A landslide geological disaster multi-precision real-time detection method and system based on geographic information constraint and visible light-laser radar fusion

Abstract

The application discloses a kind of based on geographic information constraint and visible light-laser radar fusion's landslide geological disaster multi-precision real-time detection method and system.The method first obtains the visible light image, laser radar point cloud, flight pose and sensor calibration parameter collected synchronously by unmanned aerial vehicle, and generates geographic prior field in combination with the information of topography, geological structure and disaster-bearing body in target area.Subsequently, based on pose and calibration parameter, initial mapping of image pixel and laser point is established, local registration correction is carried out using stable terrain features, and registration reliability is calculated, image texture features, point cloud geometric features and geographic prior features are extracted, adaptive fusion is realized by bidirectional projection and reliability gating.According to slope direction and landslide structure topology, the detection result is checked, and the disaster position level and unmanned aerial vehicle recheck flight path are output.The application can consider wide-range screening speed and local fine detection accuracy under limited on-board computing power, effectively reduce the false detection caused by vegetation, shadow and the like.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring, specifically to a multi-precision real-time detection method and system for landslide geological disasters based on geographic information constraints and visible light-LiDAR fusion. Background Technology

[0002] Geological hazards such as landslides, collapses, rockfalls, and debris flows are characterized by their concealment, suddenness, and uneven spatial distribution. Dangerous slopes along mountain roads, mine slopes, reservoir banks, and around residential areas typically cover large areas, have significant topographical undulations, and complex transportation conditions. Manual reconnaissance is insufficient to simultaneously meet the requirements for coverage, inspection frequency, and personnel safety. Unmanned aerial vehicles (UAVs) can acquire high-resolution visible light imagery and 3D laser point clouds in a short time, making them suitable for slope hazard investigation and emergency detection.

[0003] Visible light imagery contains rich information on color, texture, and edges, reflecting surface features such as bare soil, vegetation destruction, cracks, soil-rock boundaries, and landslide deposits. However, relying solely on visible light imagery is susceptible to variations in lighting, shadows, haze, seasonal vegetation changes, and the similarity of surface colors. For example, normal road excavations, bare quarry land, or cultivated areas may exhibit colors and textures similar to landslides; conversely, landslides covered by vegetation may lack distinct image textures.

[0004] LiDAR point clouds can directly describe surface elevation, slope, normal vector, curvature, roughness, and local elevation differences. They can also penetrate vegetation to a certain extent to obtain ground echoes, making them suitable for detecting steep slopes at the rear, bulges at the front, crack drops, and surface deformation. However, because point clouds lack continuous color texture and are affected by scanning angle, effective range, flight speed, and occlusion, the point density in different areas may vary significantly. Direct segmentation of sparse point clouds can easily miss small-scale cracks or create discontinuous disaster boundaries.

[0005] Existing multi-source detection schemes typically convert images and point clouds to the same raster and then directly stitch them together, or detect them separately and then vote on the results. Due to differences in the imaging mechanisms, spatial resolution, and sampling times of cameras and LiDAR, sensor extrinsic parameter errors, high-speed movement of UAVs, and time synchronization deviations can cause misalignment between 2D images and 3D point clouds. Direct stitching may incorrectly assign texture features from one location to a LiDAR point at another location, thus amplifying false detections.

[0006] On the other hand, the target inspection area usually has usable geographical priors. For example, slope, aspect, lithology, faults, river erosion, road excavation, historical landslides, and continuous rainfall are related to the probability of disaster occurrence; landslides also typically exhibit spatially related structures such as trailing edge cracks, main landslide walls, the landslide body, leading edge bulging, and depositional zones. Existing methods often only overlay geographical locations after detection, without directly involving geographical information in data sampling, feature fusion, inference accuracy selection, and result verification.

[0007] Furthermore, the computing units on the UAV side are limited by weight, power consumption, heat dissipation, and battery capacity. Continuously running highly complex neural networks on the entire high-resolution image and all high-density point clouds would significantly increase inference latency and shorten flight time; conversely, using lightweight models throughout would make it difficult to identify small cracks, local steep slopes, and slow deformations. Therefore, a real-time processing mechanism is needed that can quickly screen a large area first, and then improve the inference accuracy of local areas based on risk and uncertainty.

[0008] Therefore, it is necessary to provide a new technical solution so that geographic information is no longer just a result display information, but forms a geographic prior field that can participate in two-dimensional-three-dimensional registration, cross-modal fusion, point cloud sampling and multi-precision inference scheduling; at the same time, different computing paths can be selected according to modal reliability, temporal changes and edge resource status to balance detection accuracy, real-time performance and UAV endurance. Summary of the Invention

[0009] To address the problems of existing technologies, this invention provides a multi-precision real-time detection method and system for landslide geological hazards based on geographic information constraints and visible light-LiDAR fusion. This solves the problems of easy misalignment between two-dimensional images and three-dimensional point clouds, inability to dynamically evaluate the reliability of different modalities, lack of geographic information participation in the reasoning process, difficulty in real-time operation of high-precision calculations across the entire area, and difficulty in distinguishing between actual deformation and apparent surface changes in single-phase detection.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A multi-precision real-time detection method for landslide geological hazards based on geographic information constraints and visible light-lidar fusion includes the following steps: Acquire visible light images and lidar point clouds of the target inspection area simultaneously collected by the UAV, as well as the corresponding flight position data, flight attitude data, timestamps and sensor calibration parameters at the time of collection; The geographic information of the target inspection area is obtained, and the geographic information is converted into geographic prior features corresponding to the spatial units of the target inspection area. The geographic prior features include at least three of the following: topographic features, geological features, historical disaster features, environmental inducing factors, disaster-bearing body risk features, and expected sliding direction. Motion distortion compensation and surface category processing are performed on the lidar point cloud, distortion correction is performed on the visible light image, and an initial two-dimensional-three-dimensional mapping relationship between image pixels and lidar points is established based on the flight position data, flight attitude data and sensor calibration parameters. The initial two-dimensional-three-dimensional mapping relationship is locally registered and corrected using stable terrain features in the target inspection area to obtain the two-dimensional-three-dimensional mapping relationship and the local registration reliability of each spatial unit. The image features of the visible light image, the point cloud geometric features of the lidar point cloud, and the geographical prior features are extracted respectively. Based on the two-dimensional-three-dimensional mapping relationship, the image features are projected to the corresponding lidar points and the point cloud geometric features are projected to the corresponding image feature positions. The fusion weight is determined according to image quality, point cloud density, local registration reliability, and geographical risk level to obtain two-dimensional-three-dimensional fused features. The first precision reasoning path is used to perform global reasoning on the two-dimensional-three-dimensional fused features to obtain the suspected disaster spatial unit and its initial detection confidence and prediction uncertainty; Based on at least four of the following factors: the geographical risk level of the suspected disaster spatial unit, prediction uncertainty, modal difference between visible light detection results and lidar detection results, surface change relative to historical inspection data, data quality, and UAV end-side computing resource status, the target inference path is determined from multiple candidate inference paths with different accuracy and computational load. Using the target reasoning path, at least one of the following processes is performed on the suspected disaster spatial unit: landslide area segmentation, landslide boundary identification, surface crack identification, terrain abrupt change detection, and multi-temporal surface deformation calculation, to obtain disaster detection results; Based on the aforementioned geographical prior features, slope orientation, and landslide disaster structure topology, the disaster detection results are spatially consistent and the following outputs are provided: disaster category, disaster boundary, surface displacement, elevation change, volume change, risk level, geographical location, and re-inspection instruction.

[0011] The geographic information includes at least four of the following: digital elevation model or digital terrain model, slope, aspect, plane curvature, profile curvature, topographic relief, soil and rock type, fault location, distance from river, distance from road or engineering excavation area, historical disaster boundary, historical deformation record, rainfall, soil moisture content, seismic motion parameters, and distribution of settlements or roads. This geographic information, after coordinate unification, scale normalization, and reliability labeling, is encoded into a geographic risk prior map, a predicted slip direction field, and an inference priority map. Furthermore, non-uniform data processing is applied to the target inspection area based on the geographic risk level: high-risk spatial units are processed with smaller point cloud voxel sizes, higher point cloud retention ratios, and higher image cropping resolutions; low-risk spatial units are processed with larger point cloud voxel sizes, lower point cloud retention ratios, and lower image resolutions.

[0012] The stable terrain features include at least one of ridgelines, valley lines, road edges, building outlines, exposed rock edges, and terrain fault lines; the local registration correction is achieved by minimizing the reprojection residual between the image edge and the point cloud projection edge, and limiting the corrected pose deviation to no more than a preset range; the local registration reliability is determined based on the reprojection residual, effective laser point density, consistency between the image edge and the point cloud projection edge, time synchronization error, and local occlusion ratio; when the local registration reliability is lower than the first reliability threshold, the pixel-level fusion weight is reduced and feature-level or decision-level fusion is adopted.

[0013] The image features include at least two of the following: color, texture, crack edges, bare soil areas, vegetation destruction, and soil-rock boundaries; the point cloud geometric features include at least three of the following: elevation, relative elevation, slope, normal vector, curvature, roughness, local elevation difference, echo intensity, and point density; the fusion weights are also determined based on the degree of shadow or overexposure, the degree of vegetation cover, and the integrity of each modal feature.

[0014] The multiple candidate inference paths include at least: a first-precision inference path for global screening, a second-precision inference path for detecting suspected area boundaries and disaster structures, and a third-precision inference path for multi-temporal deformation analysis; the first-precision inference path, the second-precision inference path, and the third-precision inference path differ in at least three of the following: input image resolution, point cloud sampling density, voxel size, network depth, number of feature channels, number of detector heads, numerical quantization bit width, whether instance segmentation is performed, and whether historical point clouds are invoked; the inference scheduler determines the target inference path based on priority parameters, which are obtained by weighting geographical risk level, prediction uncertainty, modal difference, surface change, data quality, computing resource consumption, remaining power, and allowable inference latency; the third-precision inference path is selected when the priority parameter is higher than a first threshold, the second-precision inference path is selected when the priority parameter is between the second and first thresholds, and the first-precision inference path is maintained in other cases.

[0015] The modal differences are determined based on the disaster probability output by the visible light image branch, the disaster probability output by the lidar point cloud branch, the degree of boundary overlap, and the consistency of displacement direction. When the modal differences exceed the difference threshold, the cause of the differences is identified as at least one of vegetation occlusion, image shadow, sparse point cloud, or local registration anomaly, and the point cloud branch weight, image branch weight, or local registration is re-executed accordingly.

[0016] The multi-temporal surface deformation calculation includes: extracting stable regions from historical and current point clouds; eliminating overall translational and rotational deviations between different inspection phases based on stable regions; calculating local point-to-surface distances or three-dimensional displacement vectors of the current point cloud relative to historical point clouds; eliminating isolated change points based on displacement continuity and confidence intervals; obtaining at least one of crack propagation, boundary migration, surface displacement, deformation rate, and landslide volume change; the landslide disaster structure topology includes the spatial adjacency relationships between the rear edge crack zone, main landslide wall, landslide body, lateral edge shear zone, front edge bulging zone, and deposition zone; the spatial consistency verification includes determining whether the displacement direction of the detection area and the local slope direction satisfy a preset angle relationship, whether the deformation area is spatially continuous, and whether each disaster structure area satisfies a preset adjacency relationship.

[0017] When the point cloud density of a suspected disaster spatial unit is lower than the point density threshold, the occlusion ratio is higher than the occlusion threshold, the registration confidence is lower than the second confidence threshold, or the disaster risk is higher than the re-inspection threshold, a UAV re-inspection trajectory is generated based on the slope normal vector, the suspected disaster boundary, the current occlusion area, and the effective range of the sensor. The re-inspection trajectory includes re-inspection waypoints, flight altitude, flight speed, heading angle, and lidar scanning angle. In the case of missing visible light images, missing lidar point clouds, or communication interruption, the system switches to a single-modal degraded inference path based on the modal availability marker and reduces the confidence level of the corresponding detection results. After recovering the dual-modal data, a supplementary re-inspection is performed on the high-risk spatial units during the degraded inference period.

[0018] When training the neural network to generate disaster detection results, at least four loss functions are used, including classification loss, region segmentation loss, boundary loss, 2D-3D feature consistency loss, geographic prior constraint loss, modal reliability loss, and multi-precision distillation loss. The training samples include visible light images, laser point clouds, spatial registration relationships, geographic prior labels, disaster category labels, disaster boundary labels, and deformation labels.

[0019] The aforementioned multi-precision real-time detection system for landslide geological hazards based on geographic information constraints and visible light-lidar fusion includes: The data acquisition module is used to acquire visible light images, lidar point clouds, flight position data, flight attitude data, timestamps, and sensor calibration parameters synchronously collected by the UAV. The geographic prior module is used to generate geographic prior features based on the topography, geology, historical disasters, environmental triggering factors, and disaster-bearing body information of the target inspection area; The preprocessing and registration module is used to perform motion distortion compensation on lidar point clouds, distortion correction on visible light images, establish and correct the two-dimensional-three-dimensional mapping relationship between image pixels and lidar points, and calculate the local registration reliability. The feature fusion module is used to extract image features, point cloud geometric features, and geographic prior features, perform bidirectional projection of image features and point cloud geometric features, and generate two-dimensional-three-dimensional fused features based on data quality, local registration reliability, and geographic risk level. The multi-precision inference module includes multiple inference paths with different precision and computational load, as well as an inference scheduler. It is used to first perform a global screening, and then select the target inference path based on geographical risk, prediction uncertainty, modal differences, multi-temporal changes, data quality, and edge resource status. The verification output module is used to verify the disaster detection results based on the slope direction and landslide disaster structure topology, and output the disaster category, boundary, deformation parameters, risk level, geographical location or re-inspection instruction; The flight feedback module, communicating with the UAV flight controller, is used to adjust at least two of the UAV's waypoints, flight altitude, flight speed, heading angle, camera shooting direction, and lidar scanning direction according to the re-inspection instructions generated by the verification output module. Specifically, it determines whether to generate a re-inspection flight track based on point cloud density, occlusion ratio, registration reliability, or disaster risk.

[0020] Compared with the prior art, the beneficial effects of the invention are: 1. The invention converts information on topography, geology, historical disasters, environmental factors and disaster-bearing bodies into a geographic prior field that can be directly used in calculations, and enables this prior field to participate in candidate region screening, data sampling, feature fusion, accuracy scheduling and result verification at the same time, which can reduce false detections caused by relying solely on apparent texture.

[0021] 2. This invention avoids unconditional strong fusion when 2D images and 3D point clouds are misaligned by initial projection of sensor pose, local correction of stable terrain, and evaluation of local registration credibility; it also improves the stability of cross-modal fusion by switching between pixel-level, feature-level, and decision-level fusion based on registration credibility.

[0022] 3. This invention performs bidirectional projection from image features to point clouds and from point cloud geometric features to images, enabling texture information and 3D terrain information to be utilized in their respective spaces. It also dynamically generates fusion weights based on image quality, point density, vegetation cover, and geographical risk, thus preserving reliable modalities under conditions of shadow, vegetation occlusion, or sparse point clouds.

[0023] 4. This invention utilizes a first-precision path to quickly screen large areas, and only calls a higher-precision path for local areas with high risk, low confidence, modal conflict, or temporal changes, thereby reducing unnecessary high-resolution image processing and high-density point cloud computing. While ensuring the detection capability of high-risk areas, it reduces the average inference latency and edge power consumption.

[0024] 5. This invention combines single-phase disaster identification with multi-phase three-dimensional deformation analysis, utilizes stable regions to eliminate the overall deviation of point clouds in different phases, and outputs displacement, crack propagation, boundary migration and volume change, which can distinguish between surface color changes and true geometric deformation.

[0025] 6. This invention utilizes the structural topology between the trailing edge cracks, the main sliding wall, the landslide body, the leading edge bulge, and the accumulation zone, as well as the relationship between the displacement direction and the slope direction, to perform spatial consistency verification of the detection results. This can further suppress false alarms caused by isolated noise, road excavation surfaces, and ordinary bare soil surfaces.

[0026] 7. This invention can generate a re-inspection trajectory from the detection results, enabling the UAV to change its altitude and observation direction for point cloud blind spots, obstructed areas and high-risk uncertain areas, forming a closed loop of data collection, reasoning, risk assessment and autonomous re-inspection. Attached Figure Description

[0027] Figure 1 This is a block diagram of the overall system structure of the present invention; Figure 2 This is a flowchart of the multi-precision real-time detection method of the present invention; Figure 3 This invention generates a structural diagram of the geographical prior field. Figure 4 This is a schematic diagram of the two-dimensional-three-dimensional registration and bidirectional feature fusion of the present invention; Figure 5 This is a diagram showing the relationship between the multi-precision inference path and dynamic scheduling in this invention. Figure 6 This is a flowchart illustrating the decision-making process for selecting inference accuracy in this invention. Figure 7 This is a schematic diagram of the multi-temporal point cloud deformation analysis of the present invention; Figure 8 This is a schematic diagram of the topological verification of the landslide disaster structure according to the present invention; Figure 9 This is a schematic diagram of the UAV re-inspection trajectory driven by the detection results of this invention.

[0028] Figure labeling: 100, UAV inspection platform; 110, Visible light camera; 120, LiDAR; 130, Positioning and attitude determination component; 140, End-side computing unit; 150, Flight controller; 160, Communication component; 200, Geographic information module; 210, DEM / DTM data; 220, Geological structure data; 230, Historical disaster data; 240, Environmental inducing factor data; 250, Disaster-bearing body data; 260, Geographic prior field generation module; 300, Synchronization and preprocessing module; 310, Time synchronization unit; 320, Point cloud processing unit; 330, Image processing unit; 400, 2D-3D registration module; 410, Initial projection unit; 420, Local correction unit; 430, Registration reliability calculation unit; 500, Feature fusion. The following modules are included: 510 Image Feature Branch; 520 Point Cloud Feature Branch; 530 Geographic Prior Branch; 540 Two-way Projection Unit; 550 Credibility Gated Fusion Unit; 600 Multi-precision Inference Module; 610 First-precision Inference Path; 620 Second-precision Inference Path; 630 Third-precision Inference Path; 640 Inference Scheduler; 700 Multi-temporal Change Analysis Module; 710 Historical Baseline Data; 720 Stable Region Extraction Unit; 730 Deformation Calculation Unit; 800 Verification Output Module; 810 Disaster Structure Topology Verification Unit; 811 Trailing Edge Crack Zone; 812 Main Slide Wall; 813 Landslide Body; 814 Leading Edge Bulging Zone; 815 Accumulation Zone; 820 Disaster Detection Results; 900 Re-inspection Track Generation Module. Detailed Implementation

[0029] The present invention will be further described in detail below through embodiments. These embodiments are only used to illustrate the present invention and do not limit the scope of the present invention.

[0030] It should be noted that the "spatial unit" referred to in this invention refers to a unified spatial indexing unit used to organize geographic data, image data, point cloud data, and detection results. This unit can be a two-dimensional regular raster, a three-dimensional voxel, a topographic triangulation network, an octree node, or a geographic map node. "Multi-precision" refers not only to different numbers of neural network model parameters but also to different computational configurations such as input image resolution, point cloud sampling density, voxel size, number of feature layers, numerical quantization bit width, whether to call multi-temporal data, and whether to perform instance segmentation.

[0031] In this invention, "real-time" refers to completing data reception, screening, and risk output within the time delay constraints specified for the inspection task. These time delay constraints can be set based on the UAV's flight speed, sensor field of view, communication conditions, and task safety level, and are not limited to a fixed frame rate.

[0032] To achieve the above objectives, this invention first establishes a unified spatial reference. The UAV is equipped with a visible light camera, lidar, GNSS / RTK, IMU, flight controller, and edge computing unit; the visible light camera and lidar are synchronized through hardware triggering and unified clock or timestamp interpolation, the extrinsic parameters between the sensors are calibrated before flight, and local corrections are made during flight using stable terrain features.

[0033] This invention divides the target inspection area into multiple geospatial units. These spatial units can be two-dimensional grids, three-dimensional voxels, topographic triangulation units, or geographic map nodes with adjacency relationships. For each spatial unit, information such as DEM or DTM, slope, aspect, curvature, soil type, faults, rivers, road excavation, historical disasters, historical deformation, rainfall, and disaster-bearing bodies is read. Time validity, spatial resolution, and missing data markers are set for each data source.

[0034] The geographic prior module unifies the coordinates and normalizes the scale of various geographic data, generating a geographic risk prior map, a predicted sliding direction field, and an inference priority map. The geographic risk prior map is used to limit disaster candidate areas and adjust the model's attention; the predicted sliding direction field is used to verify the displacement direction; and the inference priority map is used to control image cropping resolution, point cloud sampling density, and subsequent inference paths.

[0035] Point cloud preprocessing includes time synchronization, motion distortion compensation, outlier removal, ground and non-ground point classification, vegetation point labeling, and point density estimation. Image preprocessing includes lens distortion correction, exposure quality assessment, motion blur assessment, and optional illumination normalization. For spatial units with sparse point clouds or blurred images, the system retains the corresponding data quality labels instead of directly filling them with normal data.

[0036] The 2D-3D registration module projects laser points onto the visible light image based on sensor intrinsic and extrinsic parameters, UAV pose, and acquisition time to obtain an initial correspondence. Subsequently, ridgelines, valley lines, road edges, building outlines, or rock edges are extracted in areas with low geographical risk, stable history, and significant edges. The local pose is then corrected by minimizing the distance between the image edge and the point cloud projection edge.

[0037] For each spatial cell, the local registration confidence level is calculated. This confidence level is jointly determined by the reprojection residual, effective point density, edge consistency, time synchronization error, and occlusion ratio. High confidence level allows pixel-to-point bidirectional fusion; medium confidence level uses local feature-level fusion; low confidence level stops forced correspondence and uses decision-level fusion or re-registration to prevent erroneous correspondences from propagating to subsequent networks.

[0038] The image branch extracts color, texture, crack edges, bare soil, vegetation destruction, and soil-rock boundary features from visible light images; the point cloud branch extracts elevation, relative elevation, slope, normal vector, curvature, roughness, local elevation difference, echo intensity, and point density features from laser point clouds; and the geographic prior branch encodes geographic risk, expected slip direction, and spatial adjacency information.

[0039] During bidirectional projection, image features are assigned to corresponding laser points according to a two-dimensional to three-dimensional mapping relationship, enabling each laser point to carry texture and crack information. Simultaneously, point cloud geometric features are projected onto the image feature plane, allowing image pixels to carry elevation and slope geometric information. The gating fusion module generates image gating coefficients and point cloud gating coefficients based on image quality, point cloud density, vegetation cover, registration reliability, and geographical risk to form two-dimensional to three-dimensional fused features.

[0040] The multi-precision inference module has at least three inference paths. The first precision inference path uses low-resolution imagery, sparse point clouds, and a lightweight detection network to quickly screen the entire inspection area. The second precision inference path uses higher-resolution imagery and denser point clouds for suspected areas to detect landslide boundaries, cracks, steep slopes, leading-edge bulging, and accumulation zones. The third precision inference path calls upon the original resolution data and historical point clouds to perform multi-temporal fine registration, three-dimensional displacement, crack propagation, and volume change calculations.

[0041] The inference scheduler calculates priority parameters based on geographical risk, prediction uncertainty, modal differences between the visible light branch and the point cloud branch, changes relative to historical data, current data quality, edge computing resource usage, remaining power, and allowable latency. High-risk or significantly deformed regions enter the third-precision inference path; medium-risk, low-confidence, or modal conflict regions enter the second-precision inference path; and continuous, stable, and low-risk regions remain on the first-precision inference path.

[0042] The result verification module converts the prediction results to a unified geographic coordinate system and verifies them based on the landslide structure topology. Verification includes: whether the displacement direction and slope direction are reasonable; whether the rear edge cracks, main landslide wall, landslide body, and leading edge deposition area satisfy spatial adjacency relationships; whether the deformation area is continuous; whether abnormal changes cross multiple adjacent spatial units; and whether the road excavation surface or normal bare land corresponds to known road and mining area layers.

[0043] When the point cloud density is insufficient, occlusion is severe, registration reliability is low, or high-risk results are still uncertain, the re-inspection module generates a re-inspection trajectory based on the slope normal vector, sensor range, and suspected boundary, controls the UAV to adjust its flight altitude, heading, speed, camera view, and laser scanning angle, and inputs the re-inspection data again into the inference path with the corresponding accuracy.

[0044] System composition and data acquisition: See Figure 1 The system in this embodiment includes an unmanned aerial vehicle (UAV) inspection platform 100, a geographic information module 200, a synchronization and preprocessing module 300, a two-dimensional to three-dimensional registration module 400, a feature fusion module 500, a multi-precision inference module 600, a multi-temporal change analysis module 700, a verification output module 800, and a re-inspection track generation module 900.

[0045] The UAV inspection platform 100 includes a visible light camera 110, a lidar 120, a positioning and attitude determination component 130, an edge computing unit 140, a flight controller 150, and a communication component 160. The visible light camera 110 can be a global shutter camera or a rolling shutter camera; when using a rolling shutter camera, the synchronization and preprocessing module 300 performs rolling shutter correction on the image based on the exposure time of each row of pixels and the IMU attitude. The lidar 120 can be a rotating, solid-state, or oscillating lidar.

[0046] The positioning and attitude determination component 130 includes GNSS, RTK, IMU, and an optional barometer. The visible light camera 110 and LiDAR 120 output timestamps via hardware triggering or a unified clock. For data with inconsistent timestamps, the synchronization and preprocessing module 300 uses adjacent IMU states for attitude interpolation, converting the image exposure center time and the sampling time of each laser point to a unified reference time.

[0047] Sensor calibration parameters include camera intrinsic parameters, camera distortion parameters, extrinsic parameters of the lidar to the body coordinate system, extrinsic parameters of the camera to the body coordinate system, and time deviation. Initial extrinsic parameters are obtained through a calibration field before flight; during flight, only local corrections to extrinsic parameters or pose are allowed within a preset deviation range to prevent dynamic corrections from deviating from the physical installation relationship.

[0048] The geographic information module 200 reads DEM / DTM data 210, geological structure data 220, historical disaster data 230, environmental inducing factor data 240, and disaster-bearing body data 250 from local storage or ground stations. For areas with unstable communication, the geographic data within the inspection range and its buffer zone are preloaded into the edge computing unit 140 first.

[0049] Generation of geographical prior fields: See Figure 3 The target inspection area is divided into spatial units i according to geographic coordinates. For each spatial unit, an original geographic feature vector Gi is constructed, which includes at least a portion of the following: slope Si, aspect Ai, curvature Ki, lithology Li, distance to fault FDi, distance to river RDi, distance to road or excavation area ODi, historical disaster frequency Hi, cumulative rainfall Pi, soil moisture content Wi, and disaster-bearing body importance Ei.

[0050] For continuous data, normalization is performed based on preset upper and lower limits; for categorical lithology and stratigraphy data, one-hot coding, embedded coding, or risk coding based on engineering geological grouping are used; a missing mask Mi is set for missing data; and a timeliness coefficient Ti is set for historical or environmental data. The geographic prior branch 530 inputs Gi, Mi, and Ti together into the encoder to avoid mistaking missing values ​​for low-risk values.

[0051] In one implementation, the initial geographic risk of spatial unit i can be expressed as: , where a1 to a10 are coefficients obtained during training or corrected during training after initialization by experts, and b is the bias. This formula is only an example; in practice, convolutional networks, graph neural networks, gradient boosting models, or rule-based models can be used to generate Ri.

[0052] The expected slip direction field Vi is calculated based on slope aspect, the direction of local maximum slope gradient, valley direction, and historical displacement direction. When geological structure or historical displacement information is reliable, the weight of the corresponding direction is increased; when the terrain is flat or the direction is uncertain, the directional confidence of Vi is decreased.

[0053] To leverage spatial adjacency relationships, the geographic prior field generation module 260 propagates geographic risks across adjacent spatial units at multiple scales, while preserving discontinuous interfaces such as faults, valleys, road boundaries, and ridges to avoid excessive smoothing across obvious terrain boundaries. The final output includes a geographic risk prior map R, a predicted slip direction field V, and an initial inference priority value P0.

[0054] Image and point cloud preprocessing: The point cloud processing unit 320 performs motion distortion compensation based on the sampling time of each laser point and the IMU attitude, transforming the point cloud to a selected reference time. Floating points are removed using statistical outlier removal, radius-based outlier removal, or outlier removal based on echo intensity.

[0055] Ground points and non-ground points in the point cloud can be determined through progressive morphological filtering, cloth simulation filtering, local slope constraints, or a trained point cloud classification network. Non-ground points do not need to be deleted all; instead, vegetation, building, vehicle, and unknown occlusion markers should be set separately. In vegetated areas, ground echoes or historical DTMs can be used preferentially.

[0056] Point cloud density di is calculated based on the number of effective points per unit surface area or per voxel. The system determines the voxel size vi based on the geographic risk prior Ri, for example, using a smaller vi for high-risk areas and a larger vi for low-risk areas, so that computing resources are tilted towards high-risk areas.

[0057] The image processing unit 330 performs distortion correction on the visible light image and calculates sharpness qi1, exposure quality qi2, shadow ratio qi3, and motion blur qi4. Image quality Qi can be obtained by weighting the above indicators. The image processing unit can also generate vegetation masks, sky masks, water masks, and masks for highly reflective areas.

[0058] 2D-3D registration and local confidence: See Figure 4 For a laser point pL, the coordinates are first transformed from the extrinsic parameters of the laser radar to the body coordinate system, the body pose at the sampling time, and the camera extrinsic parameters to the camera coordinate system. Then, the intrinsic parameters of the camera are used to project the result into a pixel u. No valid correspondence is established for laser points located behind the camera, outside the image range, or occluded by closer points.

[0059] After initial projection, image edge set EI and point cloud projection edge set EP are extracted in low-risk and historically stable regions. Point cloud projection edges can be determined by depth discontinuities, abrupt changes in normal vectors, or elevation breaks. Local correction unit 420 optimizes small-range pose increment δξ to minimize the weighted distance between EI and EP, and uses regularization terms to limit δξ from deviating from the initial calibration value.

[0060] For spatial cell i, calculate the reprojection mean square residual eri, effective point density di, edge matching ratio mi, time synchronization error ti, and occlusion ratio oi. The local registration reliability can be expressed as... Where b1 to b5 are non-negative coefficients. To avoid the point density being zero, the smallest positive number is added during the calculation.

[0061] When Cri is greater than the high confidence threshold, pixel-to-point bidirectional projection is performed; when Cri is between the high and low confidence thresholds, features are first aggregated in a larger spatial neighborhood and then fused; when Cri is lower than the low confidence threshold, individual pixel features are not directly assigned to the laser point, but the independent results of the two branches are retained and re-registration or re-examination is triggered.

[0062] Two-way projection and credibility-gated fusion: Image feature branch 510 can employ convolutional neural networks, visual Transformers, or a combination of both to output image feature maps FI at multiple scales. Point cloud feature branch 520 can employ point-level networks, sparse 3D convolutional networks, voxel networks, or neighborhood graph-based networks to output point cloud features FP. Geographic prior branch 530 outputs geographic features FG corresponding to each scale.

[0063] The bidirectional projection unit 540 performs bilinear sampling or neighborhood aggregation on the FI according to the two-dimensional-three-dimensional mapping relationship, and attaches the image texture features to the corresponding laser points; at the same time, it aggregates the elevation, normal vector, curvature and point density to the image feature plane according to the laser points in the neighborhood of each pixel. For the case where one pixel corresponds to multiple laser points, it is weighted according to depth, incident angle and point cloud confidence.

[0064] The confidence-gated fusion unit 550 generates image gating coefficients gI and point cloud gating coefficients gP. The inputs to the gating coefficients include image quality Q, point cloud density d, vegetation cover v, registration confidence Cr, geographic risk R, and modality missing markers. gI and gP can be normalized using softmax to... Geographic features (FG) can be used as additional bias or attention queries in the fusion process.

[0065] In areas with significant vegetation obstruction and effective surface echoes, increase gP; in areas with sparse point clouds but clear images, increase gI; when both modes are unreliable, do not force the generation of high-confidence results, but instead increase the prediction uncertainty and proceed to a re-examination.

[0066] Multi-precision inference paths and dynamic scheduling: See Figure 5 and Figure 6 The first-precision inference path 610 is designed for rapid screening over a large area. Its input can be a scaled-down image, a point cloud downsampled to larger voxels, and a low-resolution geographic prior map. The output is the hazard probability, rough boundary, and initial uncertainty for each spatial unit. The first-precision inference path can employ a pruned or quantized lightweight network.

[0067] The second-precision inference path 620 is for suspected area structure detection. Based on the primary output and the geographic risk map, the system cuts silhouette image areas, retains denser point clouds and uses finer voxels to perform semantic segmentation or instance segmentation of landslide areas, while detecting trailing cracks, main landslide walls, lateral edges, leading edge bulges and accumulation zones.

[0068] The third-precision inference path 630 is designed for high-risk or deformable areas. It calls upon the original resolution imagery, high-density point cloud, and at least one historical temporal data to perform local high-precision registration, point-level segmentation, crack size estimation, and calculation of three-dimensional displacement and volume change. The third-precision path can be completed independently by the end-side computing unit, or it can send local data or intermediate features to the ground station for collaborative processing when communication conditions are met.

[0069] The inference scheduler 640 calculates the priority parameters of spatial element i: Where Ui is the uncertainty, Di is the modal difference, Mi is the phase variation, Ti is the task time urgency, Ei is the normalized value of remaining battery power, and Ci is the computational resource usage; each λ is a non-negative weight. The computational resource usage term can also be set as a negative penalty, and the specific sign is determined according to the normalization definition at the time of implementation.

[0070] When Pi is higher than the high threshold, or either Ri or Mi exceeds the safety mandatory threshold, the third-precision inference path is directly selected; when Pi is in the middle range, or Di or Ui exceeds the corresponding threshold, the second-precision inference path is selected; otherwise, the first-precision inference path is maintained. To prevent the model from frequently switching near the threshold, different hysteresis intervals can be set for the entry and exit thresholds.

[0071] When there are multiple suspected areas requiring upgrades at the same time, the scheduler establishes a task queue based on the importance of the disaster-bearing body, disaster risk, estimated time to reach the edge of the frame, and computational cost. Priority is given to slopes above high-risk residential areas; for low-risk areas that are stable across multiple consecutive frames, delays or reduced accuracy are allowed.

[0072] Modal conflict review: The visible light image branch outputs the image hazard probability pI, image boundary BI, and apparent crack CI; the lidar branch outputs the point cloud hazard probability pP, geometric boundary BP, and deformation direction VP. The modal difference Di can be determined by |pI-pP|, the crossover ratio of BI and BP, the difference between crack location and terrain fault location, and the angle between VP and the expected slip direction V.

[0073] When an image branch is identified as a landslide but no geometric anomalies are found in the point cloud branch, the system determines whether the area corresponds to road construction, a quarry, bare farmland, or seasonal vegetation changes. If it corresponds to a known human activity layer, the disaster confidence level is lowered. If there are no records of human activity and the geographical risk is high, the system proceeds to the second-precision path review.

[0074] When a point cloud branch detects a steep slope or displacement while the image branch does not show obvious texture, the system determines whether there is vegetation occlusion or shadow. If the point cloud density and registration reliability are high, the weight of the point cloud branch is increased. If the point cloud density is low or the incident angle is unfavorable, a re-examination track with a changed scanning direction is generated.

[0075] When the image edge and the point cloud projection edge are misaligned, the conflict is attributed to local registration anomalies first, and local pose optimization is re-executed, rather than directly determining a certain modal error.

[0076] Multi-phase deformation analysis: See Figure 7From the historical baseline data 710, a historical point cloud P0 that is compatible with the current seasonal conditions, scanning direction, and point density of the region is selected, and the current point cloud Pt is obtained. The stable region extraction unit 720 excludes historical disaster areas, high-risk areas, vegetation, vehicles, water bodies, and newly constructed project areas to obtain a stable region set Ωs.

[0077] Rigid fine registration is performed on P0 and Pt based on Ωs to eliminate residual errors from GNSS or extrinsic parameters. Subsequently, point-to-surface distances, local surface differences, and changes in three-dimensional displacement vectors or voxel occupancy are calculated in the suspected disaster area, and the confidence intervals of these changes are estimated by combining local point density and scan incidence angle.

[0078] Spatial connectivity and temporal continuity are verified for isolated change points. Only when the change area spans a predetermined number of adjacent spatial units, or is aligned with the direction of cracks, steep slopes, and gradients, is it considered a valid deformation.

[0079] Crack width can be estimated based on the shortest distance between the point cloud surfaces on both sides of the crack or the combined drop between the image crack edge and the point cloud; landslide volume change can be obtained by integrating the elevation difference between the current surface and the historical benchmark surface within the disaster boundary; deformation rate is obtained by removing multi-time phases at time intervals.

[0080] Disaster structure topology verification: See Figure 8 The verification output module 800 constructs a topological map of the landslide disaster structure. The map nodes include the rear edge crack zone 811, the main landslide wall 812, the landslide body 813, the front edge bulging zone 814, and the deposition zone 815. The map edges represent spatial adjacency, uphill / downhill directional relationships, or deformation transmission relationships.

[0081] For each detected structural region, its relative position in the slope coordinate system is calculated. The trailing crack zone 811 should be located on the uphill side of the main landslide body 813; the leading bulge zone 814 and the deposition zone 815 should be located on the downhill side; the main landslide wall 812 is generally related to abrupt changes in slope elevation. These relationships are used to adjust the confidence level, not to require all landslides to simultaneously possess all structures.

[0082] The system also compares the predicted displacement direction with the expected slip direction field V. When the angle is less than a direction threshold, the consistency score is improved; when the angle is too large, it checks for lateral slippage, local collapse, or registration anomalies. After topology verification, results corresponding to the same geographic object in multiple frames or flight strips are merged to avoid duplicate alarms.

[0083] Re-inspection of flight track generation: See Figure 9 When a spatial unit meets any of the following conditions: high risk and low confidence, insufficient point cloud density, excessive occlusion ratio, or low registration confidence, the re-examination track generation module 900 determines the target surface that needs to be observed.

[0084] The re-inspection track generation module 900 calculates candidate observation directions based on the local slope normal vector and the effective incident angle range of the lidar, calculates candidate flight distances based on the focal length of the visible light camera and the desired ground resolution, and selects re-inspection waypoints by combining no-fly zones, terrain elevation, remaining battery power and track length.

[0085] The re-inspection track prioritizes covering areas not effectively covered by laser points in the primary scan, and ensures sufficient overlap between adjacent re-inspection viewpoints. After executing the re-inspection track, the flight controller 150 associates the newly acquired data with the original suspected area and directly enters the second or third precision inference path, without needing to perform a global screening of the entire inspection area again.

[0086] Neural network training: Training data is organized using uniform geospatial units. Each training sample includes at least visible light imagery, laser point cloud, sensor calibration parameters, 2D-3D correspondence, geographic prior features, disaster category labels, and disaster area labels; when multi-temporal data is available, displacement labels, crack propagation labels, or volume change labels are also included.

[0087] To enhance the model's robustness to modality loss and registration errors, during training, partial image occlusion, random downsampling of point clouds, simulation of illumination changes, small-scale extrinsic parameter perturbations, and temporal biases can be added. The real reliability label or the simulated reliability label is then input into the gating fusion module.

[0088] The total loss function can be weighted by classification loss Lcls, region segmentation loss Lseg, boundary loss Lbd, 2D-3D consistency loss Lcon, geographical prior constraint loss Lgeo, reliability gating loss Lgate, and multi-precision distillation loss Ldistill. Multi-precision distillation loss enables the first and second precision paths to learn the soft labels of the third precision path in the same region, while preserving the computational differences between the paths.

[0089] To prevent geographical priors from dominating the model and ignoring real sensor evidence, geographical priors can be randomly weakened or conflicting prior samples can be introduced during training. Even when sensor evidence is significantly inconsistent with geographical priors and the sensor quality is high, the model can still output a high-risk result for newly occurring disasters.

[0090] Exemplary execution process: In step S101, the UAV flies along a preset route, continuously acquiring visible light images, laser point clouds, and pose data. The edge computing unit 140 generates synchronization data packets according to time windows.

[0091] In step S102, the geographic prior field generation module 260 reads the terrain, geology, historical disasters and rainfall information covered by the current field of view, and generates a risk map R, a expected sliding direction field V and an initial priority value P0.

[0092] In step S103, the preprocessing module 300 synchronously completes image distortion correction, point cloud motion compensation, ground point classification, and data quality assessment, and performs non-uniform voxel sampling on the point cloud based on the risk map.

[0093] In step S104, the two-dimensional-three-dimensional registration module 400 establishes an initial projection relationship based on the pose and extrinsic parameters, performs local corrections in the stable region, and outputs the local registration confidence level Cr.

[0094] In step S105, the feature fusion module 500 extracts image, point cloud and geographic prior features, and performs bidirectional projection and gated fusion based on Cr, image quality and point density.

[0095] In step S106, the first precision inference path 610 screens the entire field of view and outputs the suspected disaster spatial units, initial confidence level, and uncertainty.

[0096] In step S107, the inference scheduler 640 selects the target inference path by combining geographical risk, modal differences, historical changes, resource status and task latency.

[0097] In step S108, for suspected medium-risk areas, the second-precision inference path 620 outputs the boundary, cracks, and disaster structures; for high-risk areas or areas with significant historical changes, the third-precision inference path 630 further outputs three-dimensional displacement and volume changes.

[0098] In step S109, the disaster structure topology verification unit 810 verifies the uphill and downhill positions, structural adjacency, and displacement direction, and merges the results from multiple frames.

[0099] In step S110, the system outputs geographic coordinates, boundary polygons, deformation parameters, and risk level; for high-risk areas that are still uncertain, the re-inspection track generation module 900 generates a re-inspection command.

[0100] Anomaly and Degradation Handling: When the visible light camera 110 is temporarily unavailable, the system performs single-modal inference using point cloud geometric features and geographic priors, and sets the image gating coefficients to zero. When the lidar 120 is temporarily unavailable, the system performs degraded inference using visible light features, low-resolution terrain features generated from the DEM, and geographic priors. The degraded results are marked as lower confidence levels.

[0101] When edge computing resources are insufficient, the system prioritizes second or third precision inference for high-risk spatial units, reducing the frame rate or point cloud density for low-risk areas. When communication is interrupted, the edge computing unit saves the original data, key intermediate features, and geographic indexes of suspected high-risk areas, and uploads them after communication is restored.

[0102] When there is a significant seasonal difference between the historical point cloud and the current point cloud, the stable region extraction unit excludes areas with obvious vegetation height changes; when a sufficient stable region cannot be obtained, it does not output definite absolute deformation variables, but outputs state markers that require supplementary baseline data.

[0103] Applicable methods for other geological disasters: This invention focuses on landslides as the primary detection target, but while maintaining geographical prior constraints, reliable fusion of two-dimensional and three-dimensional structures, and multi-precision scheduling mechanisms, it can be applied to collapse, rockfall hazard sources, and debris flow source areas.

[0104] When used for landslide detection, the disaster structure topology can be replaced with the spatial relationship between unstable rock mass, trailing edge fissures, free face, and accumulation body; when used for rockfall hazard source detection, protrusion, block volume, slope toe road location, and potential movement path can be added; when used for debris flow source area detection, geographical priors such as valley confluence, loose material source, gully slope, and rainfall threshold can be added.

[0105] All of the above replacements can use the same multi-precision inference scheduling method: first screen a large area, and then perform local high-precision point cloud and image analysis for high-risk or modal conflict areas.

[0106] Obviously, the above embodiments are only used to illustrate the technical concept of the present invention, and are not intended to limit the present invention. For those skilled in the art, equivalent substitutions or modifications made to the spatial unit form, network structure, loss function, threshold, weight, sensor type, and edge-cloud collaboration method without departing from the spirit and essence of the present invention should all fall within the protection scope of the present invention.

Claims

1. A multi-precision real-time detection method for landslide geological hazards based on geographic information constraints and visible light-lidar fusion, characterized in that, include: Acquire visible light images, lidar point clouds, flight position data, flight attitude data, timestamps, and sensor calibration parameters of the target inspection area simultaneously collected by the UAV; Geographic information of the target inspection area is acquired and converted into geographic prior features corresponding to spatial units. The geographic prior features include at least three of the following: topographic features, geological features, historical disaster features, environmental inducing factors, disaster-bearing body risk features, and expected sliding direction. Motion distortion compensation and surface classification are performed on the point cloud, distortion correction is performed on the image, and an initial two-dimensional-three-dimensional mapping between image pixels and laser points is established based on pose and calibration parameters. The initial mapping is locally registered and corrected using stable terrain features to obtain the updated two-dimensional-three-dimensional mapping and the local registration reliability of each spatial unit; Image features, point cloud geometric features, and geographic prior features are extracted. Bidirectional projection is performed based on the two-dimensional-three-dimensional mapping. The fusion weights are determined according to image quality, point cloud density, registration reliability, and geographic risk to obtain two-dimensional-three-dimensional fused features. The fused features are globally inferred using the first precision inference path to obtain suspected disaster spatial units and their initial confidence and prediction uncertainty; The target inference path is determined from multiple candidate inference paths based on at least four of the following factors: geographical risk, prediction uncertainty, modal differences, surface change, data quality, and edge computing resources of the suspected disaster spatial unit. The target reasoning path is used to perform at least one of the following processing methods on the suspected disaster spatial unit: landslide segmentation, boundary identification, crack identification, terrain change detection, and multi-temporal deformation calculation, to obtain the disaster detection result; Based on geographical prior features, slope orientation, and landslide structure topology, the detection results are spatially consistent and output at least one of the following: disaster category, boundary, displacement, elevation change, volume change, risk level, location, and re-inspection instruction.

2. The method for multi-precision real-time detection of landslide geological hazards based on geographic information constraints and visible light-lidar fusion as described in claim 1, is characterized in that, The geographic information includes at least four of the following: DEM / DTM, slope, aspect, curvature, topographic relief, soil and rock type, fault location, distance from river / road / excavation area, historical disaster boundary, historical deformation, rainfall, soil moisture content, seismic motion parameters, and distribution of settlements / roads; after coordinate unification, normalization, and reliability labeling, it is encoded into a geographic risk prior map, a expected slip direction field, and an inference priority map.

3. The method for multi-precision real-time detection of landslide geological hazards based on geographic information constraints and visible light-lidar fusion as described in claim 1, is characterized in that, Non-uniform processing is applied based on geographical risk level: high-risk units are processed with smaller point cloud voxels, higher point cloud retention, and higher image resolution, while low-risk units are processed in the opposite way.

4. The method for multi-precision real-time detection of landslide geological hazards based on geographic information constraints and visible light-lidar fusion as described in claim 1, characterized in that, The stable terrain features include at least one of ridgelines, valley lines, road edges, building outlines, exposed rock edges, and terrain fault lines; the local registration correction is accomplished by minimizing the reprojection residual between the image edge and the point cloud projection edge and limiting the pose deviation range; the registration reliability is determined based on the reprojection residual, effective point density, edge consistency, time synchronization error, and occlusion ratio; when the registration reliability is lower than the threshold, the pixel-level fusion weight is reduced, and feature-level or decision-level fusion is switched.

5. The method for multi-precision real-time detection of landslide geological hazards based on geographic information constraints and visible light-lidar fusion as described in claim 1, characterized in that, The image features include at least two of color, texture, crack edges, bare soil, vegetation destruction, and soil-rock boundary; the point cloud geometric features include at least three of elevation, relative elevation, slope, normal vector, curvature, roughness, local elevation difference, echo intensity, and point density; the fusion weights are also determined based on shadow / overexposure degree, vegetation coverage, and modal integrity.

6. The method for multi-precision real-time detection of landslide geological hazards based on geographic information constraints and visible light-lidar fusion as described in claim 1, is characterized in that, The candidate inference path includes at least a first, second, and third precision path, which differ in at least three of the following: image resolution, point cloud density, voxel size, network depth, number of feature channels, number of detector heads, quantization bit width, whether instance segmentation is performed, and whether historical point clouds are called. The inference scheduler selects paths based on priority parameters weighted by geographical risk, uncertainty, modal differences, variability, data quality, computing resource usage, remaining power, and allowable latency: paths above the first threshold are selected with the third precision, paths between the thresholds are selected with the second precision, and paths otherwise maintain the first precision.

7. The method for multi-precision real-time detection of landslide geological hazards based on geographic information constraints and visible light-lidar fusion as described in claim 1, characterized in that, The modal differences are determined based on the disaster probability, boundary overlap, and displacement direction consistency between the visible light branch and the point cloud branch. When the difference threshold is exceeded, the cause is identified as vegetation shading, shadows, sparse point clouds, or registration anomalies, and the corresponding branch weight is increased or re-registered accordingly.

8. The method for multi-precision real-time detection of landslide geological hazards based on geographic information constraints and visible light-lidar fusion as described in claim 7, is characterized in that, The multi-temporal deformation calculation includes: extracting stable regions from historical and current point clouds, eliminating overall deviations between temporal phases, calculating local point-to-surface distances or three-dimensional displacement vectors, eliminating isolated change points based on continuity and confidence intervals, and obtaining at least one of crack propagation, boundary migration, displacement, deformation rate, and volume change; the landslide structural topology includes the spatial adjacency relationships of the rear edge crack zone, main landslide wall, landslide body, lateral edge shear zone, front edge bulging zone, and deposition zone; the consistency verification includes determining whether the displacement direction and slope direction satisfy a preset angle relationship, whether the deformation region is continuous, and whether each structural zone satisfies the adjacency relationship.

9. The method for multi-precision real-time detection of landslide geological hazards based on geographic information constraints and visible light-lidar fusion as described in claim 8, is characterized in that, When the point cloud density is below the threshold, the occlusion ratio exceeds the threshold, the registration confidence is below the threshold, or the disaster risk exceeds the threshold, a re-inspection track is generated based on the slope normal vector, suspected boundary, occlusion area, and sensor range, including re-inspection waypoints, altitude, speed, heading angle, and scanning angle; when a single mode is missing or communication is interrupted, the path is switched to a degraded inference path and the confidence level is reduced, and high-risk units are re-inspected after recovery.

10. A multi-precision real-time detection system for landslide geological hazards based on geographic information constraints and visible light-lidar fusion, as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire synchronously collected visible light images, point clouds, poses, timestamps, and calibration parameters. The geographic prior module is used to generate geographic prior features based on topography, geology, historical disasters, environmental factors, and disaster-bearing body information; The preprocessing and registration module is used for point cloud motion distortion compensation, image distortion correction, establishing and correcting two-dimensional to three-dimensional mapping, and calculating local registration reliability. The feature fusion module is used to extract image features, point cloud geometric features and geographic prior features, perform bidirectional projection, and generate two-dimensional-three-dimensional fused features based on data quality, registration reliability and geographic risk. The multi-precision inference module includes multiple inference paths and schedulers with different precision and computational loads. It is used to select the target path based on geographical risk, uncertainty, modal differences, variability, data quality, and resource status after global screening. The verification output module is used to output the disaster category, boundary, deformation parameters, risk level, location, or re-inspection command based on the slope direction and landslide structure topology verification and detection results. The flight feedback module communicates with the flight controller and is used to adjust at least two of the waypoint, altitude, speed, heading angle, camera orientation, and scan orientation according to the retest command.

Images