance coefficient is a parameter that adjusts the strictness of the tolerance threshold, ranging from 0.5 to 2.0 (a larger coefficient indicates a wider tolerance range and a more conservative adjustment). Specifically, the process of setting the difference tolerance threshold and identifying areas requiring grade adjustment is as follows:

[0135] Step S411: Calculate the arithmetic mean of the field values ​​of all grid cells in the spatial difference field. Calculate the sum of squares of the differences between all field values ​​and the mean, divide by the total number of grid cells to obtain the variance, and calculate the square root of the variance to obtain the standard deviation.

[0136] Step S412: Multiply the standard deviation by the balance coefficient to obtain the tolerance threshold for variability (the balance coefficient is determined according to the strictness of risk control: 0.5 is used near residential areas to make the tolerance threshold more stringent and adjustments more frequent; 2.0 is used in remote mountainous areas to make the tolerance threshold more lenient and adjustments more conservative).

[0137] This adaptive threshold setting method allows the sensitivity of level adjustment to be dynamically adjusted according to the application scenario, taking into account the risk management needs of different scenarios.

[0138] Step S413: Traverse the difference value corresponding to each geographic unit in the spatial difference field, identify areas where the absolute value of the difference exceeds the difference tolerance threshold, and mark them as areas that need level adjustment.

[0139] Step S42: Based on the adjustment area identified in step S41, the spatial difference field generated in step S3, and the weighted decision grid generated in step S2, perform a level correction operation to obtain the corrected risk level distribution map.

[0140] The risk level correction function is a piecewise function that maps the original risk level to the adjusted risk level. The median weight is the value in the middle position after sorting the comprehensive weight values ​​of all nodes in the weighted decision grid. The global mean is the average of the sum of the comprehensive weights of the child nodes within the coverage area of ​​all nodes in the weighted decision grid. The spatial continuity constraint means that the adjusted risk level distribution in the local neighborhood should meet the smoothness condition to avoid abrupt changes in risk levels between adjacent locations (mathematically, for any two hazard points whose geographical coordinate distance is less than the smoothing radius, the absolute value of their corrected risk level difference should be less than the risk level difference tolerance value, the smoothing radius is taken as 10 times the spatial resolution of the digital orthophoto map, and the risk level difference tolerance value is taken as level 1). The local smoothing filter is a spatial filtering operator used to eliminate abrupt changes in risk levels. Specifically, the risk level correction process is as follows:

[0141] Step S421: For each geographic unit that needs adjustment, find the smallest coverage node corresponding to it in the weighted decision grid, and obtain the comprehensive weight value of the node, the sum of the comprehensive weights of its child nodes, and the original risk level (determined by the standard hazard classification decision surface).

[0142] Step S422: Calculate the ascending order of the comprehensive weight values ​​of all nodes in the weighted decision lattice, and take the value at the middle position as the median weight; calculate the arithmetic mean of the sum of the comprehensive weights of all child nodes as the global mean.

[0143] Step S423: Perform level correction: If the node's overall weight is greater than the median weight or the sum of the child node's overall weights is greater than the global mean, and the spatial difference field value is positive, then the original level is increased by one level; if the node's overall weight is greater than the median weight or the sum of the child node's overall weights is greater than the global mean, and the spatial difference field satisfies the condition and the spatial difference field value is negative, then the original level is decreased by one level; otherwise, the original level remains unchanged (the level value range is 1-5, and the boundary value is taken when it exceeds the range).

[0144] The hierarchical correction function couples the field value of spatial difference field with the weighted decision lattice node weights, so that the correction direction and magnitude simultaneously reflect the topological differences and the uncertainty of observation conditions, realizing a complete mapping from geometric differences to decision adjustment.

[0145] Step S424: Perform a local smoothing filter operation on the corrected rank distribution map: initialize the blank smoothing result map, traverse each grid cell, collect the rank values ​​of its 8 neighboring cells, calculate the geographical distance between the current cell and the neighboring cells and take the square of the inverse distance as the weight, and weight the rank value of the current cell and the rank values ​​of the neighboring cells according to the weight to obtain the smoothed rank value and write it into the result map.

[0146] This filtering process eliminates abrupt changes in risk levels caused by abnormal weights of individual nodes, ensuring that the final risk level distribution conforms to the spatial continuity pattern of geological disasters and meets spatial continuity constraints.

[0147] Step S43: Based on the corrected risk level distribution map generated in step S42, the spatial difference field generated in step S3, and the weighted decision lattice generated in step S2, calculate the posterior confidence assessment and output the final result.

[0148] The posterior credibility assessment is a quantitative index of credibility for level determination calculated based on spatial dissimilarity and node weights (with a value range of 0 to 1, the closer to 1, the higher the credibility). The risk grading map is stored in GeoTIFF format, with pixel grayscale values ​​encoding the risk level (levels 1 to 5 correspond to different grayscale values). The metadata area stores the credibility assessment value as an additional band, ensuring that each pixel contains not only level information but also credibility information. Specifically, the process for calculating the posterior credibility assessment and outputting the final result is as follows:

[0149] Step S431: For each potential hazard point, extract the absolute value of its corresponding spatial difference field value and the comprehensive weight value of the weighted decision grid node.

[0150] Step S432: Calculate the credibility score: The score = (1 - absolute value of field value / maximum absolute value of field value) × (1 - normalized value of node comprehensive weight).

[0151] The smaller the absolute value of the field value, the smaller the topological difference at that point, and the more reliable the grade determination. The smaller the comprehensive weight of the nodes, the better the observation conditions, and the more reliable the grade determination. This score can comprehensively reflect the reliability of the grade determination. The score ranges from 0 to 1, and the closer the value is to 1, the higher the reliability of the grade determination.

[0152] Step S433: Attach the credibility score as a label to each potential hazard point in the corrected risk level distribution map to generate a risk grading map with credibility labels. Store the map in GeoTIFF format (pixel grayscale values ​​correspond to risk levels 1-5, and the metadata area stores the credibility assessment value). Finally, output the final risk level and credibility assessment results.

[0153] The core advantage of this step lies in directly transforming the topological mismatch information of the spatial difference field into correction constraints for hierarchical decision-making, solving the problem of discrepancies between the credibility of traditional hierarchical results and nominal levels. Traditional methods lack topological consistency checks on hierarchical results, still using a uniform standard in areas with significant spatial differences, leading to a systematic underestimation of the actual credibility of risk quantification results. This step automatically identifies areas with significant topological mismatch through a difference tolerance threshold mechanism, correcting levels only for key nodes with high comprehensive weights, avoiding over-adjustment while ensuring the accuracy of hierarchical classification in important areas. The level correction function couples the spatial difference field value with the weighted decision lattice node weights, ensuring that the correction direction and magnitude simultaneously reflect both topological differences and observational uncertainties, achieving a complete mapping from geometric differences to decision adjustments. The local smoothing filter enforces spatial continuity constraints, ensuring that the corrected level distribution is consistent with the spatial continuity of geological hazards, avoiding level islands caused by data noise. The posterior credibility assessment provides a quantitative basis for each hazard point, allowing risk management personnel to focus on reviewing high-risk but low-credibility areas based on credibility values, achieving a dual output of risk quantification and credibility. In complex terrains such as high mountains and canyons, this method can automatically identify systematic level deviations caused by drastic fluctuations in the base height ratio. It corrects the risk level by weighting the confidence level through the weighting mechanism of the weighted decision lattice, so that the final risk level has both spatial topological consistency and observation condition adaptability. It completes the paradigm shift from discrete grading to difference-driven grading, and fundamentally solves the technical problem of mismatch between the confidence field and the spatial topology of the grading decision boundary in the background technology.

[0154] Example 2:

[0155] Based on Example 1, this embodiment provides a system for intelligent identification and classification of geological hazard hazard points using UAV aerial surveying, as shown in Figure 5, including:

[0156] A system for intelligent identification and classification of geological hazard hazard points using UAV aerial surveys, comprising the aforementioned method for intelligent identification and classification of geological hazard hazard points using UAV aerial surveys, the system including:

[0157] Image processing and confidence quadtree construction module: used to acquire UAV aerial survey image data, generate pixel-level hazard identification confidence field and construct confidence quadtree;

[0158] Partial order relation establishment and weighted decision lattice generation module: used to establish partial order relations and generate weighted decision lattices based on confidence quadtrees and aerial survey parameters;

[0159] Topological Dissimilarity Calculation and Spatial Dissimilarity Field Generation Module: Used to calculate topological dissimilarity and generate spatial dissimilarity fields based on weighted decision lattices and standard hierarchical decision surfaces;

[0160] Adaptive grading judgment and evaluation result output module: Based on the spatial difference field and weighted decision lattice, it performs adaptive grading judgment and outputs the final risk level and credibility evaluation results.

Claims

1. A method for intelligent identification and classification of geological hazard hazard points using UAV aerial surveying, characterized in that, The method includes: S1: acquiring UAV aerial survey image data, generating a pixel-level hazard identification confidence field, and constructing a confidence quadtree; S2: establishing a partial order relation and generating a weighted decision lattice based on the confidence quadtree and aerial survey parameters; S3: calculating the topological difference degree and generating a spatial difference field based on the weighted decision lattice and a standard hierarchical decision surface; S4: performing adaptive hierarchical judgment and outputting the evaluation result based on the spatial difference field and the weighted decision lattice; wherein, the step of establishing a partial order relation and generating a weighted decision lattice includes: receiving the confidence quadtree, extracting node information and acquiring UAV aerial survey technical parameters, generating a node basic dataset and normalized aerial survey parameters; the UAV aerial survey technical parameters include basal-to-height ratio, image overlap, and terrain texture complexity; defining a priority relation and performing a transitive closure operation based on the node basic dataset and normalized aerial survey parameters to generate a partial order relation matrix; constructing a decision lattice and calculating the node comprehensive weight based on the partial order relation matrix to generate a weighted decision lattice; the comprehensive weight refers to the variance of the confidence quadtree node corresponding to the node multiplied by the depth. The normalization factor is then multiplied by the geometric mean of three normalized parameters: the reciprocal of the basal-to-height ratio, the image overlap complement, and the terrain texture complexity entropy. The steps of defining the priority relationship and performing the transitive closure operation include: defining the priority relationship; for any two nodes, determining the spatial inclusion relationship and the observation condition weight inferiority relationship; establishing a priority connection when both the spatial inclusion relationship and the observation condition weight inferiority relationship are satisfied; the spatial inclusion relationship refers to the fact that the geographic range rectangle of node A completely contains the geographic range rectangle of node B; the observation condition weight inferiority relationship refers to obtaining the observation condition weight inferiority index by calculating the weighted sum of the reciprocal of the basal-to-height ratio, the image overlap complement, and the terrain texture complexity entropy of the node; if the observation condition weight inferiority index of node A is greater than that of node B, then the observation condition weight of node A is determined to be inferior to that of node B; initializing the priority relationship matrix by setting the matrix elements corresponding to the established priority connections to 1; performing the transitive closure operation on the priority relationship matrix, iterating until no new priority connections are generated, and obtaining the partial order relationship matrix.

2. The method for intelligent identification and classification of geological hazard hazard points by UAV aerial survey according to claim 1, characterized in that, The steps for generating a pixel-level hazard identification confidence field and constructing a confidence quadtree include: acquiring multi-view images from UAV aerial surveys, performing aerial triangulation and orthorectification operations to generate a standardized digital orthophoto map; based on the standardized digital orthophoto map, calling an edge-enhanced convolutional neural network to perform pixel-level hazard identification operations and generate a pixel-level hazard identification confidence field; based on the pixel-level hazard identification confidence field, performing a quadtree construction operation according to a geospatial recursive partitioning rule to generate a confidence quadtree; the geospatial recursive partitioning rule refers to dividing the rectangular geographical area of ​​the current node into four equal sub-rectangular areas according to the principle of spatial four-eighths division.

3. The method for intelligent identification and classification of geological hazard hazard points by UAV aerial survey according to claim 2, characterized in that, The processing steps of the edge-enhanced convolutional neural network include: inputting a standardized digital orthophoto image into the first layer of the network and extracting basic visual features through a standard convolutional kernel; in the second layer of the network, introducing a Laplacian edge operator and multiplying it element-wise with the feature map of the first layer to enhance the feature response of the boundary area of ​​the hazard; expanding the receptive field of view through dilated convolution in the third layer of the network to capture information on the surface deformation area; concatenating the features of the third layer with the features of the first layer through skip connections in the fourth layer of the network to retain multi-scale contextual information; and using a 1×1 convolution to reduce the dimensionality of the concatenated features in the fifth layer of the network, outputting a dual-channel probability map of the hazard category and the non-hazard category after inputting it into the classification head, and taking the hazard category channel as the pixel-level hazard identification confidence field.

4. The method for intelligent identification and classification of geological hazard hazard points by UAV aerial survey according to claim 1, characterized in that, The steps for calculating topological dissimilarity and generating a spatial dissimilarity field include: obtaining the quantitative index thresholds in the geological hazard hazard classification standard and constructing a standard hazard classification decision surface; the quantitative index thresholds include deformation threshold, hazard area threshold, and slope threshold; the standard hazard classification decision surface refers to the combination of boundary hyperplanes formed by the deformation threshold, hazard area threshold, and slope threshold in a three-dimensional feature space composed of the deformation axis, hazard area axis, and slope axis; calculating the topological dissimilarity index of each node based on the weighted decision grid and the standard hazard classification decision surface to obtain a node topological dissimilarity set; and performing bilinear interpolation based on the node topological dissimilarity set to generate a spatial dissimilarity field; the bilinear interpolation operation refers to the operation of using the dissimilarity values ​​of four adjacent grid nodes and performing a weighted average according to distance weights to calculate the field value of non-node locations.

5. The method for intelligent identification and classification of geological hazard hazard points by UAV aerial survey according to claim 4, characterized in that, The steps for calculating the topology dissimilarity index for each node include: traversing all nodes in the weighted decision lattice, extracting the average confidence score and confidence score interval, and mapping the average confidence score to coordinate points in the three-dimensional feature space; calculating the vertical distance from the coordinate point to the standard hazard classification decision surface as the spatial distance component; extracting the sum of the node's comprehensive weight value and the comprehensive weight of its child nodes as the basic weight component, and calculating the weight adjustment coefficient in combination with the confidence score interval width to obtain the final weight component; calculating the topology dissimilarity index, which is the spatial distance component multiplied by the weight component, and then multiplied by the depth normalization factor, and integrating them to form the node topology dissimilarity set.

6. The method for intelligent identification and classification of geological hazard hazard points by UAV aerial survey according to claim 5, characterized in that, The steps for performing adaptive classification judgment and outputting evaluation results include: receiving a spatial difference field, setting a difference tolerance threshold, and identifying areas that require level adjustment; the difference tolerance threshold is obtained by calculating the standard deviation of all grid cell values ​​in the spatial difference field and then multiplying it by a balance coefficient; based on the adjustment area, the spatial difference field, and the weighted decision grid, performing a level correction operation to obtain a corrected risk level distribution map; and based on the corrected risk level distribution map, the spatial difference field, and the weighted decision grid, calculating the posterior confidence assessment and outputting the final result.

7. The method for intelligent identification and classification of geological hazard hazard points by UAV aerial survey according to claim 6, characterized in that, The steps for performing risk level correction include: for each geographic unit requiring adjustment, finding the corresponding minimum coverage node, obtaining the node's comprehensive weight value, the sum of the comprehensive weights of its sub-nodes, and the original risk level; calculating the median of the comprehensive weights of all nodes in the weighted decision grid and the global mean of the sum of the comprehensive weights of its sub-nodes; performing risk level correction based on the node's comprehensive weight, the sum of the comprehensive weights of its sub-nodes, and the spatial difference field value; the rules for risk level correction are as follows: if the node's comprehensive weight is greater than the median weight or the sum of the comprehensive weights of its sub-nodes is greater than the global mean, and the spatial difference field value is positive, then the original risk level is increased by one level; if the node's comprehensive weight is greater than the median weight or the sum of the comprehensive weights of its sub-nodes is greater than the global mean, and the spatial difference field value is negative, then the original risk level is decreased by one level; otherwise, the original risk level remains unchanged; and performing a local smoothing filter operation on the corrected risk level distribution map to obtain the corrected risk level distribution map.

8. A system for intelligent identification and classification of geological hazard hazard points using UAV aerial surveying, which is used to implement the method for intelligent identification and classification of geological hazard hazard points using UAV aerial surveying as described in any one of claims 1-7, characterized in that, The system includes: an image processing and confidence quadtree construction module for acquiring UAV aerial survey image data, generating a pixel-level hazard identification confidence field, and constructing a confidence quadtree; a partial order relation establishment and weighted decision lattice generation module for establishing a partial order relation and generating a weighted decision lattice based on the confidence quadtree and aerial survey parameters; a topological difference calculation and spatial difference field generation module for calculating topological difference and generating a spatial difference field based on the weighted decision lattice and standard hierarchical decision surface; and an adaptive hierarchical judgment and evaluation result output module for performing adaptive hierarchical judgment and outputting the final risk level and credibility evaluation result based on the spatial difference field and weighted decision lattice.

Images