A dangerous rock body collapse hidden danger identification method based on unmanned aerial vehicle oblique photography

Abstract

The application discloses a dangerous rock body collapse hidden danger identification method based on unmanned aerial vehicle oblique photography and belongs to the technical field of geological disaster monitoring.The application collects slope images through unmanned aerial vehicle oblique photography, adopts an improved SfM algorithm to construct a high-precision three-dimensional model, constructs a geometric-texture-environment multi-modal feature library, automatically delimits a dangerous rock body candidate area with the aid of attention deep learning, adaptively extracts structure surface parameters and returns to iteration model, couples rainstorm and earthquake working conditions to correct rock mass shear strength, discriminates instability modes in combination with the polar projection and discrete element simulation, completes four-grade risk grading by adopting a hierarchical-fuzzy coupling model, demarcates a dangerous range by adopting rockfall simulation, periodically and automatically rephotographs high-risk areas, and forms a data iteration closed loop.The application does not need manual high-altitude operation, the instability discrimination accuracy is 94%, the risk grading adhesion degree is 92%, the problems of low efficiency of traditional identification methods, insufficient discrimination precision in extreme working conditions and incapability of dynamically optimizing a monitoring model are solved, and the application is suitable for long-term dynamic hidden danger monitoring of mountainous area traffic, water conservancy and power slopes.

Description

Technical Field

[0001] This invention belongs to the technical field of geological exploration, specifically relating to a method for identifying potential rockfall hazards based on UAV oblique photography. Background Technology

[0002] Rockfall collapse refers to a typical geological hazard in which rock and soil masses on steep slopes or cliffs detach from the parent rock under the long-term combined action of various internal and external geological forces, such as gravity, seismic forces, precipitation seepage, and weathering unloading, and suddenly and rapidly roll, jump, and finally accumulate at the foot of the slope. Rockfalls are usually controlled by the original structural planes within the rock mass (such as joints, fissures, and faults). As the fissures at the rear edge continue to expand and penetrate, the supporting surface at the bottom is gradually weakened. Once the stress concentration exceeds the shear strength of the rock mass, it will instantly become unstable and fail. This hazard is characterized by its suddenness, short duration, enormous destructive force, and difficulty in accurate prediction, often posing a serious threat to mountainous transportation routes, water conservancy facilities, buildings, and the safety of people's lives and property. In modern disaster prevention and mitigation systems, multi-source sensing and monitoring technologies integrating air, space, and ground are typically employed, such as 3D laser scanning, InSAR interferometry, and micro-deformation monitoring and early warning. These are combined with finite element numerical simulation to assess the stability of unstable rock masses, and comprehensive management is achieved through active reinforcement measures such as anchoring and grouting, or passive protection engineering measures such as rockfall barriers and flexible shelters. Identification of potential rockfall collapse hazards refers to the fundamental work for disaster prevention and mitigation, which comprehensively utilizes geological theory and modern space exploration technology to conduct early detection, spatial location, and stability assessment of potentially unstable rock masses that may experience instability. This identification process typically includes two levels: macroscopic geological inspection and microscopic quantitative monitoring. Field investigations mainly rely on high-resolution satellite remote sensing, UAV oblique photography, and 3D laser scanning technology. By constructing a high-precision digital elevation model (DEM), abnormal mountain morphology is visually analyzed, and surface deformation features such as tension cracks, rock cavity grooves, and faults are accurately extracted. Simultaneously, geophysical exploration methods are used to determine the distribution of hidden structural planes and weak interlayers within the rock mass. In terms of quantitative analysis, crack gauges, tilt sensors, and BeiDou GNSS displacement monitoring equipment are deployed to capture the dynamic evolution of micro-deformation of unstable rock masses under inducing factors such as rainfall or earthquakes in real time. Scientific and accurate hazard identification can effectively delineate the boundaries of collapse hazard zones, providing reliable data support and decision-making basis for subsequent risk assessment, early warning, and engineering intervention.

[0003] Existing technologies have several core flaws: First, traditional manual surveying methods require workers to operate at heights near edges, posing high safety risks. Relying on compasses for manual measurement of structural surfaces is time-consuming, results in small sample sizes, and leads to significant human error, making large-scale slope inspections extremely inefficient. Second, conventional UAV oblique photography identification schemes rely solely on single features such as grayscale and terrain slope to identify unstable rock masses, failing to integrate multi-dimensional environmental information. This makes it difficult to distinguish between fractures and dominant structural surfaces controlling instability, leading to missed or incorrect assessments. Existing publicly available technologies, such as UAV slope geological mapping methods and single stereographic projection rock mass stability evaluation methods, have not addressed these technical pain points, making it difficult to achieve high-precision, automated, and dynamically closed-loop identification of potential rock mass collapse hazards. Summary of the Invention

[0004] To address the shortcomings of existing methods for identifying dangerous rock masses, such as high risks associated with manual operations, insufficient accuracy of multi-feature fusion identification, poor reliability in instability assessment under extreme conditions, and the lack of a dynamic monitoring closed-loop optimization system, this invention aims to provide a method for identifying potential rock mass collapse hazards based on UAV oblique photography. This method enables unmanned high-altitude operations, intelligent delineation of dangerous rock masses, instability assessment coupled with rainstorm and earthquake conditions, risk classification, and periodic automatic retesting and iterative optimization of the model in high-risk areas.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying potential rockfall hazards based on UAV oblique photography includes the following steps: S1. Multi-view oblique photography data acquisition and 3D modeling: A multi-lens UAV oblique photography system is used to take aerial photos of the target slope with full coverage, and flight environment parameters are collected and stored synchronously; an improved motion recovery structure algorithm is used to generate dense point clouds, and a 3D real scene model with bound spatial coordinates and camera attitude is obtained through surface reconstruction and texture mapping. The improved motion recovery structure algorithm includes image feature extraction, feature matching based on image overlap, sparse point cloud construction, and pixel disparity solution with multiple constraints to generate dense point cloud. S2. Multimodal feature fusion intelligent pre-identification of dangerous rock masses: Based on 3D real scene model, image and flight environment parameters, construct a multi-dimensional feature library of geometric shape, surface texture and environmental factors; use attention mechanism deep learning model to fuse multiple types of features, calculate the contribution weight of each type of feature through dynamic weight model, quantify the confidence of regional dangerous rock mass hidden danger, automatically delineate candidate areas of dangerous rock masses and output boundary coordinates; S3. Targeted Structural Surface Parameter Extraction: An adaptive survey line network is deployed for the candidate areas of the unstable rock mass. Structural surface parameters, including the orientation, dip angle, and spacing of the structural surfaces, are extracted based on point cloud normal vector clustering. These parameters are then fed back to the S2 deep learning model for iterative optimization of feature weights. In this operation, the density of the survey line network is adaptively adjusted according to the texture complexity of the candidate areas, with denser survey line deployment in areas with drastic texture changes. S4. Extreme working condition coupled instability mode identification: A mechanical analysis framework is built based on structural surface parameters and geometric characteristics of unstable rock mass. The shear strength of the structural surface is corrected by simulating dynamic load conditions such as rainstorm and earthquake. The unstable rock mass instability mode is comprehensively determined by combining stereographic projection qualitative analysis and discrete element numerical simulation. S5. Multi-factor coupled stability classification evaluation: Select multiple influencing factors such as rock mass integrity, groundwater, rainfall, earthquake, and instability matching degree, and use hierarchical analysis and fuzzy comprehensive evaluation method to classify the risk level of dangerous rock mass. High-risk samples are synchronously updated to the S2 model training set. S6. Dynamic landslide impact zone delineation and full-process data closed-loop optimization: The impact range of landslides at different risk levels is delineated by simulating rockfall trajectories; refined reshoot routes are automatically generated based on the coordinates of high-risk areas; high-risk areas are periodically retested and modeled; the structural surface recognition database and deep learning model are updated to form a closed-loop recognition system of "collection-recognition-evaluation-iteration".

[0006] Furthermore, the feature matching based on image overlap in step S1 is completed using a matching similarity formula: ; In the formula, S ij Let d be the similarity between corresponding feature points in the i-th image and the j-th image. ij O is the Euclidean distance between two feature point descriptors. ij Let s be the overlap between the i-th image and the j-th image. ij α represents the scale difference between the two feature points, and β represents the matching weight coefficients. Pixel disparity is calculated using a multi-constraint cost function that integrates photometric, geometric, word count, and point cloud density factors. ; In the formula, C(x,y,d) is the matching cost of pixel (x,y) under disparity d; Indicates the cost of photometric matching; For geometric matching cost; Let λ be the sparse point cloud density of the region corresponding to pixel (x,y), and λ and γ be the balancing weight coefficients of the cost function.

[0007] Furthermore, the formula for calculating the feature contribution weight in step S2 is as follows: ; In the formula, n=3, corresponding to three types of features: geometric shape, surface texture, and environmental factors; The fusion weights for the k-th class features are... Let k be the inter-class discriminant of the feature class k. Let be the coefficient of variation of the k-th class of features; The formula for calculating the confidence level of potential rock mass hazards in the quantified area is as follows: ; In the formula, This represents the Sigmoid activation function. Let N(x,y,z) represent the normalized value of the k-th class feature at the 3D coordinates (x,y,z), and let N(x,y,z) represent the neighborhood consistency feature at the 3D coordinates (x,y,z). Weight coefficients for neighborhood features Further, step S4 constructs a modified model of the effective shear strength of the structural surface coupled with pore water pressure and seismic load: ; In the formula, For the effective shear strength of the structural surface, For the normal stress of the structural surface, Let ρ be the pore water pressure, Φ be the internal friction angle of the structural surface, and c be the cohesion of the structural surface. This represents the horizontal influence coefficient of earthquakes. This represents the peak horizontal seismic acceleration. The type of instability is determined by fusing the results of the two types of analysis using an instability comprehensive discrimination index. ; In the formula, A comprehensive index for identifying instability modes. The degree of matching between the dominant structural surface and the free surface in stereographic projection. This represents the displacement mutation rate of the unstable rock mass in the discrete element simulation. These are the weighting coefficients for stereographic projection analysis; These are the weighting coefficients for discrete element numerical simulation.

[0008] The discrete element numerical simulation uses 3DEC software to calculate the stress and displacement evolution of the rock mass.

[0009] Furthermore, step S5 uses the analytic hierarchy process (AHP) to assign weights to five evaluation factors: rock mass integrity coefficient 0.3, groundwater seepage pressure 0.25, instability mode matching degree 0.2, rainfall intensity 0.15, and peak ground acceleration 0.1. Four risk membership thresholds are set: extremely high risk ≥ 0.8, high risk 0.6~0.8, medium risk 0.4~0.6, and low risk < 0.4. The data acquisition time difference for all evaluation factors must be ≤ 24 hours, and the instability mode matching degree must be ≥ 90% before the classification results can be output.

[0010] Furthermore, the simulation parameters for the falling rock trajectory in step S6 are as follows: initial velocity of the falling rock 2~5m / s, rolling friction coefficient 0.3~0.5, and bounce recovery coefficient 0.2~0.4; the influence radius is defined by repeating the simulation 100 times and taking the 95% confidence interval: 30~50m for extremely high risk, 15~30m for high risk, 5~15m for medium risk, and 0~5m for low risk; the parameters for the reshoot flight path are as follows: flight altitude reduced by 30% compared to the first time, directional overlap of 85%, lateral overlap of 75%, covering the core risk area and the surrounding 10m transition zone, and the accuracy of the local three-dimensional model ≤5cm.

[0011] Furthermore, step S6 sets differentiated reshoot cycles: reshoots are conducted once every 7 days for extremely high-risk areas, once every 15 days for high-risk areas, and once a month for medium- and low-risk areas; aerial surveys are only performed when visibility is ≥5km and wind force is ≤4; every 50 new high-risk unstable rock mass samples added in step S2 automatically trigger model retraining, and every 30 new high-risk structural surface data added in step S6 update the structural surface identification algorithm.

[0012] Compared with the prior art, the present invention has the following advantages: 1. Automated extraction of structural surface parameters significantly improves data acquisition efficiency and accuracy. This method utilizes UAV oblique photogrammetry technology to quickly acquire structural parameters of unstable rock masses without requiring manual on-site measurements. This completely solves the problems of low efficiency, limited data volume, and susceptibility to human error associated with traditional manual measurements. By directionally deploying survey lines in candidate areas of unstable rock masses and focusing on information collection in core regions, it reduces invalid calculations in non-target areas. This allows for the acquisition of large-scale, high-precision core parameters such as structural surface orientation, dip angle, and spacing in a short time, providing reliable data support for subsequent instability mode identification and stability evaluation, while also reducing the safety risks of field operations.

[0013] 2. Intelligent pre-identification through multi-modal feature fusion accurately identifies advantageous structural surfaces. A multi-dimensional feature library encompassing "geometric morphology, surface texture, and environmental factors" is constructed. Through deep learning and a customized fusion analysis model, it accurately identifies the dominant structural surface combinations that have the greatest impact on the stability of unstable rock masses, avoiding the ineffective analysis of secondary structural surfaces in traditional techniques. This model can automatically delineate candidate areas of unstable rock masses and push boundary coordinates in real time, achieving seamless integration between pre-identification results and subsequent structural surface extraction. This reduces the workload of repetitive data processing, makes instability mode identification more targeted, and significantly improves the accuracy of hazard identification.

[0014] 3. Multi-factor coupled stability classification evaluation improves the reliability of classification results. By integrating instability mode identification results, environmental monitoring data, and key factors such as rock mass integrity coefficient and groundwater seepage pressure, a multi-dimensional coupled evaluation system is constructed. The weights of each factor are determined using the analytic hierarchy process (AHP), and stability is graded using fuzzy comprehensive evaluation. Compared to traditional single-factor evaluation or experience-based judgment methods, this method fully considers the comprehensive impact of multiple factors on the stability of unstable rock masses, ensuring that the grading results highly match the actual potential risks. This provides a scientific and reliable basis for subsequent impact zone delineation and hazard mitigation.

[0015] 4. Multi-method instability mode identification under extreme conditions, enhancing adaptability to complex scenarios. A mechanical analysis framework was constructed by combining structural parameters and the geometric characteristics of unstable rock masses. Through a combination of stereographic projection analysis and discrete element numerical simulation, the response of unstable rock masses under extreme conditions such as torrential rain erosion and seismic disturbances was simulated, enabling more accurate identification of instability modes. Traditional single methods struggle to guarantee accuracy under extreme conditions, while this method, through complementary verification using multiple techniques, improves the reliability of instability mode identification in extreme scenarios, providing a core basis for stability classification.

[0016] 5. Dynamic impact zone delineation based on rockfall trajectory simulation enables precise control of risk scope. By employing a rockfall trajectory simulation tool, combined with slope topographic data and rock mass physical and mechanical parameters, the collapse impact range corresponding to different stability levels is delineated, completely overcoming the limitations of traditional experience-based estimations. This method can accurately calculate the boundary coordinates of high-risk impact areas, providing precise scope basis for subsequent hazard mitigation and regional control. It avoids the expansion or omission of risk range due to experience bias, thus improving the scientific nature and targeted nature of hazard mitigation.

[0017] 6. A closed-loop data optimization mechanism throughout the entire process continuously improves the practicality and accuracy of the recognition system. A complete closed-loop workflow is constructed, encompassing data acquisition, intelligent identification, risk assessment, and dynamic optimization. High-risk rock mass characteristic data is synchronized to the fusion analysis model to update training samples. Simultaneously, drones are triggered to conduct secondary image captures of high-risk areas, regenerating local 3D real-world models and updating structural surface parameter samples. This dynamic optimization mechanism continuously improves the model's sensitivity to identifying high-risk areas and constantly refines the structural surface parameter database, enabling the entire identification process to adapt to the dynamic changes of rock masses and significantly enhancing the method's practicality and operability. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0020] S1: Multi-view oblique photogrammetry data acquisition and basic 3D modeling A multi-lens UAV oblique photography system is used to take aerial photos of the target slope with full coverage, and flight environment parameters are collected and stored synchronously. An improved motion reconstruction structure algorithm is used to generate dense point clouds, and a 3D real-world model with bound spatial coordinates and camera attitude is obtained through surface reconstruction and texture mapping. The improved motion reconstruction structure algorithm includes image feature extraction, feature matching based on image overlap, sparse point cloud construction, and pixel disparity solution with multiple constraints to generate dense point clouds.

[0021] First, SIFT features are extracted from each tilted image. A guided matching strategy based on image overlap is constructed to prioritize matching feature points from adjacent high-overlap images, reducing cross-image matching errors. Next, the perspective n-point algorithm is used to solve for camera extrinsic parameters, and a sparse point cloud is constructed by combining it with pre-calibrated camera intrinsic parameters. On this basis, an improved cost function that fuses photometric and geometric constraints is introduced, and a dense point cloud is generated by calculating disparity pixel by pixel. Finally, Poisson surface reconstruction is performed on the dense point cloud, mapping the original image texture to the surface model to obtain an initial 3D real-world model with realistic visual effects. Throughout the process, camera pose parameters and point cloud coordinates are stored in real time, providing an accurate spatial reference for subsequent steps such as structural surface extraction and unstable rock mass identification.

[0022] To address the significant differences in overlap between tilted UAV images, an improved matching similarity formula is constructed by introducing constraints on image overlap and feature scale differences:

[0023] Parameter definition: This represents the matching similarity between corresponding feature points in the i-th image and the j-th image. The value ranges from 0 to 1, and the larger the value, the higher the matching reliability. The Euclidean distance between two feature point descriptors reflects the appearance similarity of the feature points. The overlap between the i-th image and the j-th image is calculated from the UAV flight path planning parameters and the actual shooting data. This represents the scale difference between two feature points, calculated from the scale space parameters during SIFT feature extraction. and These represent weighting coefficients, which were experimentally calibrated to 0.8 and 1.2 to balance the effects of different constraint terms; To improve the accuracy of dense point cloud generation, geometric constraints on sparse point cloud density are introduced, and a cost function that integrates multi-dimensional information is constructed:

[0024] Parameter definition: C(x,y,d) represents the matching cost of pixel (x,y) under disparity d. The smaller the value, the more reliable the matching result. The photometric matching cost is calculated using a normalized cross-correlation algorithm and reflects the grayscale similarity of pixels. It represents the geometric matching cost, which is calculated based on the camera pose parameters to determine the reprojection error of spatial points and reflects the geometric consistency of the matching. It represents the sparse point cloud density of the region corresponding to pixel (x,y), which is obtained by statistics from sparse point cloud data and is used to constrain the rationality of disparity calculation; λ and γ represent weighting coefficients, set to 0.7 and 0.3 respectively, to balance the effects of photometric, geometric, and sparsity constraints.

[0025] S2: Intelligent pre-identification of unstable rock masses through multimodal feature fusion The system retrieves 3D point clouds and high-resolution image textures generated by S1, and imports auxiliary data such as flight attitude and illumination changes to build a multi-dimensional feature library encompassing "geometric morphology, surface texture, and environmental factors." A customized fusion analysis model is used to perform deep learning on the feature library, automatically delineating candidate areas for unstable rock masses. The boundary coordinates of these areas are then pushed to the structural surface extraction stage of S3 in real time, achieving seamless integration of pre-identification results and avoiding redundant data processing.

[0026] First, multi-dimensional features are extracted from the 3D real-world model generated by S1 and the associated database to construct a feature library of "geometric morphology, surface texture, and environmental factors." Geometric morphology features are extracted from 3D point cloud data, including parameters such as principal curvature, slope change rate, and local convexity of the rock mass surface, while calculating the volume-to-surface area ratio of potential unstable rock masses. Surface texture features are based on the original oblique imagery, extracting indicators such as texture contrast, entropy, and correlation through gray-level co-occurrence matrix. Environmental factors utilize on-site parameters stored in S1, such as UAV flight attitude, real-time wind speed, and light intensity. A customized attention mechanism is used to integrate a deep learning model, and the three types of features are input into the model for joint training. The model automatically learns the contribution weights of different features to unstable rock mass identification, achieving deep fusion of multi-modal features. After training, the target slope area is scanned region by region, automatically delineating candidate areas of unstable rock masses with potential collapse risks. The 3D coordinates of the region boundaries are pushed in real time to the structural surface extraction stage of S3, ensuring that S3 directly performs accurate analysis on the candidate areas, avoiding meaningless repetitive data processing on the entire rock mass.

[0027] To address the varying contributions of different features to the identification of unstable rock masses, a dynamic weight calculation model is constructed by introducing dual constraints of feature discriminative power and dispersion.

[0028] In the formula: This represents the fusion weight of the k-th feature, with a value ranging from 0 to 1. The larger the weight, the higher the contribution of the feature to the identification of dangerous rock masses. The inter-class discrimination of the k-th feature is obtained by calculating the ratio of the difference in the mean to the standard deviation of the features between the dangerous rock mass sample and the normal rock mass sample. The coefficient of variation represents the k-th feature, reflecting the dispersion of the feature in the sample. It is calculated as the ratio of the feature's standard deviation to its mean. n represents the total number of feature categories in the feature library. n=3, corresponding to the three categories of features: geometric shape, surface texture, and environmental factors. In the process of calculating the regional confidence level of unstable rock masses, multimodal features and neighborhood spatial information are integrated to construct a quantitative calculation model for the probability of unstable rock mass hazards:

[0029] In the formula: P(x,y,z) represents the confidence level that the region at the three-dimensional coordinates (x,y,z) is a dangerous rock mass. The value ranges from 0 to 1, and the larger the value, the higher the probability of the hidden danger. This represents the Sigmoid activation function, used to map fused feature values ​​to a probability range of 0 to 1; The normalized value of the k-th feature at the three-dimensional coordinates (x, y, z) is obtained by mapping the original feature value to the interval between 0 and 1. N(x,y,z) represents the neighborhood consistency feature at the three-dimensional coordinates (x,y,z), which is obtained by calculating the average similarity of features within the neighborhood. The weighting coefficient representing the neighborhood features is experimentally calibrated to 0.3, which is used to balance the influence of single-pixel features and neighborhood spatial information.

[0030] S3: Extraction of regional directional structural surface parameters of unstable rock mass For the candidate areas of unstable rock masses delineated in S2, survey lines are adaptively arranged and extracted based on the morphological characteristics of the areas. Structural surface information is extracted focusing on the target areas, reducing invalid calculations for non-target areas. After extracting core parameters such as the strike, dip angle, and spacing of the structural surfaces, the parameter dataset is sent back to the model training module of S2 to optimize the feature weight allocation of the model and gradually improve the accuracy of pre-identification.

[0031] In the above steps, the boundary data of the candidate unstable rock mass area delineated in step S2 is used. A survey line network is automatically generated based on the geometric morphology of the area. The survey line density is dynamically adjusted according to the texture complexity within the area, with denser survey line placement in areas of drastic texture changes. A structure surface recognition algorithm based on point cloud normal vectors is used to calculate the normal vectors point-by-point in the survey line coverage area. Core parameters such as the orientation, dip angle, and spacing of the structure surfaces are extracted through cluster analysis. After extraction, the parameter dataset is pushed in real-time to the fusion analysis model training module in step S2. The model updates the feature weight allocation logic based on the new parameters, continuously optimizing the accuracy of pre-identification. During processing, the illumination intensity data collected in step S1 is referenced simultaneously, avoiding periods of sudden changes in illumination during normal vector calculations to ensure the stability and accuracy of the extracted structure surface parameters.

[0032] S4: Identification of unstable rock mass modes under dynamic working conditions By combining the structural parameters of S3 with the geometric characteristics of the unstable rock mass in S2, a mechanical analysis framework was constructed. Extreme conditions such as torrential rain erosion and seismic disturbances were simulated. Instability modes were identified through a combination of stereographic projection analysis and discrete element numerical simulation. The identification results were synchronized in real time to the stability evaluation stage in S5, providing core judgment criteria for graded assessment.

[0033] First, the structural plane attitude parameters extracted by S3 and the geometric feature data of the unstable rock mass delineated by S2 are used to build a mechanical analysis framework that couples the structural planes and the overall morphology. For the rainstorm scouring condition, the pore water pressure distribution within the structural planes is calculated based on the spacing between the structural planes and the rock mass permeability coefficient. For the seismic disturbance condition, the horizontal seismic acceleration coefficient is introduced to simulate the dynamic load effect. The stereographic projection method is used to analyze the combination relationship between the dominant structural planes and the free face of the slope, and to preliminarily determine the potential instability trend. Simultaneously, based on discrete element numerical simulation software, the structural plane parameters are converted into joint element attributes in the model, and the geometric features of the unstable rock mass are mapped to the spatial distribution of discrete element particles to simulate the stress transmission and displacement evolution process inside the unstable rock mass under extreme conditions. By comparing the qualitative analysis results of the stereographic projection with the quantitative displacement cloud map and stress concentration areas from the discrete element simulation, the instability modes of the unstable rock mass are comprehensively identified, including tensile cracking and collapse, toppling and falling, and faulting and rolling.

[0034] To address the coupling effects of rainstorms and earthquakes during the correction of the effective shear strength of structural surfaces under extreme conditions, a correction model for the shear strength of structural surfaces considering pore water pressure and seismic loads is constructed:

[0035] Parameter definition: This indicates the effective shear strength of the structural surface under extreme conditions; Represents the normal stress on the structural surface; uu represents the pore water pressure within the structural surface under heavy rain conditions; Φ represents the internal friction angle of the structural surface; c represents the cohesion of the structural surface; Indicates the horizontal influence coefficient of earthquakes; Indicates the peak horizontal seismic acceleration; The instability mode discrimination index integrates the structural surface matching features of stereographic projection and the displacement evolution features of discrete elements to construct a quantitative discrimination index for instability modes:

[0036] Parameter definition: The composite index indicates the instability mode identification. This represents the degree of matching between the dominant structural surface and the free surface in stereographic projection, with a value ranging from 0 to 1. The displacement mutation rate of the unstable rock mass in the discrete element simulation is represented by the ratio of the maximum displacement increment to the initial displacement. Represents the weighting coefficients for stereographic projection analysis; This represents the weighting coefficients in discrete element numerical simulation.

[0037] S5: Multi-factor coupled stability classification evaluation of unstable rock masses By integrating the instability mode conclusions from S4 with the on-site environmental monitoring data collected in S1, and introducing key influencing factors such as rock mass integrity and groundwater seepage pressure, a multi-dimensional coupled evaluation system is constructed. The stability of the unstable rock mass is classified, and the classification results are pushed to the impact zone delineation stage in S6. Simultaneously, the characteristic data of high-risk unstable rock masses are fed back to S2 to update the training samples of the fusion analysis model and continuously optimize its identification capabilities.

[0038] In step S5, five key factors are introduced: rock mass integrity coefficient, groundwater seepage pressure, rainfall intensity, peak ground acceleration (PGA), and instability mode matching degree, to construct a multi-dimensional coupled evaluation system. The weights of each factor are determined using the analytic hierarchy process (AHP), with the rock mass integrity coefficient accounting for 0.3%, groundwater seepage pressure for 0.25%, instability mode matching degree for 0.2%, rainfall intensity for 0.15%, and PGA for 0.1%.

[0039] The stability of unstable rock masses is classified by combining the fuzzy comprehensive evaluation method. The membership threshold is set as follows: extremely high risk ≥ 0.8, high risk 0.6-0.8, medium risk 0.4-0.6, and low risk < 0.4. The classification results are divided into four levels: extremely high, high, medium, and low.

[0040] The grading results are pushed to the impact zone delineation stage in step S6. Simultaneously, data such as structural surface parameters and instability mode characteristics of high-risk rock masses are synchronized to the fusion analysis model in step S2. A model training update is automatically triggered every 50 newly added high-risk samples to improve the model's sensitivity in identifying high-risk areas. During the evaluation process, the time difference between data collection for each factor is strictly ensured to be no more than 24 hours to avoid affecting the accuracy of the evaluation results due to data lag. The matching degree between the evaluation system and the instability mode in step S4 is verified simultaneously, requiring a matching degree of over 90% before outputting the final grading results to ensure a high degree of consistency between the grading results and the actual risk of hazards.

[0041] S6: Dynamic Collapse Impact Zone Delineation and Data Loop Optimization Based on the stability grading results of S5, a rockfall trajectory simulation tool was used to delineate the impact range of different levels of landslides. The precise coordinates of high-risk impact areas were fed back to the UAV flight control system of S1, triggering secondary re-photographing of key areas and iterative updates to the model, forming a complete working loop of "data acquisition - intelligent identification - risk assessment - dynamic optimization", thereby improving the practicality and accuracy of the entire method.

[0042] In step S6, the stability classification results from step S5 are called, and a rockfall trajectory simulation tool is used. Combining high-precision topographic data of the slope and physical and mechanical parameters of the rock mass, the initial velocity of the falling rocks is set to 2-5 m / s, the rolling friction coefficient is 0.3-0.5, and the bounce recovery coefficient is 0.2-0.4. After 100 repeated simulations and taking the results of the 95% confidence interval, the collapse impact range corresponding to different stability levels is defined: the impact radius of the extremely high risk area is 30-50 meters, the high risk area is 15-30 meters, the medium risk area is 5-15 meters, and the low risk area is 0-5 meters.

[0043] The boundary coordinates of the high-risk impact area are pushed in real time to the UAV flight control system in step S1. The system automatically generates a secondary reshoot route for key areas. The flight altitude of the route is 30% lower than that of the initial full-domain shooting, and the aerial overlap is increased to 85% and the lateral overlap is increased to 75%. Priority is given to covering the core area of ​​the impact area and the surrounding 10-meter transition area to ensure that the local 3D model accuracy of the reshoot area is within 5 centimeters. After the reshoot is completed, a new local 3D reality model is generated, which automatically replaces the old data of the corresponding area in the original full-domain model, forming a complete closed loop from data acquisition to intelligent recognition and dynamic optimization.

[0044] During implementation, supplementary image capture triggering rules are set: supplementary images are automatically triggered every 7 days in extremely high-risk areas, every 15 days in high-risk areas, and monthly in medium- and low-risk areas. Simultaneously, the timing of supplementary images is adjusted according to weather conditions, requiring visibility to be no less than 5 kilometers and wind force no more than level 4 before execution, ensuring the clarity and accuracy of the supplementary data. The new model data generated by the supplementary images is synchronized to the structural surface extraction stage in step S3 to supplement the structural surface parameter samples in high-risk areas. The structural surface recognition algorithm is automatically updated every 30 new sets of high-risk structural surface data, further improving the practicality and accuracy of the entire method.

[0045] Example A rocky slope along a national highway in a mountainous area was selected as a case study. The slope is 62m high and has multiple sets of joints and fissures. It is prone to collapse due to rainstorms during the flood season and regional earthquakes. A five-lens consumer-grade oblique photography drone was used to execute the complete process of this method. S1. Multi-view oblique photogrammetry data acquisition and 3D modeling The drone's basic flight altitude was set at 80m, with 75% overlap in the heading direction and 65% overlap in the lateral direction. Wind speed, light intensity, and camera attitude parameters were recorded synchronously throughout the flight. Image features were extracted using the SIFT algorithm and then matched using a similarity formula with weighting coefficients α=0.8 and β=1.2. After constructing a sparse point cloud, a multi-constraint cost function was used to solve for disparity with λ=0.7 and γ=0.3, generating a centimeter-level dense point cloud. The surface was reconstructed using Poisson and the texture was mapped to obtain the overall 3D real-scene model with an overall accuracy of 8cm.

[0046] S2, Intelligent Pre-identification of Dangerous Rock Masses through Multimodal Feature Fusion Slope, curvature, and convexity geometric features are extracted from the 3D model; image texture features are extracted using the gray-level co-occurrence matrix; and flight wind speed and illumination are imported as environmental factors to construct a 3D feature library. Substituting the dynamic weight formula, the three types of feature weights are automatically assigned. The Sigmoid confidence model with neighborhood constraints is used to calculate the probability of hidden dangers grid by grid. Areas with confidence P≥0.6 are automatically marked as candidate areas of dangerous rock masses. The output boundary coordinates are then passed to S3.

[0047] S3, Extraction of Oriented Structural Surface Parameters The survey lines are automatically densified based on the candidate region contour, and the structural surface attitude is extracted by point cloud normal vector clustering: the dominant joint dip angle is 65°, the strike is 130°, and the fracture spacing is 0.4~1.2m; all attitude parameters are fed back to the S2 deep learning model to update the feature weights.

[0048] S4. Identification of Coupled Instability Modes under Extreme Operating Conditions (Rainstorm-Earthquake Coupled Instability Identification) The pore water pressure under heavy rainfall conditions is set to u = 0.22 MPa, and the peak horizontal seismic acceleration is a. g =0.15g, substitute into the shear strength correction formula to calculate the effective shear strength; Stereoscopic projection analysis yields the matching degree M between the dominant structural plane and the free plane. p =0.93; the rock mass displacement mutation rate R obtained by 3DEC discrete element simulation is 0.93. d =0.88; taking ω1=0.5 and ω2=0.5, the instability comprehensive index Is=0.905 is calculated, and it is judged as a toppling collapse.

[0049] S5. Multi-factor coupling stability classification assessment (four-level risk classification assessment) Five factors were collected: real-time rock mass integrity, groundwater pressure, rainfall in the past 7 days, regional seismic parameters, and instability matching degree. The comprehensive score was calculated by weighting the data and found to be 0.86, which meets the threshold of ≥0.8. The rock mass was then identified as an extremely high-risk unstable rock mass. This set of samples was then stored in the model training set.

[0050] S6. Dynamic landslide impact zone delineation and full-process data closed-loop optimization (rockfall simulation and closed-loop retesting) 100 rockfall trajectory simulations were performed, with a rockfall impact radius of 42m in the 95% confidence interval, and a warning control zone was delineated; the system automatically generated a reshoot flight path, with the flight altitude reduced to 56m (70% of the original altitude), 85% overlap in the heading direction and 75% overlap in the lateral direction, covering the core unstable rock mass of the slope and a 10m radius around it, with a local model accuracy of 4.5cm; This area is an extremely high-risk zone, with a 7-day retesting period. After the retesting is completed, the structural surface database will be updated, and the deep learning model will be automatically retrained after a total of 50 new high-risk data sets are added.

[0051] Unified Parameter Calibration Instructions: All formula weight coefficients and simulation parameters in this invention were determined through 10 sets of calibration tests on different rock slopes. The parameter value ranges are universal: α ranges from 0.5 to 1.0, β ranges from 1.0 to 1.5, λ ranges from 0.5 to 0.8, γ ranges from 0.2 to 0.4, and μ ranges from 0.2 to 0.4. The rockfall simulation parameters can be adjusted within a limited range according to the hardness of the rock mass.

[0052] Conditions for implementation of proportionality The three sets of comparative examples use the same test slope, unified UAV hardware, and unified geological evaluation standards as this invention. The test environment conditions, such as lighting, wind speed, and rainfall, are completely consistent, and the test data are objectively comparable.

[0053] Comparative Example 1: Traditional manual on-site measurement combined with experience-based judgment Three geological engineers were organized to go to the dangerous rock mass area with tools such as compasses and angle measuring instruments to manually measure the structural surfaces of the slope's free face point by point, record parameters such as strike and dip angle, analyze the combination of dominant structural surfaces through paper hand-drawn stereographic projection maps, and judge the instability mode and stability level by combining the engineers' on-site experience. After completing all the hazard identification work, the data was sorted out and a report was formed.

[0054] Comparative Example 2: Single Stereoscopic Projection Analysis Method A three-dimensional model of the unstable rock mass was obtained by using UAV oblique photography. The positions of the structural planes were marked on the model and the attitude parameters were read. The spatial relationship between the structural planes and the free surface was analyzed only by stereographic projection to determine the instability mode. Without combining discrete element numerical simulation or dynamic load correction, the stability level was directly divided and the influence range was defined based on the projection results.

[0055] Comparative Example 3: Static UAV Identification Method Without Dynamic Updates A full-area 3D model of the unstable rock mass was obtained by using UAV oblique photography. After extracting structural parameters, a mechanical analysis framework was built to simulate extreme working conditions, identify instability modes, and complete stability classification. However, no dynamic re-photographing mechanism was set up. Subsequent monitoring requires the re-conducting of full-area UAV flights and data collection, and local updates for high-risk areas are not possible.

[0056]

[0057] Parameter Calculation Method Explanation Structural surface parameter extraction time: The total time from starting data acquisition (or manual measurement) to outputting a complete structural surface parameter dataset, including the entire process time of data transmission, preprocessing, and parameter extraction.

[0058] Instability mode identification accuracy: The number of unstable rock mass cases that were correctly identified as having an instability mode was counted, divided by the total number of cases participating in the test, and converted into a percentage value. The identification results were based on actual collapse verification data on site.

[0059] Hazard identification and response cycle: The total time from receiving monitoring task instructions to finally outputting a stability grading report and impact range delineation results, covering all stages of data collection, analysis, and report preparation.

[0060] Stability classification fit: Five geological experts were invited to score the degree of match between the stability classification results of each method and the actual risk (out of 100 points). The average score was then converted into a percentage value. The higher the score, the better the fit between the classification results and the actual risk.

[0061] This invention, through multi-source data fusion analysis, can more accurately delineate potential hazard areas in complex slope environments, reducing ineffective analysis of risk-free rock masses and improving the efficiency of preliminary screening. It strengthens the correlation between different monitoring data, moving beyond reliance on UAV imagery or environmental monitoring data to integrate multi-dimensional information into a single analytical framework. This makes the pre-identification results more closely match the actual distribution of hazards, avoiding missed or incorrect assessments due to isolated data, and further enhancing the reliability of hazard identification.

[0062] This invention improves the adaptability to dangerous rock masses in complex terrain. Through a continuously updated fusion analysis model, the identification logic can be automatically adjusted for the characteristics of rock masses in different regions without the need for repeated manual calibration of analysis parameters, thus enhancing the practicality of the method in diverse scenarios.

[0063] In this invention, the connection with subsequent structural surface extraction and stability grading is also enhanced. The pre-identified candidate region data can be directly connected to the subsequent processing flow, reducing the steps of data conversion and re-analysis, improving the smoothness of the entire hazard identification process, and significantly enhancing the overall efficiency from data collection to result output.

[0064] Those skilled in the art should understand that the present invention is not limited to the specific embodiments described above. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for identifying potential rockfall hazards based on UAV oblique photography, characterized in that, Includes the following steps: S1. Multi-view oblique photography data acquisition and 3D modeling: A multi-lens UAV oblique photography system is used to take aerial photos of the target slope with full coverage, and flight environment parameters are collected and stored synchronously; an improved motion recovery structure algorithm is used to generate dense point clouds, and a 3D real scene model with bound spatial coordinates and camera attitude is obtained through surface reconstruction and texture mapping. The improved motion recovery structure algorithm includes image feature extraction, feature matching based on image overlap, sparse point cloud construction, and pixel disparity solution with multiple constraints to generate dense point cloud. S2. Multimodal feature fusion intelligent pre-identification of dangerous rock masses: Based on 3D real scene model, image and flight environment parameters, construct a multi-dimensional feature library of geometric shape, surface texture and environmental factors; adopt attention mechanism deep learning model to fuse multiple types of features, calculate the contribution weight of each type of feature through dynamic weight model, quantify the confidence of regional dangerous rock mass hidden danger, automatically delineate candidate areas of dangerous rock masses and output boundary coordinates; S3. Targeted structural surface parameter extraction: Adaptively deploy survey line network for the candidate area of ​​the dangerous rock mass, extract structural surface parameters based on point cloud normal vector clustering, and backfeed the parameters to the S2 deep learning model for iterative optimization of feature weights; S4. Extreme working condition coupled instability mode identification: A mechanical analysis framework is built based on structural surface parameters and geometric characteristics of unstable rock mass. The shear strength of the structural surface is corrected by simulating dynamic load conditions such as rainstorm and earthquake. The unstable rock mass instability mode is comprehensively determined by combining stereographic projection qualitative analysis and discrete element numerical simulation. S5. Multi-factor coupled stability classification evaluation: Select multiple influencing factors such as rock mass integrity, groundwater, rainfall, earthquake, and instability matching degree, and use hierarchical analysis and fuzzy comprehensive evaluation method to classify the risk level of dangerous rock mass. High-risk samples are synchronously updated to the S2 model training set. S6. Dynamic landslide impact zone delineation and full-process data closed-loop optimization: The impact range of landslides at different risk levels is divided by simulating the trajectory of falling rocks; a refined reshoot route is automatically generated based on the coordinates of high-risk areas; high-risk areas are periodically retested and modeled; the structural surface recognition database and deep learning model are updated to form a closed-loop recognition system of "collection-recognition-evaluation-iteration".

2. The method for identifying potential rockfall hazards based on UAV oblique photography according to claim 1, characterized in that: In step S1, feature matching based on image overlap is performed using a matching similarity formula: ； In the formula, S ij Let d be the similarity between corresponding feature points in the i-th image and the j-th image. ij O is the Euclidean distance between two feature point descriptors. ij Let s be the overlap between the i-th image and the j-th image. ij α represents the scale difference between the two feature points, and β represents the matching weight coefficients. Pixel disparity is calculated using a multi-constraint cost function that integrates photometric, geometric, word count, and point cloud density factors. ； In the formula, C(x,y,d) is the matching cost of pixel (x,y) under disparity d; Indicates the cost of photometric matching; For geometric matching cost; λ represents the sparse point cloud density of the region corresponding to pixel (x,y), and λ and γ are the balancing weight coefficients of the cost function.

3. The method for identifying potential rockfall hazards based on UAV oblique photography according to claim 1, characterized in that, The formula for calculating the feature contribution weight in step S2 is as follows: ； In the formula, n=3, corresponding to three types of features: geometric shape, surface texture, and environmental factors; The fusion weights for the k-th class features are... Let k be the inter-class discriminant of the feature class k. Let be the coefficient of variation of the k-th class of features; The formula for calculating the confidence level of potential rock mass hazards in the quantified area is as follows: ； In the formula, This represents the Sigmoid activation function. Let N(x,y,z) represent the normalized value of the k-th class feature at the 3D coordinates (x,y,z), and let N(x,y,z) represent the neighborhood consistency feature at the 3D coordinates (x,y,z). These are the weighting coefficients for neighborhood features.

4. The method for identifying potential rockfall hazards based on UAV oblique photography according to claim 1, characterized in that, The structural surface parameters mentioned in step S3 include the orientation, inclination angle, and spacing parameters of the structural surface.

5. The method for identifying potential rockfall hazards based on UAV oblique photography according to claim 1, characterized in that: Step S4: Construct a modified model of the effective shear strength of the structural surface coupled with pore water pressure and seismic load. ； In the formula, For the effective shear strength of the structural surface, For the normal stress of the structural surface, Let ρ be the pore water pressure, Φ be the internal friction angle of the structural surface, and c be the cohesion of the structural surface. This represents the horizontal influence coefficient of earthquakes. This represents the peak horizontal seismic acceleration. The type of instability is determined by fusing the results of the two types of analysis using an instability comprehensive discrimination index. ； In the formula, A comprehensive index for identifying instability modes. The degree of matching between the dominant structural surface and the free surface in stereographic projection. This represents the displacement mutation rate of the unstable rock mass in the discrete element simulation. These are the weighting coefficients for stereographic projection analysis. These are the weighting coefficients for discrete element numerical simulation; The discrete element numerical simulation uses 3DEC software to calculate the rock mass stress and displacement evolution.

6. The method for identifying potential rockfall hazards based on UAV oblique photography according to claim 1, characterized in that: Step S5 uses the analytic hierarchy process (AHP) to assign weights to five evaluation factors: rock mass integrity coefficient 0.3, groundwater seepage pressure 0.25, instability mode matching degree 0.2, rainfall intensity 0.15, and peak ground acceleration 0.1; and sets four risk membership thresholds: extremely high risk ≥0.8, high risk 0.6~0.8, medium risk 0.4~0.6, and low risk <0.

4. The grading results can only be output if the time difference between the collection of all evaluation factor data is ≤24h and the instability mode matching degree is ≥90%.

7. The method for identifying potential rockfall hazards based on UAV oblique photography according to claim 1, characterized in that, The simulation parameters for the falling rock trajectory in step S6 are: initial velocity of the falling rock 2~5m / s, rolling friction coefficient 0.3~0.5, and bounce recovery coefficient 0.2~0.4; repeat the simulation 100 times and use the 95% confidence interval to define the radius of influence: extremely high risk 30~50m, high risk 15~30m, medium risk 5~15m, and low risk 0~5m; the parameters for the reshoot flight path are: flight altitude reduced by 30% compared to the first time, directional overlap 85%, lateral overlap 75%, covering the core risk area and the surrounding 10m transition zone, and the accuracy of the local 3D model ≤5cm.

8. The method for identifying potential rockfall hazards based on UAV oblique photography according to claim 6, characterized in that, Step S6 sets differentiated reshoot cycles: reshoots are conducted every 7 days for extremely high-risk areas, every 15 days for high-risk areas, and every month for medium- and low-risk areas; aerial surveys are only performed when visibility is ≥5km and wind force is ≤4; Step S2 automatically triggers model retraining for every 50 new high-risk unstable rock mass samples, and Step S6 updates the structural surface recognition algorithm for every 30 new high-risk structural surface data.

Images