An unmanned aerial vehicle image geological disaster hidden danger automatic checking method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING MUNICIPAL ROAD & BRIDGE
- Filing Date
- 2026-03-12
- Publication Date
- 2026-05-29
Smart Images
Figure CN122116208A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition technology, and more specifically, to a method and system for automatically investigating potential geological hazards using drone images. Background Technology
[0002] Due to their high mobility and coverage efficiency, drone aerial photography has become a common data source for field investigations and post-disaster emergency assessments of geological hazards such as landslides, collapses, and debris flows. With the development of computer vision and pattern recognition technologies, automatic identification, localization, and classification based on image content are gradually becoming more widespread in this field, supporting the rapid formation of suspected hazard area distribution and risk warnings, reducing reliance on manual point-by-point interpretation.
[0003] However, some existing technologies focus on acquiring images and identifying and outputting the location and category confidence level of risks. For example, invention patent CN113065455A uses a drone to capture images of riverbank slopes and record the shooting location. The images are then input into a prediction model to output the location and category confidence level of landslide risks. However, its identification output usually remains at the level of a single identification result, and it cannot accumulate identification results from multiple frames. This can easily lead to false alarms that cannot be quickly verified in the same task, or missed alarms that lack supplementary evidence. Other geological disaster detection solutions emphasize panoramic stitching, generating digital elevation models and orthophotos before identifying potential hazards, such as CN112923904B. This approach is more inclined to model first and then identify. In emergency investigation scenarios, this can easily lead to long processing links, insufficient timeliness, and the inability to allocate the same acquisition and processing resources to low-risk areas. As a result, it is difficult to focus acquisition resources on suspected hazard areas and form verification evidence within limited flight time.
[0004] Therefore, it is necessary to design an automatic method and system for investigating potential geological hazards using UAV imagery to solve the problems existing in the current technology. Summary of the Invention
[0005] In view of this, the present invention proposes an automatic method and system for investigating geological disaster hazards using UAV imagery, aiming to solve the problems in the current technology of UAV geological disaster hazard investigation, such as the difficulty in accumulating single identification results across frames, the inability to focus on suspected areas within a limited flight time and achieve rapid verification of the same task, resulting in false alarms that are difficult to verify and missed alarms that lack evidence.
[0006] This invention proposes an automatic method for detecting potential geological hazards using UAV imagery, comprising:
[0007] Construct a geographic grid for the area to be investigated, control the drone to perform coarse-scale flight, collect coarse-scale images and obtain pose information;
[0008] The coarse-arranged image is used to identify potential hazards, thereby obtaining candidate hazard regions, hazard categories, and identification confidence levels. The candidate hazard regions are then projected onto a geographic grid based on the pose information and spatially accumulated. A set of suspected hazard regions is then selected based on the accumulated occurrence frequency threshold and the accumulated identification confidence level threshold.
[0009] Based on the set of suspected potential hazard areas, a review route is planned and a drone is controlled to collect review images and review pose information. Review candidate identification is performed on the review images, and the review pose information is projected onto the geographic coordinate system to obtain the review geographic candidate area.
[0010] Under the geographic coordinate system, a consistency determination is performed to determine whether the same suspected hidden danger area forms a verification geographic candidate area that overlaps in no less than two verification images and reaches a preset overlap threshold; based on the consistency determination result, the hidden danger investigation result is output, which includes the location of the suspected hidden danger area, the hidden danger category, and the verification conclusion.
[0011] Furthermore, distortion correction and denoising are performed on the coarse-arrayed image; anomaly maps are generated based on the grayscale differences, texture differences, and edge intensity of the coarse-arrayed image; threshold segmentation is performed on the anomaly maps and connected component extraction is used to obtain the potential hazard candidate regions;
[0012] For each candidate hazard region, calculate its area, aspect ratio, boundary curvature, texture uniformity, and contrast with its surroundings, and match it with a preset hazard category discrimination rule set. Determine the hazard category based on the matching results.
[0013] When all discrimination rules are met, a preset high confidence value is used as the recognition confidence value; when the key discrimination rules are met, a preset medium confidence value is used as the recognition confidence value; and for the rest, a preset low confidence value is used as the recognition confidence value.
[0014] Furthermore, the spatial accumulation includes: mapping the projection results of all the coarse-arranged images to the same grid area and counting the number of occurrences, and accumulating statistics on the identification confidence; the set of suspected hidden danger areas is obtained by merging consecutively adjacent grid areas.
[0015] Furthermore, when accumulating and statistically analyzing the identification confidence scores, the identification confidence scores mapped to the same grid area are weighted and accumulated, with the weights determined based on the intersection area of the projections of the hazard candidate area and the grid area.
[0016] Furthermore, when screening the set of suspected potential hazard areas, the method also includes: after merging the consecutive adjacent grid areas, performing a threshold screening on the area of the merged area to remove merged areas with an area smaller than a preset area threshold.
[0017] Furthermore, when planning the review route, the process includes: obtaining a return safety margin, determining the review priority based on the return safety margin and the cumulative occurrence and cumulative identification confidence of the merged area, and selecting at least one merged area for review flight based on the review priority.
[0018] Furthermore, the review priority is determined according to the rule of prioritizing the cumulative occurrence frequency and then the cumulative identification confidence level. Under the premise of satisfying the return safety margin, a preset number of merged areas with the highest review priority are selected to enter the review route.
[0019] Furthermore, the consistency determination includes: matching the candidate geographic regions for verification of all the verification images with the merged region; outputting a verification pass conclusion when the match is successful, and outputting a pending verification conclusion when the match is unsuccessful.
[0020] Furthermore, the matching is valid if at least one of the following conditions is met:
[0021] The ratio of the overlapping area of the verification geographic candidate region and the corresponding merged region to the area of the merged region reaches the preset overlap threshold, or the geographic center point of the verification geographic candidate region falls into the merged region and the distance between the geographic center point of the merged region and the geographic center point of the merged region is not greater than the preset distance threshold.
[0022] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a geographic grid and projecting the candidate areas of potential hazards in the coarse-sorting stage, combined with pose information, onto a unified spatial reference, evidence accumulation and dual-threshold screening across multiple frames and perspectives are achieved, suppressing the spread of false alarms caused by single-frame false detections, and elevating suspected potential hazard areas from single-identification results to spatial objects with repeated occurrence and accumulated confidence support; for the set of suspected potential hazard areas, a review route is automatically planned and review images are acquired, shifting limited flight time and acquisition bandwidth from low-risk areas to high-suspect areas, reducing invalid coverage and redundant processing; in the review stage, the candidate areas for review are projected again onto the geographic coordinate system, and the geographic overlap of no less than two review images is used as the consistency judgment criterion, transforming the review conclusion from subjective judgment into objective evidence, thereby shortening the link time from acquisition to conclusion output without relying on panoramic stitching and full-domain modeling, improving the timeliness of emergency investigation, and forming verifiable output within the same task, reducing the risk spillover caused by missed reports and reducing the review cost caused by false reports.
[0023] On the other hand, this application also provides an automatic geological hazard hazard investigation system based on UAV imagery, used to apply the above-mentioned automatic geological hazard hazard investigation method based on UAV imagery, including:
[0024] The data acquisition unit is configured to construct a geographic grid of the area to be investigated, control the drone to perform coarse-scale flight, acquire coarse-scale images, and obtain pose information.
[0025] The identification unit is configured to perform hazard candidate identification on the coarse-grained image, obtain hazard candidate areas, hazard categories and identification confidence levels; project the hazard candidate areas onto a geographic grid according to the pose information and perform spatial accumulation; and filter a set of suspected hazard areas according to a cumulative occurrence threshold and a cumulative identification confidence level threshold.
[0026] The re-collection unit is configured to plan a re-examination route based on the set of suspected hidden danger areas and control the UAV to collect re-examination images and re-examination pose information, perform re-examination candidate identification on the re-examination images, and project the re-examination pose information onto the geographic coordinate system to obtain the re-examination geographic candidate area.
[0027] The determination unit is configured to perform consistency determination in the geographic coordinate system, and determine whether the same suspected hidden danger area forms a verification geographic candidate area that overlaps in no less than two verification images and reaches a preset overlap threshold; and output the hidden danger investigation result based on the consistency determination result, which includes the location of the suspected hidden danger area, the hidden danger category and the verification conclusion.
[0028] It is understandable that the aforementioned method and system for automatically identifying potential geological hazards using drone imagery have the same beneficial effects, and will not be elaborated upon here. Attached Figure Description
[0029] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:
[0030] Figure 1 A flowchart of the automatic geological hazard investigation method using UAV imagery provided in this embodiment of the invention;
[0031] Figure 2 A functional block diagram of the UAV imagery-based automatic geological hazard hazard investigation system provided in this embodiment of the invention. Detailed Implementation
[0032] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0033] In some embodiments of this application, see Figure 1 As shown, this application proposes an automatic method for investigating potential geological hazards using UAV imagery, comprising:
[0034] S100: Construct a geographic grid for the area to be investigated, control the UAV to perform coarse-scale flight, collect coarse-scale images, and obtain pose information.
[0035] S200: Perform hazard candidate identification on coarse-grained images to obtain hazard candidate regions, hazard categories, and identification confidence levels. Project hazard candidate regions onto a geographic grid based on pose information and perform spatial accumulation. Filter the set of suspected hazard regions based on accumulated occurrence thresholds and accumulated identification confidence thresholds.
[0036] S300: Based on the set of suspected potential hazard areas, plan the verification route and control the UAV to collect verification images and verification pose information. Perform verification candidate identification on the verification images and project the verification pose information onto the geographic coordinate system to obtain the verification geographic candidate area.
[0037] S400: Perform consistency determination in the geographic coordinate system to determine whether the same suspected hazard area overlaps in no fewer than two review images to reach a preset overlap threshold as a candidate geographic area for review. Output the hazard investigation results based on the consistency determination results. The hazard investigation results include the location of the suspected hazard area, the hazard category, and the review conclusion.
[0038] Specifically, in this application, "pose information" refers to the instantaneous exterior orientation elements acquired by the UAV corresponding to each image, including at least three-dimensional position coordinates and attitude angles (or equivalent rotation representations), and may further include timestamps and camera intrinsic parameters. "Geographic coordinate system" refers to a unified geographic reference coordinate system or engineering coordinate system used to calculate the area and distance of candidate regions, grids, and verification results under the same plane / spatial reference. "Geographic grid" refers to a set of regular or irregular grids obtained by discretizing the area to be investigated under a geographic coordinate system, where each grid region has a unique grid identifier, boundary, and area. "Projection" refers to the process of converting the candidate region boundaries in the image pixel space into candidate geographic region boundaries under a geographic coordinate system based on pose information. Projection includes: back-projecting multiple spatial rays based on pixels within the camera; intersecting the spatial rays with the ground elevation surface to obtain the corresponding geographic coordinate point sequence, and forming the candidate geographic region boundary from the geographic coordinate point sequence; wherein, the ground elevation surface can be provided by a pre-obtained digital elevation model, or approximated by the average ground elevation of the area to be investigated in an emergency coarse-scale scenario, and the coarse-scale stage and the review stage use the same elevation caliber to ensure the comparability of spatial accumulation and consistency judgment. "Hazardous candidate region" refers to the pixel connected region obtained by threshold segmentation and connected region extraction in a single image. "Review geographic candidate region" refers to the spatial region in the geographic coordinate system of the candidate region in the review image after projection transformation. "Cumulative occurrence count" refers to the count result of the same grid area being effectively covered by the candidate geographic region. "Cumulative identification confidence" refers to the statistical quantity obtained by accumulating the identification confidence of the candidates covering the grid area according to a preset weighting rule. "Overlap threshold" refers to the proportional threshold used to judge the degree of overlap between two geographic regions, and its overlap ratio can be calculated as the ratio of the intersection area of the two regions to the reference area. "Distance threshold" refers to the upper limit of the distance used to determine the spatial proximity of two geographic center points.
[0039] Step S100: Divide the area to be investigated into a geographic grid under a unified geographic coordinate system. Each geographic grid consists of several grid regions, each corresponding to a fixed geographic area and possessing a unique grid identifier. Control the UAV to fly along a preset coarse-array flight path, acquire coarse-array images, and record the corresponding pose information for each coarse-array image. The pose information characterizes the spatial position and attitude at the instant of image acquisition and corresponds one-to-one with the coarse-array images. To ensure the repeatability of projection and accumulation, the coarse-array images and pose information are associated using the same time reference, and the geographic grid and pose information use the same coordinate reference system or can be converted to the same coordinate reference system.
[0040] Without limiting the scope of protection of this application, the grid side length (or equivalent resolution) of the geographic grid can be set comprehensively based on the ground resolution GSD of the coarse-laid image and the UAV positioning error, the human-machine positioning error, and the minimum identifiable scale of the target hazard. Preferably, the grid side length is not less than a preset multiple of the positioning error and not less than a preset multiple of the minimum identifiable scale of the target, and can be selected within a range of 5–30m. This ensures that the same hazard target has a stable probability of repeated hits in the grid area projected in adjacent flight paths and adjacent frames. The identification confidence can be expressed using discrete numerical levels, with high / medium / low confidence levels of 0.9 / 0.6 / 0.3 or equivalent 3 / 2 / 1, respectively, to facilitate subsequent cumulative statistics. The cumulative occurrence threshold can be 2–4 times, and the cumulative identification confidence threshold can be 1.2–2.0 (taking 0–1 confidence level as an example) or an equivalent threshold (taking 1–3 confidence level as an example) to suppress occasional false detections in single frames. The preset intersection condition can be set to ensure that the area of the intersection between the candidate geographic region and the grid region is no less than 0.05–0.2 times the area of the grid region, in order to exclude invalid hits caused by boundary skimming. The area threshold for merging regions can be set to no less than N grid regions (e.g., N=3–10) or no less than 50–500m. 2 The selectable range can be determined based on the minimum identifiable hazard scale and the resolution of the coarse-grained image. The overlap threshold can be selected from 0.3 to 0.7, and the distance threshold can be selected from 0.5 to 2 times the grid side length or 5 to 30 m, and set in relation to the pose positioning error tolerance. The number of verification images required for consistency judgment can be no less than 2, and can be increased to no less than 3 images to enhance robustness when the positioning error is large or the occlusion is heavy. The return safety margin can be determined by at least one of the following: remaining battery power threshold, remaining flight time threshold, or remaining range threshold. For example, setting the remaining battery power to be no less than 20% or the remaining flight time to be no less than the estimated return time plus the safety margin.
[0041] Step S200: Hazard candidate identification is performed on the coarse-scaled image to obtain hazard candidate regions, hazard categories, and identification confidence levels. In this embodiment, hazard candidate identification is achieved based on image anomaly saliency and morphological rules: First, distortion correction and denoising are performed on the coarse-scaled image, and then anomaly maps are generated based on grayscale differences, texture differences, and edge intensity. Threshold segmentation is performed on the anomaly maps, and connected component extraction is used to obtain hazard candidate regions, where the hazard candidate regions are pixel connected regions in the coarse-scaled image. For each hazard candidate region, the area, aspect ratio, boundary tortuosity, texture uniformity, and contrast with the surrounding area are calculated, and matched with a preset hazard category discrimination rule set to determine the hazard category. The discrimination rule set is used to distinguish different hazard categories such as landslides, collapses, and debris flow channels. The rule set provides discrimination conditions corresponding to the above features for each hazard category. The identification confidence level is generated using a discrete level output method: high confidence is assigned when the hazard candidate region meets all discrimination conditions of the target hazard category, medium confidence is assigned when the key discrimination conditions are met, and low confidence is assigned in other cases. The key discrimination criteria are necessary conditions pre-marked in the rule set, used to limit feature combinations that are prone to false alarms. The hazard category discrimination rule set can be stored in a structure of "category rule group + key rule label". Each category rule group includes at least several items from area range, aspect ratio range, boundary tortuosity range, texture uniformity range, and contrast range. The key discrimination rules contain at least the necessary criteria that can distinguish this category from other categories. For example, for the collapse / landslide category, "boundary tortuosity greater than the threshold and contrast greater than the threshold" can be marked as a key rule. For the debris flow gully category, "aspect ratio greater than the threshold and edge strength continuity satisfying the threshold" can be marked as a key rule. When a candidate region satisfies all key rules but a few non-key rules are not satisfied, a medium confidence score can still be output to improve the recall rate and then handed over to subsequent spatial accumulation and verification to suppress false alarms. Based on the pose information, the hazard candidate region is projected onto the geographic grid and spatial accumulation is performed.
[0042] Specifically, based on pose information, candidate hazard regions are converted from image pixel space to candidate geographic regions in a geographic coordinate system. These candidate geographic regions are then spatially mapped to geographic grids to determine one or more grid regions covered by the candidate geographic regions. The occurrence count and cumulative identification confidence are then statistically analyzed for each covered grid region. For each grid region, the occurrence count and cumulative identification confidence are accumulated. The occurrence count represents the number of times the grid region is hit by a candidate hazard region during the coarse-sorting flight, and the cumulative identification confidence is the cumulative statistical result of the identification confidence corresponding to that grid region. To reduce statistical bias caused by candidate regions crossing grid boundaries, the cumulative identification confidence is accumulated using a weighted accumulation method based on the projected intersection area. That is, the larger the proportion of the intersection area between the grid region and the candidate geographic region, the higher the corresponding cumulative identification confidence contribution. If the identification confidence is numerical, it is accumulated for each covered grid region using Ci = Ci + wi * c, where c is the identification confidence value of the candidate region, and wi is the proportion of the intersection area between the candidate geographic region and grid region i to the total area of the candidate geographic region. If the identification confidence level is graded, the level can be mapped to a numerical value first (e.g., high / medium / low mapped to 0.9 / 0.6 / 0.3 or 3 / 2 / 1) and then accumulated as described above. To avoid incomparable accumulated values due to differences in the number of images, the average identification confidence level can be calculated simultaneously during screening, or the accumulated identification confidence level can be normalized. The accumulated identification confidence level threshold is set one-to-one with the selected caliber. After spatial accumulation is completed, a dual-threshold screening method is used to select the set of suspected hazard areas: when the cumulative occurrence frequency of a grid area reaches a preset frequency threshold and the accumulated identification confidence level reaches a preset confidence level threshold, the grid area is considered a suspected grid area. Subsequently, consecutive adjacent suspected grid areas are merged into a merged area to form a set of suspected hazard areas. The merged area is then screened using an area threshold to remove small noise areas, thus obtaining a spatially continuous set of suspected hazard areas with sufficient evidence accumulation.
[0043] Step S300: Plan a review route based on the set of suspected potential hazard areas and control the UAV to collect review images and review pose information. In this embodiment, the review route planning is constrained by a return-to-home safety margin, which is used to ensure that the UAV still has the minimum flight capability required for a safe return after the review flight. Under the premise of meeting the return-to-home safety margin, the review priority is determined based on the cumulative occurrence frequency and cumulative identification confidence of the merged areas. The review priority is sorted according to the rule of cumulative occurrence frequency first, followed by cumulative identification confidence, and a preset number of merged areas with the highest review priority are selected to enter the review route. The UAV collects review images and records review pose information along the review route, and then performs review candidate identification on the review images to obtain review geographic candidate areas. The review candidate identification is consistent with the coarse sorting stage to ensure that the criteria for coarse sorting and review are consistent. Based on the review pose information, the review candidate areas are projected onto the geographic coordinate system to obtain review geographic candidate areas that can be directly spatially matched with the suspected potential hazard areas.
[0044] Step S400: Perform consistency determination in the geographic coordinate system and output the hidden danger investigation results. In this embodiment, consistency determination includes matching the review geographic candidate area with the corresponding merged area. The matching condition must meet at least one of the following: the ratio of the overlapping area of the review geographic candidate area and the merged area to the area of the merged area reaches a preset overlap threshold, or the geographic center point of the review geographic candidate area falls into the merged area and the distance between the geographic center point of the review geographic candidate area and the geographic center point of the merged area is not greater than a preset distance threshold. The overlapping area ratio can be defined as the ratio of the intersection area A∩ of the review geographic candidate area and the merged area to the area A_merged of the merged area, A∩ / A_merged, or, when more stringent constraints are required, as the ratio of A∩ to the union area A∪ of the two areas, A∩ / A∪. This application preferably adopts A∩ / A_merged as a criterion consistent with the "degree of review coverage of the suspected area". At the same time, it is required that the same merged area forms a review geographic candidate area that meets the above conditions in no less than two review images, as objective evidence of review success. The results of the hazard investigation are output based on the consistency judgment results. The hazard investigation results include at least the location of the suspected hazard area, the hazard category, and the review conclusion. The review conclusion includes whether the review is passed or pending review, which is used to support the priority arrangement and review decision-making for on-site handling.
[0045] Understandably, compared to investigation methods relying on single-image recognition, this embodiment normalizes candidate hazard areas from the coarse screening stage to a geographic grid and performs cross-frame spatial accumulation through pose projection. It utilizes dual-threshold filtering and adjacent merging to suppress sporadic noise and isolated false detections, ensuring that the set of suspected hazard areas possesses evidence of recurrence and accumulated confidence, thereby improving the reliability of suspected areas. The review route, under the constraint of return safety margin, focuses on high-suspect areas according to review priority, reducing invalid resampling of low-risk areas. Furthermore, it forms traceable review evidence through multi-image consistency judgment in the geographic coordinate system, realizing the output from suspected alerts to review conclusions. Spatial accumulation and dual-threshold filtering provide a high-quality target set for the review route. The review route and consistency judgment further transform the target set into verifiable conclusions, improving the timeliness and accuracy of emergency investigations and reducing the cost of false positives and the spillover risk of missed reports.
[0046] In some embodiments of this application, distortion correction and denoising are performed on the coarse-formatted image. Distortion correction is used to eliminate the influence of radial and tangential distortion of the lens on the calculation of morphological features, and denoising is used to suppress false edge enhancement caused by sensor noise and compression noise. Distortion correction can be performed based on the inherent parameters of the camera, and denoising can use edge-preserving filtering methods to avoid excessive smoothing of candidate boundaries. Subsequently, an anomaly map is generated from the processed coarse-formatted image: grayscale difference, texture difference, and edge intensity are calculated separately at the same spatial scale. Grayscale difference is used to reflect the abrupt changes in brightness between the candidate region and the surrounding background, texture difference is used to reflect the abrupt changes in surface roughness or graininess, and edge intensity is used to reflect the continuity and sharpness of structural boundaries. The three are combined into an anomaly map according to a preset fusion rule, so that the value of the anomaly map can characterize the "degree of anomaly relative to the surroundings". For example, grayscale differences, texture differences, and edge intensity can be normalized to 0–1 and then weighted and summed to form an anomaly map. The weights can be set to w1, w2, and w3, satisfying w1+w2+w3=1. For example, 0.4 / 0.3 / 0.3 can be used to balance abrupt changes in brightness and boundary continuity. Thresholding can be performed using an adaptive thresholding method based on anomaly map statistics. For example, a threshold T=μ+kσ can be constructed using the anomaly map mean μ and standard deviation σ, where k can be selected from 1.0–2.0. Alternatively, the equivalent threshold can be obtained using the maximum inter-class variance method. The width of the surrounding neighborhood ring can be set to 0.5–2 times the equivalent radius of the candidate region or selected from 3–15 pixels to ensure the stability of neighborhood statistics and avoid interference from too far away regions. After the anomaly map is generated, thresholding is performed to obtain an anomaly binary map, and then connected component extraction is used to obtain potential hazard candidate regions. The threshold for threshold segmentation is used to separate background fluctuations from real anomalies. It can be set to an adaptive threshold driven by anomaly map statistics to adapt to different lighting and surface reflection conditions. Connected region extraction is used to aggregate spatially adjacent anomalous pixels into candidate regions, and can further remove obviously too small connected regions to reduce noisy candidates.
[0047] After obtaining candidate hazard regions, regional features and contrast features are calculated for each candidate region for category classification. Regional features include area, aspect ratio, and boundary tortuosity. Area reflects the candidate's scale, aspect ratio reflects its elongated or blocky shape, and boundary tortuosity reflects whether the boundary is irregularly fragmented. Texture features include texture uniformity, reflecting whether the texture within the candidate region is homogeneous or banded. Contrast features include contrast with surrounding areas, reflecting the intensity of the difference in grayscale or texture between the candidate region and its extended neighborhood. In this embodiment, the surrounding neighborhood is determined by "forming a ring around the candidate region," with the ring width ensuring statistical stability of the neighborhood and avoiding interference from distant objects. Subsequently, the above features are matched with a preset hazard category classification rule set to determine the hazard category. The discrimination rule set consists of rule groups for multiple hazard categories. Each rule group comprises several discrimination rules, which use the range of feature values or feature relationships as judgment conditions. For example, area and boundary tortuosity are used to distinguish between "blocky damage" and "linear gully damage," and texture uniformity and contrast are used to distinguish between "fresh bare body" and "vegetation cover." To improve the interpretability and maintainability of rule matching, the rule set further labels key discrimination rules. These key discrimination rules are used to characterize the necessary criteria that distinguish this hazard category from other categories and to identify the confidence level classification.
[0048] After determining the hazard category, an identification confidence score is generated. This embodiment uses a tiered confidence score output: for the same hazard candidate region, the number of matching discrimination rules and the satisfaction status of key discrimination rules are statistically analyzed. When the hazard candidate region satisfies all discrimination rules of the rule group corresponding to the hazard category, a preset high confidence score is assigned. When the hazard candidate region satisfies all key discrimination rules of the rule group but not all non-key discrimination rules, a preset medium confidence score is assigned. In other cases, a preset low confidence score is assigned. This tiered mechanism directly correlates the identification confidence score with the "degree of rule compliance," facilitating objective cumulative statistics and regional screening in spatial accumulation and dual-threshold screening, and reducing the risk of category misjudgment due to fluctuations in a single feature.
[0049] Understandably, the process of anomaly graphs, connected regions, rule matching, and graded confidence levels enables the coarse-sorting stage to output interpretable candidate hazard regions, hazard categories, and identification confidence levels with low computational complexity. This provides a consistent and traceable data foundation for spatial accumulation and dual-threshold screening based on pose projection. Graded confidence levels explicitly quantify the uncertainty of candidate identification by the degree of rule conformity, which is beneficial for achieving the effect of "high-evidence candidates are more likely to form suspected areas, and low-evidence candidates are more likely to be diluted" during spatial accumulation, thereby reducing the spread of false alarm candidates in grid accumulation. This embodiment, on the one hand, rapidly produces structured candidates in a low-cost manner, providing input for spatial accumulation screening of suspected hazard region sets. On the other hand, through rule-based category discrimination and confidence leveling, it improves the category consistency and evidence strength of the suspected hazard region set, thereby indirectly improving the efficiency of route focusing and the stability of geographical consistency judgment, and enhancing the timeliness and accuracy of closed-loop investigation.
[0050] In some embodiments of this application, a geographic grid is established and a unique grid identifier is assigned to each grid region. Simultaneously, two statistical quantities—cumulative occurrence count and cumulative identification confidence—are initialized for each grid region. For any coarse-array image, the candidate hazard region in the coarse-array image is projected onto a geographic coordinate system based on the pose information corresponding to the coarse-array image, obtaining a candidate geographic region. The candidate geographic region is spatially mapped to the geographic grid to determine one or more grid regions covered by the candidate geographic region. For each covered grid region, an occurrence count is performed: when the intersection area of the projection of the candidate geographic region and the grid region reaches a preset intersection condition, the cumulative occurrence count of the grid region is increased by one. The preset intersection condition is used to exclude accidental hits caused by boundary crossings, ensuring that the count reflects effective coverage.
[0051] When accumulating the identification confidence score, this embodiment adopts a weighted accumulation method based on the projected intersection area to address the evidence allocation problem when candidate geographic regions are distributed across grids. Specifically, for each grid region covered by a candidate geographic region, the projected intersection area between the grid region and the candidate geographic region is calculated. The proportion of this projected intersection area in the total area of the candidate geographic region is used as a weight to allocate the identification confidence score corresponding to the candidate region to the cumulative identification confidence score of each grid region. When the same grid region is hit in multiple coarse-sorted images, the corresponding weighted cumulative value is continuously accumulated over that grid region, thereby forming a cumulative identification confidence score that reflects the "degree of repetition across multiple frames and the strength of evidence". To ensure consistency in the screening criteria, this embodiment forms a set of statistics for each grid region after accumulation, including at least the cumulative occurrence count and cumulative identification confidence score. An average identification confidence score can be further generated as an equivalent statistical metric. The average identification confidence score is used to eliminate the influence of differences in the number of images on the cumulative value.
[0052] After completing the above spatial accumulation, a set of suspected potential hazard areas is generated. In this embodiment, each grid area is first filtered using a dual threshold based on a preset frequency threshold and a preset confidence threshold to obtain suspected grid areas. Then, adjacent merging is performed on the suspected grid areas to form a set of suspected potential hazard areas. The adjacent merging is based on the grid topology, merging suspected grid areas with adjacent boundaries or corners into the same suspected potential hazard area. To avoid incorrectly connecting spatially discrete suspected grid areas, this embodiment can use connectivity constraints during merging, that is, merging is only allowed when there are continuous adjacent links between suspected grid areas, thus making the set of suspected potential hazard areas spatially continuous and interpretable. After the set of suspected potential hazard areas is generated, the geographical boundary of each suspected potential hazard area can be represented as the outer envelope of the grid areas it contains, and the cumulative occurrence frequency and cumulative identification confidence value of the corresponding suspected potential hazard area are recorded for use in review priority ranking and review route planning.
[0053] Understandably, this embodiment unifies the multi-frame identification results of potential hazard candidate areas onto a geographic grid through pose projection, and constructs a spatial evidence accumulation mechanism using "occurrence count" and "weighted identification confidence accumulation based on the intersection area of projections." This effectively suppresses the spatial spread of occasional false detections in a single frame, while avoiding evidence distortion caused by cross-grid segmentation of candidate areas, making the screening of suspected grid areas more stable. Furthermore, by merging consecutive adjacent suspected grid areas, the set of suspected hazard areas is elevated from discrete grids to spatially continuous objects, facilitating targeted planning and resource focus for route review. Spatial accumulation and adjacent merging provide a high-quality, sortable set of suspected hazard areas, providing clear target input for route review planning. The weighted accumulation mechanism improves the interpretability of the evidence strength of suspected areas, making it easier to form stable review conclusions based on geographic coordinate system consistency judgments, thereby improving the overall timeliness and reliability of emergency investigation results.
[0054] In some embodiments of this application, after merging adjacent suspected grid areas, the area of each merged area is determined as the sum of the areas of the grid areas contained within it, or as the area of the geographic boundary envelope of the merged area. The area of each grid area is determined by the grid resolution of the geographic grid in the geographic coordinate system, and the envelope area reflects the overall footprint of the merged area. Subsequently, area threshold filtering is performed: the area of the merged area is compared with a preset area threshold, and merged areas with areas smaller than the preset area threshold are removed. The preset area threshold is used to exclude small-scale merged areas caused by isolated false detections, boundary noise, or sporadic anomalies, and can be set according to the ground resolution of the coarse-formatted image and the minimum identifiable scale of the target hazard, thereby ensuring that the retained merged areas have engineering-significant verification value. The merged areas after area threshold filtering constitute a set of screened suspected hazard areas, and the cumulative occurrence frequency and cumulative identification confidence of each merged area are retained for verification priority calculation.
[0055] During the route planning review phase, the return-to-home safety margin is first obtained and used as a constraint for the review flight. The return-to-home safety margin characterizes the safe return capability that the UAV must retain after completing the review flight. It can be determined by at least one of the following: remaining battery power threshold, remaining flight time threshold, or remaining range threshold, and is correlated with the UAV's current state and the estimated return-to-home path. Specifically, the return-to-home safety margin can be defined as the lower limit of the difference between "resources required to complete the return" and "actual remaining resources," thereby ensuring that the review route planning does not encroach on necessary return-to-home resources. Subsequently, the review priority is determined based on the cumulative occurrence frequency and cumulative identification confidence of the merged area: In this embodiment, the cumulative occurrence frequency is prioritized to reflect the "strength of evidence repeated across multiple frames," and the cumulative identification confidence is used to reflect the "quality of evidence across single frames and cross-frames." When the cumulative occurrence frequencies are the same, the cumulative identification confidence is used for ranking. After the review priorities are determined, merging regions are selected sequentially from highest to lowest priority to enter the review task set, until the expected flight resource consumption corresponding to the review task set reaches the return safety margin constraint boundary. This ensures that at least one merging region is included in the review flight and the overall review flight can be completed safely. Finally, a review flight path is generated based on the review task set. The review flight path must at least cover the geographical boundary range of the selected merging regions. During the execution of the review flight, review images and review pose information corresponding to the merging regions are collected to provide a data foundation for consistency determination.
[0056] Understandably, this embodiment, by applying area threshold screening to the merged area, further constrains the set of suspected potential hazard areas from "spatial objects whose accumulated evidence meets the threshold" to "spatial objects whose scale reaches a verifiable level." This reduces invalid supplementary sampling caused by small-scale noise areas entering the verification process, improving the target quality and utilization efficiency of the verification route. Simultaneously, by introducing a return-to-base safety margin constraint and combining the cumulative occurrence frequency and cumulative identification confidence to determine the verification priority, priority verification of highly suspected areas is achieved under the premise of safe accessibility. This avoids the verification task affecting return-to-base safety and maximizes the verification value within limited flight time. Area threshold screening improves the "verifiability" of the suspected area set, while verification priority and return-to-base safety margin constraints improve the "executability and focus" of the verification route. Together, they promote the scheme to concentrate collection resources on high-value suspected areas and form stable verification evidence, further improving the timeliness and reliability of emergency investigations.
[0057] In some embodiments of this application, based on the cumulative occurrence count and cumulative identification confidence of each merged region, the merged regions are prioritized for review: the cumulative occurrence count is used as the first ranking factor, the higher the cumulative occurrence count, the more sufficient the evidence that the merged region was repeatedly hit in the coarse-sorting stage, and the higher the review priority. When the cumulative occurrence counts are the same, the cumulative identification confidence is used as the second ranking factor, the higher the cumulative identification confidence, the higher the candidate identification quality of the merged region, and the higher the review priority. Subsequently, under the premise of satisfying the return safety margin constraint, the set of merged regions entering the review route is determined: a preset number is set to limit the target number upper limit of a single review flight, and the preset number can be set according to the available review flight time of the UAV and the review collection cost of a single merged region. Merged regions are selected in descending order of review priority until the selected number reaches the preset number or the selection will lead to insufficient return safety margin, thereby ensuring that high-priority merged regions enter the review route under safe and reachable conditions.
[0058] After the verification route is completed, a consistency determination is performed and a verification conclusion is output. In this embodiment, verification candidate identification is performed on each verification image to obtain verification geographic candidate regions. The verification geographic candidate regions are spatial regions projected onto the geographic coordinate system based on the verification pose information. For each merged region, all verification geographic candidate regions spatially adjacent to the merged region in all verification images are collected, and the matching relationship between the "verification geographic candidate region and the merged region" is determined. The matching determination is based on the spatial relationship under the geographic coordinate system, including at least the determination criteria of overlap or inclusion relationship. To avoid misjudgment due to occasional false detections of a single verification image, this embodiment requires that the same merged region has a matching verification geographic candidate region in no less than two verification images before outputting the verification pass conclusion. When the above matching condition is met, the verification pass conclusion is output; when not met, a pending verification conclusion is output, and the corresponding merged region is entered into the manual verification or re-verification process as a pending verification object. The final output of the hazard investigation results corresponds one-to-one with the merged areas, including the geographical location of the merged area, the hazard category, and the review conclusion, which is used to support the priority and review decision-making for emergency response.
[0059] Understandably, by clearly defining the review priority as a two-factor ranking rule of "cumulative occurrence frequency first, followed by cumulative identification confidence," the selection of review targets has a stable basis, thereby avoiding arbitrariness in the selection of review route targets. Furthermore, through the joint constraints of preset quantities and return-to-base safety margins, review resources are concentrated in the merging areas with the most accumulated evidence, improving the output per unit flight time for review routes, while ensuring flight safety. Simultaneously, consistency judgment adopts matching relationships under a geographic coordinate system and introduces the evidence requirement of no less than two review images, transforming the review conclusion from a single-frame judgment to a multi-image consistency verification result, reducing the risk of occasional false detections during the review stage. Review priority ranking and safety constraints ensure that "what to review and what to do first" is executable and efficient, while multi-image matching judgment ensures that "how the review conclusion is formed" is objective and traceable. Together, these enhance the timeliness, reliability, and verifiability of the overall closed-loop investigation plan in emergency scenarios.
[0060] In some embodiments of this application, both the merged region and the verification geographic candidate region are represented as planar spatial regions in a geographic coordinate system. The merged region is determined by the outer envelope or boundary point sequence of the grid regions it contains, while the verification geographic candidate region is determined by the boundary point sequence obtained by combining the verification candidate recognition result with the verification pose information projection. To ensure consistency in the calculation of area and distance, the two types of regions use the same area calculation method and the same distance measurement method in the same geographic coordinate system. Subsequently, based on the overlap area ratio criterion, the spatial intersection area between the verification geographic candidate region and the corresponding merged region is calculated, and the area of the intersection area and the area of the merged region are calculated separately. The ratio of the intersection area area to the area of the merged region is compared with a preset overlap threshold. When the ratio is not less than the preset overlap threshold, a match is determined. The preset overlap threshold is used to constrain "the degree of overlap must reach a verifiable coverage ratio." Its value can be set according to the scale of the merged region, the resolution of the geographic grid, and the positioning error tolerance of the verification image. The optional range or setting principle is given in the specification in the implementation method to avoid the threshold being too strict, leading to misjudgment as pending verification, or the threshold being too wide, leading to false positives.
[0061] In another criterion, a combination of geographic center point inclusion and center point distance constraints is used. Specifically, the geographic center points of the candidate geographic regions and the merged region are determined separately. The geographic center point can be determined using the region's geometric center or centroid. It is then determined whether the geographic center point of the candidate geographic region falls within the boundary of the merged region. If it does, the geographic distance between the two center points is calculated and compared with a preset distance threshold. A match is considered successful when the geographic distance is not greater than the preset distance threshold. The preset distance threshold constrains that "the deviation between the location center of the candidate geographic region and the suspected region's center should not exceed an acceptable deviation." Its setting can be related to the grid edge length of the geographic grid, the ground resolution of the verification image, and the pose positioning error, thereby ensuring that the criterion can tolerate small-scale positioning drift while excluding occasional false detections far from the suspected region. To improve the robustness of the matching determination, this embodiment allows a match to be considered successful if either criterion is met. However, in the consistency determination, at least two verification images are still required to meet the matching requirement to avoid misjudgments caused by occasional matching of a single image.
[0062] Understandably, this embodiment specifies "matching" into two conditions: overlapping area ratio criterion and center point constraint criterion, reducing the uncertainty caused by subjective human judgment. The overlapping area ratio criterion directly reflects the spatial coverage of the suspected area by the verification evidence, and is suitable for scenarios where the merged area has a relatively regular shape or high positioning accuracy. The center point constraint criterion is more tolerant of scenarios with irregular shapes or certain boundary drift, and can reduce false rejections caused by slight boundary offsets while maintaining positioning constraints. By placing the core link of consistency determination on geospatial calculation, the objectivity and stability of the verification conclusion are improved, making the overall closed-loop solution more likely to generate reliable and verifiable output results in emergency investigation scenarios, and further reducing the probability of false positives and missed verifications.
[0063] In summary, this application constructs a geographic grid and projects candidate hazard areas from the coarse-grained stage, combined with pose information, onto a unified spatial reference. This achieves evidence accumulation and dual-threshold screening across multiple frames and viewpoints, suppressing the spread of false alarms caused by single-frame false detections. It also elevates suspected hazard areas from single-identification results to spatial objects with recurring occurrences and accumulated confidence support. For sets of suspected hazard areas, a review route is automatically planned and review images are acquired, shifting limited flight time and acquisition bandwidth from low-risk areas to high-suspect areas, reducing invalid coverage and redundant processing. In the review stage, candidate review areas are projected again onto a geographic coordinate system, and the geographic overlap of at least two review images is used as the consistency criterion. This transforms the review conclusion from subjective judgment to objective evidence, thereby shortening the link time from acquisition to conclusion output without relying on panoramic stitching and global modeling. This improves the timeliness of emergency investigations and generates verifiable output within the same task, reducing the risk spillover caused by missed detections and lowering the review costs associated with false alarms.
[0064] Based on another preferred embodiment described above, see [link to preferred embodiment]. Figure 2 As shown, this embodiment provides an automatic geological hazard hazard investigation system based on UAV imagery, used to apply the aforementioned automatic geological hazard hazard investigation method based on UAV imagery, including:
[0065] The data acquisition unit is configured to construct a geographic grid of the area to be investigated, control the drone to perform coarse-scale flight, acquire coarse-scale images, and obtain pose information.
[0066] The identification unit is configured to identify potential hazards in coarse-grained images, obtaining candidate hazard regions, hazard categories, and identification confidence levels. Based on pose information, the candidate hazard regions are projected onto a geographic grid and spatially accumulated. A set of suspected hazard regions is then selected based on accumulated occurrence thresholds and accumulated identification confidence thresholds.
[0067] The re-collection unit is configured to plan a re-examination route based on the set of suspected hidden danger areas and control the UAV to collect re-examination images and re-examination pose information, identify re-examination candidates for the re-examination images, and project the re-examination pose information onto the geographic coordinate system to obtain the re-examination geographic candidate area.
[0068] The judgment unit is configured to perform consistency judgment in a geographic coordinate system, determining whether the same suspected hazard area forms a verification geographic candidate area that overlaps in no fewer than two verification images and reaches a preset overlap threshold. Based on the consistency judgment result, the hazard investigation result is output, including the location of the suspected hazard area, the hazard category, and the verification conclusion.
[0069] It is understandable that the aforementioned method and system for automatically identifying potential geological hazards using drone imagery have the same beneficial effects, and will not be elaborated upon here.
[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for automatically investigating potential geological hazards using unmanned aerial vehicle (UAV) imagery, characterized in that, include: Construct a geographic grid for the area to be investigated, control the drone to perform coarse-scale flight, collect coarse-scale images and obtain pose information; The coarse-arranged image is used to identify potential hazards, thereby obtaining candidate hazard regions, hazard categories, and identification confidence levels. The candidate hazard regions are then projected onto a geographic grid based on the pose information and spatially accumulated. A set of suspected hazard regions is then selected based on the accumulated occurrence frequency threshold and the accumulated identification confidence level threshold. Based on the set of suspected potential hazard areas, a review route is planned and a drone is controlled to collect review images and review pose information. Review candidate identification is performed on the review images, and the review pose information is projected onto the geographic coordinate system to obtain the review geographic candidate area. Under the geographic coordinate system, a consistency determination is made to determine whether the same suspected hidden danger area forms an overlapping geographic candidate area in no less than two review images that reaches a preset overlap threshold. The hazard investigation results are output based on the consistency judgment results. The hazard investigation results include the location of the suspected hazard area, the hazard category, and the review conclusion.
2. The method for automatic investigation of geological disaster hazards using UAV imagery according to claim 1, characterized in that, Distortion correction and denoising are performed on the coarse-arrayed image; anomaly maps are generated based on the grayscale differences, texture differences, and edge intensity of the coarse-arrayed image; threshold segmentation is performed on the anomaly maps and candidate regions for potential hazards are obtained through connected component extraction; For each candidate hazard region, calculate its area, aspect ratio, boundary curvature, texture uniformity, and contrast with its surroundings, and match it with a preset hazard category discrimination rule set. Determine the hazard category based on the matching results. When all discrimination rules are met, a preset high confidence value is used as the recognition confidence value; when the key discrimination rules are met, a preset medium confidence value is used as the recognition confidence value; and for the rest, a preset low confidence value is used as the recognition confidence value.
3. The method for automatic investigation of geological disaster hazards using UAV imagery according to claim 2, characterized in that, The spatial accumulation includes: mapping the projection results of all the coarse-arranged images to the same grid area and counting the number of occurrences, and accumulating the identification confidence; the set of suspected hidden danger areas is obtained by merging consecutively adjacent grid areas.
4. The method for automatic investigation of geological disaster hazards using UAV imagery according to claim 3, characterized in that, When accumulating and statistically analyzing the identification confidence scores, the identification confidence scores mapped to the same grid area are weighted and accumulated, with the weights determined based on the intersection area of the projections of the candidate hazard area and the grid area.
5. The method for automatic investigation of geological disaster hazards using UAV imagery according to claim 3, characterized in that, When screening a set of suspected potential hazard areas, the method further includes: after merging the consecutive adjacent grid areas, performing a threshold screening on the area of the merged area to remove merged areas with an area smaller than a preset area threshold.
6. The method for automatic investigation of geological disaster hazards using UAV imagery according to claim 5, characterized in that, When planning the review route, the process includes: obtaining a return safety margin, determining the review priority based on the return safety margin and the cumulative occurrence and cumulative identification confidence of the merged area, and selecting at least one merged area for review flight based on the review priority.
7. The method for automatic investigation of geological disaster hazards using UAV imagery according to claim 6, characterized in that, The review priority is determined according to the rule of prioritizing the cumulative occurrence frequency and then the cumulative identification confidence level. Under the premise of meeting the return safety margin, a preset number of merged areas with the highest review priority are selected to enter the review route.
8. The method for automatic investigation of geological disaster hazards using UAV imagery according to claim 6, characterized in that, The consistency determination includes: matching the candidate geographic regions for verification of all the verification images with the merged region; outputting a verification pass conclusion when the match is successful, and outputting a pending verification conclusion when the match is unsuccessful.
9. The method for automatic investigation of geological disaster hazards using UAV imagery according to claim 8, characterized in that, The matching is valid if at least one of the following conditions is met: The ratio of the overlapping area of the verification geographic candidate region and the corresponding merged region to the area of the merged region reaches the preset overlap threshold, or the geographic center point of the verification geographic candidate region falls into the merged region and the distance between the geographic center point of the merged region and the geographic center point of the merged region is not greater than the preset distance threshold.
10. An automatic geological hazard hazard investigation system based on UAV imagery, used to apply the automatic geological hazard hazard investigation method based on UAV imagery as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is configured to construct a geographic grid of the area to be investigated, control the drone to perform coarse-scale flight, acquire coarse-scale images, and obtain pose information. The identification unit is configured to perform hazard candidate identification on the coarse-grained image, obtain hazard candidate areas, hazard categories and identification confidence levels; project the hazard candidate areas onto a geographic grid according to the pose information and perform spatial accumulation; and filter a set of suspected hazard areas according to a cumulative occurrence threshold and a cumulative identification confidence level threshold. The re-collection unit is configured to plan a re-examination route based on the set of suspected hidden danger areas and control the UAV to collect re-examination images and re-examination pose information, perform re-examination candidate identification on the re-examination images, and project the re-examination pose information onto the geographic coordinate system to obtain the re-examination geographic candidate area. The determination unit is configured to perform consistency determination in the geographic coordinate system and determine whether the same suspected hidden danger area forms a verification geographic candidate area that overlaps in no less than two verification images and reaches a preset overlap threshold. The hazard investigation results are output based on the consistency judgment results. The hazard investigation results include the location of the suspected hazard area, the hazard category, and the review conclusion.