A highway slope modeling method based on unmanned aerial vehicle inspection route

Abstract

The application discloses a highway slope modeling method based on unmanned aerial vehicle (UAV) inspection route, and relates to the technical field of UAV inspection.The steps of the method comprise the following steps: measuring the length of a highway slope, topography and geomorphology and the complexity of geology, selecting a suitable UAV type and sensor combination according to the measurement; planning a UAV ground surface inspection and shallow geological exploration route, ensuring spatial matching, and synchronously acquiring high-resolution image and ground penetrating radar data; processing the image data through photogrammetry technology, generating a digital elevation model and an orthographic image map, and combining to build a ground surface three-dimensional model; performing geophysical inversion on the ground penetrating radar data, extracting reflection characteristics and projecting them to the ground surface three-dimensional model to establish the spatial correlation between the ground surface and the shallow structure; digitizing the reflection characteristics and geometrically reconstructing them into three-dimensional surface elements, embedding the three-dimensional surface elements into the ground surface model to form a three-dimensional geological body model, introducing a time dimension, and constructing a dynamic evolution model.The application realizes dynamic analysis of the stability of a slope according to the dynamic evolution model.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) inspection engineering, specifically a method for modeling highway slopes based on UAV inspection routes. Background Technology

[0002] As a common geological structure in various engineering constructions, slope stability directly affects the safe operation of engineering facilities and the safety of the surrounding environment and the lives and property of people. In construction scenarios such as highways, railways, and open-pit mines, geological disasters such as landslides and collapses caused by slope instability are common. These disasters can not only cause traffic disruptions and damage to engineering facilities, but also, in severe cases, lead to major casualties and huge economic losses.

[0003] However, traditional highway slope modeling mainly relies on manual on-site surveys, such as geological drilling and tape measurement, to obtain slope data. The accuracy and scope of data collection are limited, making it difficult to fully cover the complex terrain and geological structure of the slope. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for modeling highway slopes based on UAV inspection routes, which solves the problems of limited accuracy and scope of data acquisition in traditional methods, making it difficult to fully cover the complex terrain and geological structure of slopes.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for modeling highway slopes based on UAV inspection routes, comprising the following steps:

[0006] Step S1: Measure and conduct on-site investigation of the road slope to be measured to obtain the length, topographic features and geological complexity of the road slope to be measured. Based on the length, topographic features and geological complexity of the road slope to be measured, select the type of UAV and sensor combination.

[0007] Step S2: Design a spatial matching scheme between the UAV surface information inspection route and the shallow geological exploration scanning route. Based on the spatial matching scheme and sensor combination, the UAV simultaneously acquires high-resolution image data and ground-penetrating radar data, wherein the ground-penetrating radar data includes the coordinates collected by the ground-penetrating radar.

[0008] Step S3: Process the high-resolution image data using photogrammetry to generate a digital elevation model and orthophoto map of the highway slope. Combine the digital elevation model and orthophoto map using 3D modeling software to obtain a 3D model of the highway slope surface.

[0009] Step S4: Use geophysical inversion method to obtain the reflection features in the ground penetrating radar data. Project the reflection features onto the three-dimensional model of the roadside slope surface according to the coordinates collected by the ground penetrating radar for spatial registration and mutual verification to obtain the spatial logical connection between the surface and the shallow interior.

[0010] Step S5: Based on the spatial logical connection between the surface and the shallow interior, the reflection features are digitized and geometrically reconstructed using 3D modeling software to obtain 3D surface elements representing the reflection features. These 3D surface elements representing the reflection features are then embedded into the 3D surface model of the highway slope to form a 3D geological model of the highway slope.

[0011] Step S6: By introducing the time dimension, high-resolution image data and ground-penetrating radar data are repeatedly acquired to update the three-dimensional geological model of the highway slope and construct a dynamic evolution model of the highway slope.

[0012] Preferably, the selection of UAV type and sensor combination based on the length, topographic features, and geological complexity of the roadside slope to be measured includes:

[0013] Select the type of drone and the combination of sensors it carries:

[0014] Combination 1: Long-endurance fixed-wing UAV + medium-to-high precision visible light camera + lightweight ground-penetrating radar;

[0015] Applicable scenarios: Slope length <1000 meters, relatively regular terrain; Geological conditions: Geological complexity level C / D;

[0016] Drone type and sensor configuration:

[0017] Drones: Utilizing long-endurance fixed-wing aircraft, with a 24-hour flight time, providing efficient coverage over a wide area;

[0018] Sensors: Medium-to-high precision visible light camera: Select a medium-to-high precision visible light camera with 16-20 megapixels; Integrated lightweight ground-penetrating radar: Select a lightweight ground-penetrating radar with 800-1000MHz, which can detect shallow water-bearing areas and loose overburden layers, meeting the basic needs of shallow geological structure detection.

[0019] Combination 2: Large multi-rotor UAV + high-resolution visible light camera + ultra-lightweight ground-penetrating radar:

[0020] Applicable scenarios: Slope length: 500~1000 meters, complex terrain; Geological conditions: Geological complexity level B / C;

[0021] Drone and sensor configuration:

[0022] Drone: A hexacopter drone is selected, with a hovering accuracy of ±0.1 meters, supporting low-altitude slope-skimming flight;

[0023] Sensors: Visible light camera: Equipped with an aerial survey camera of over 20 megapixels, working in conjunction with the POS system to generate a DOM with centimeter-level resolution, accurately identifying surface cracks and weathering;

[0024] Ultralight Ground Penetrating Radar (GPR): Integrated with a 1000MHz high-frequency antenna, weighing less than 1kg, with a detection depth of 3-5 meters, suitable for shallow water-bearing areas and loose overburden detection;

[0025] Combination 3: Vertical Take-Off and Landing Compound Wing UAV + High-Resolution Visible Light Camera + Small Ground Penetrating Radar

[0026] Applicable scenarios: edge length < 800 meters, mixed terrain; geological conditions: geological complexity level A / B / C;

[0027] Drone and sensor configuration:

[0028] The drone uses a vertical take-off and landing hybrid wing, which combines the endurance of a fixed wing with the take-off and landing flexibility of a multi-rotor, making it suitable for complex airspace in mountainous areas.

[0029] Sensors: Visible light camera is the same as that of the combined system, ensuring three-dimensional accuracy of the Earth's surface; Small GPR: Integrated 500MHz antenna, detection depth 5~8 meters;

[0030] Preferably, the simultaneous acquisition of high-resolution image data and ground-penetrating radar data by the UAV based on the spatial matching scheme and sensor combination includes:

[0031] Spatial matching scheme:

[0032] Surface information inspection route:

[0033] Principles of direction:

[0034] If the geological structure of the slope is known, such as if the on-site investigation report clearly states that the slope orientation is north-south, the drone equipped with a visible light camera is parallel to the slope orientation, with the orientation being east-west, to ensure continuous coverage of the three-dimensional morphology of the slope.

[0035] If the geological structure of the slope is unknown and the on-site survey report cannot clearly define the geological structure of the slope, the direction of the surface information inspection route can be determined based on the geological survey report during the highway design stage. For example, if the slope runs east-west along the highway, the surface information inspection route is designed to run back and forth in an east-west direction.

[0036] Key parameters:

[0037] Flight altitude: The UAV equipped with a visible light camera should be 100-200m above the slope to ensure image resolution ≤5cm / pixel; ensure sufficient image overlap, with lateral overlap ≥60% and longitudinal overlap ≥80%;

[0038] Shallow geological exploration scanning route:

[0039] Principles of direction:

[0040] If the geological structure of the slope is known, the UAV equipped with ground-penetrating radar (GPR) is perpendicular to the geological structure, making the GPR reflection wave group clearer. The reflection signal is strongest when the structure is cut perpendicularly.

[0041] If the geological structure orientation is unknown, it is assumed to be at a 30° angle to the surface information inspection route. The surface information inspection route is the surface information inspection route with the unknown geological structure orientation of the slope mentioned above, which intersects the surface information inspection route at a 30° angle and is densified laterally along the slope.

[0042] Key parameters: Shallow geological exploration scanning route spacing: 5-10 meters for sections with geological complexity A / B; 10-20 meters for sections with moderate geological complexity C; 20-30 meters for sections with simple geological complexity D; flight altitude: 10-20 cm above the slope.

[0043] Trigger mode synchronizes GNSS positioning, and records a set of GPR data every 1-2cm to ensure the spatial resolution of underground structures;

[0044] Synchronous data acquisition: Through integrated collaborative inspection path planning, near-synchronous acquisition of surface information and shallow geological information is achieved, reducing data inconsistency caused by time differences. This ensures that while the UAV acquires surface morphology data, GPR can simultaneously acquire ground radar data of shallow slopes, which can indirectly reflect underground structural characteristics such as soil and rock stratification, water content, potential cavities, or loose areas.

[0045] Preferably, the generation of the digital elevation model and orthophoto map of the highway slope includes:

[0046] Digital elevation model generation:

[0047] Initialize the cloth mesh:

[0048] The process of separating ground points and non-ground points from a DSM typically employs filtering algorithms. This invention uses cloth simulation filtering to remove vegetation and building features, resulting in a point cloud that contains only the terrain surface.

[0049] Define a virtual cloth mesh on the DSM, and set the initial elevation of the mesh points to a fixed value higher than the highest point of the DSM to ensure that the cloth is initially above the DSM;

[0050] Define physical parameters:

[0051] Define the physical properties of the fabric, including virtual mass and elasticity coefficient. These parameters will affect the fabric's motion behavior during the iteration process.

[0052] Iterative calculation:

[0053] Calculate the resultant force:

[0054] For each point in the cloth mesh, calculate the gravity acting on it and the tension from the surrounding points;

[0055] The formula for calculating gravity is:

[0056]

[0057] in, It is gravity, m is virtual mass, and g is gravitational acceleration;

[0058] The calculation of tension is relatively complex, as it depends on the relative positions and distances between adjacent grid points. Hooke's law is used to approximate the calculation of tension, i.e.

[0059]

[0060] in, It is the tension from surrounding points, and k is the elastic coefficient. It represents the change in distance between adjacent points;

[0061] Calculate acceleration, velocity, and position:

[0062] According to Newton's second law Calculate the acceleration at each point using the following formula:

[0063]

[0064] in, It is the acceleration at each point. It is gravity. It is the tension from surrounding points, and m is the virtual mass;

[0065] The velocity of each point is updated based on acceleration, and the velocity update formula is:

[0066]

[0067] in, The updated speed It is the acceleration at each point. The velocity before the update, where n is the number of iterations and Δt is the time step;

[0068] The position of each point is updated based on the velocity; the position update formula is:

[0069]

[0070] in, It represents the updated position of each point, where Δt is the time step. This is the position before the update. This is the updated speed;

[0071] Convergence criteria:

[0072] After each iteration, check the position changes of the cloth mesh points. If the position changes of all points are less than a certain threshold, then the iteration is considered to have converged.

[0073] Extract ground points

[0074] Once the iteration converges, the points where the cloth adheres to the terrain surface are the ground points. These ground points are extracted from the original DSM to form a point cloud containing only the terrain surface.

[0075] Interpolation and meshing:

[0076] The extracted ground point cloud is interpolated to generate a regular grid;

[0077] Each grid point is assigned an elevation value to form a digital elevation model (DEM), which can reflect the true elevation changes of the terrain.

[0078] Orthophoto generation:

[0079] Based on the generated digital elevation model (DEM) and the internal and external orientation elements of the imagery, the internal and external orientation elements of the imagery include the camera position, attitude, and camera intrinsic parameters.

[0080] Perform orthophoto transformation on each image to eliminate distortion caused by terrain undulations and image tilt;

[0081] The orthophoto image is mosaicked and stitched together, and the color differences at the stitching seams are processed to generate an orthophoto map, which is an orthophoto image with geographic coordinates.

[0082] Preferably, the step of combining the digital elevation model (DEM) and the orthophoto map (DOM) using 3D modeling software to obtain a 3D model of the highway slope surface includes:

[0083] Construction of a 3D model of the roadside slope surface:

[0084] Elevation grid foundation establishment:

[0085] Using a DEM as a terrain framework, a terrain surface is formed in three-dimensional space; the DEM only contains the elevation information of the terrain, removing the influence of the height of features such as vegetation and buildings; this makes it more accurate when performing pure terrain analysis.

[0086] Mapping and Visual Rendering:

[0087] By using the DOM as a texture layer and registering it with geographic coordinates, it is overlaid onto the elevation grid of the DEM, giving the 3D model the color and texture features of a realistic surface and improving the visualization accuracy of the model.

[0088] Geometric Reconstruction and Detail Integration:

[0089] In 3D modeling software, the elevation data of the DEM is used to divide the grid, and the texture information of the DOM is combined for surface rendering. For key features, the point cloud density information in the DSM can be used to enhance the detail depiction.

[0090] Preferably, the geophysical inversion method is used to obtain the reflection characteristics in the ground penetrating radar (GPR) data, including:

[0091] Using geophysical inversion methods, effective reflection wave groups, in-phase axes, and anomalous reflection regions in GPR are identified and extracted.

[0092] Input data: GPR data after filtering, gain, and offset preprocessing, including the intensity, phase, frequency, and propagation time information of the reflected wave signal;

[0093] Technical implementation process of the inversion algorithm:

[0094] First, the underground space is divided into... There are three small grid cells, each with an unknown dielectric constant, denoted as . ( );

[0095] Establish a system of equations:

[0096] According to the theory of GPR wave propagation in underground media, for each pair of transmit-receive antenna positions, the wave propagation time is... It can be represented as:

[0097]

[0098] in, It is a coefficient that represents the wave originating from the transmitting antenna. After grid cells Reaching the receiving antenna Path coefficients, Is it a wave in the grid cell? The propagation path length in Is it a wave in the grid cell? The speed of propagation in;

[0099] By rearranging the above equations, we can obtain information about the dielectric constant. For multiple transmit-receive antenna pairs, a set of equations can be obtained;

[0100] The equations are usually nonlinear. To facilitate solving, Taylor series expansion is used for linearization. This Taylor series expansion method is used to linearize each nonlinear equation.

[0101] After linearization, all equations can be written as a system of linear equations. ,in It is a coefficient matrix. It is an unknown vector. It is a vector of constant terms;

[0102] This linearization method allows the approximate solution dielectric constant vector to be obtained by solving the originally complex nonlinear equations using the method of solving linear equations. The dielectric constant vector ε plays a crucial role in geophysical inversion;

[0103] Location of abnormal reflection areas:

[0104] Using the dielectric constant anomaly vector obtained from the inversion, combined with geological experience, the following can be identified:

[0105] High dielectric constant region: water-rich areas or saturated soil and rock masses; Low dielectric constant region: cavities, dry sand or loose deposits; Region with drastic fluctuations in dielectric constant: uneven density of soil and rock masses or fracture zones;

[0106] Effective reflected wave group extraction:

[0107] When the dielectric constant vector is in a state of abrupt change, that is, when the dielectric constant vector of the grid point region suddenly increases or decreases from the next grid point region, it corresponds to the interface between soil and rock layers, such as the interface between sandstone and clay, or the top surface of bedrock.

[0108] In-phase axis tracking:

[0109] In-phase axis refers to the trajectory formed by connecting points with the same phase in the electromagnetic wave reflection signal recorded by ground penetrating radar. In layman's terms, it is the line formed by points that reach the same vibration state at the same time in a series of continuous reflected wave signals.

[0110] By performing three-dimensional spatial extension analysis on the continuous reflection wave sequence, the orientation and variation trend of the phase axis are obtained. Combined with the prior knowledge of geological structure, the orientation and variation trend of the phase axis are constrained to determine the spatial distribution of groundwater level and continuous weak interlayers, and their burial depth and dip are inverted.

[0111] By analyzing effective reflection wave groups, in-phase axes, and anomalous reflection areas, reflection characteristics are obtained. These reflection characteristics can indicate the interfaces between different soil and rock layers, groundwater levels, density variations, and shallow geological structure information of potential cavities or water-rich areas.

[0112] Preferably, the step of obtaining three-dimensional surface elements representing the reflection features after digitizing and geometrically reconstructing the reflection features using three-dimensional modeling software includes:

[0113] For reflection features, they are discretized into a series of points, and the final three-dimensional coordinates of the geological reflection features are obtained based on these points. Three-dimensional surface elements are constructed using a triangulation interpolation algorithm for three points. , and This allows the construction of a triangular element.

[0114] Preferably, the step of embedding the three-dimensional surface elements representing the reflection characteristics into the three-dimensional surface model of the highway slope to form a three-dimensional geological model of the highway slope includes:

[0115] The shallow geological reflection characteristics are precisely embedded into the three-dimensional surface model of the highway slope to form a three-dimensional geological model of the highway slope.

[0116] Ensure the embedding process conforms to geometric constraints:

[0117] Layer continuity: For adjacent soil-rock layer boundaries, during the embedding process, it is necessary to ensure that the interface between adjacent layers is continuous; assuming the distance between two adjacent soil-rock layer boundaries in a certain direction is l, by adjusting the position of geological reflection features, so that... Approaching 0 ensures continuity at different levels;

[0118] Geological attributes and visualization parameters are assigned as follows:

[0119] Assigning properties to different soil and rock units: Based on the soil and rock type, assign corresponding physical properties to different geological reflection characteristics;

[0120] Visualization parameter settings: Set colors for geological reflection characteristics based on soil and rock types. Texture and transparency Visual parameters;

[0121] The final generated 3D geological model of the highway slope allows users to "see through" the interior of the slope by adjusting the transparency of different geological units, making arbitrary cuts, or partially hiding certain layers. This provides a deeper understanding of the spatial distribution of the main soil and rock layers, the location and shape of potential weak structural surfaces, and provides a cognitive depth for slope stability analysis that far exceeds that of traditional two-dimensional or simple surface models.

[0122] Preferably, the step of updating the three-dimensional geological model of the highway slope by introducing a time dimension and repeatedly acquiring high-resolution image data and ground-penetrating radar data includes:

[0123] Duplicate data acquisition and model update

[0124] Data Acquisition:

[0125] Repeatedly acquire the following data: high-resolution imagery data and ground-penetrating radar data collected by UAVs;

[0126] Historical data series:

[0127] The three-dimensional geological model is updated based on the acquired data, and high-precision registration technology is used to accurately map data from different periods to a unified coordinate system.

[0128] The updated 3D geological models are organized chronologically, and key features are extracted to form a historical data sequence. By introducing the time dimension, data is collected at fixed time intervals. Each collected 3D geological model is used as a time slice to construct a dataset containing the time dimension, namely historical time series data, which provides historical information for model input.

[0129] Preferably, the constructed dynamic evolution model of the highway slope includes:

[0130] Model architecture design:

[0131] Input layer: The input consists of future time period information and historical time series data; key features of the 3D geological model are extracted and arranged in chronological order as input features;

[0132] LSTM layer: Multiple LSTM units are set up to capture the time dependencies and complex evolution patterns in historical data; LSTM units can effectively process long-term time series data and remember the state changes of slope geological bodies at different times;

[0133] Intermediate layer: Add a fully connected layer to further extract and fuse the features output by the LSTM layer, in preparation for outputting a future 3D geological model;

[0134] Output layer: Outputs a 3D geological model for a specified future time period, including information such as the spatial distribution of each rock layer, changes in groundwater level, and fracture development;

[0135] Model training:

[0136] Training data partitioning: The preprocessed historical data is divided into training set, validation set and test set, generally in a ratio of 7:1:2;

[0137] Loss function and optimizer: Mean squared error is used as the loss function to measure the difference between the predicted future 3D geological body model and the actual situation; the Adam optimizer is used to update model parameters and accelerate training convergence speed.

[0138] Training process: Through continuous iterative training, the model parameters are adjusted to minimize the loss of the model on the validation set, thereby improving the model's prediction accuracy and generalization ability.

[0139] This invention provides a method for modeling highway slopes based on UAV inspection routes, involving machine learning and deep learning technologies, which has the following beneficial effects:

[0140] (1) This method for modeling highway slopes based on UAV inspection routes customizes the combination of UAVs and sensors according to the actual situation of the slope and designs a spatially matched data acquisition scheme to simultaneously acquire high-resolution images and ground-penetrating radar data. This multi-source data acquisition method can capture the subtle features of the slope surface and detect shallow underground structures, which greatly enriches the basic information for model construction compared with traditional single data sources.

[0141] (2) This method for modeling highway slopes based on UAV inspection routes establishes a spatial logical connection between the surface and shallow interior by spatially registering and fusing ground-penetrating radar reflection characteristics with a three-dimensional surface model. This overcomes the limitations of traditional modeling, which only focuses on a single dimension of the surface or underground. This method enables researchers to comprehensively analyze the correlation between slope surface deformation and underground structural changes from a three-dimensional spatial perspective. For example, it allows for the rapid location of the correspondence between surface cracks and underground cavities, providing an intuitive and comprehensive analytical tool for in-depth exploration of slope instability mechanisms and helping to more accurately assess slope stability.

[0142] (3) The highway slope modeling method based on UAV inspection route introduces a time dimension to construct a dynamic evolution model of highway slope through periodic repeated data collection and model update. This feature breaks the constraints of traditional static modeling and can track the slope change trend over time in real time, capture the subtle changes of slope caused by factors such as rainfall and earthquakes in a timely manner. Managers can predict the future state of the slope based on the dynamic model, formulate targeted protection measures in advance, realize dynamic control of slope risk, effectively reduce the probability of geological disasters, ensure the safe operation of highways, and reduce the economic losses and casualties caused by slope instability. Attached Figure Description

[0143] Figure 1 This is a flowchart of a highway slope modeling method based on UAV inspection routes proposed in this invention.

[0144] Figure 2 This is a hierarchical diagram of a highway slope modeling method based on UAV inspection routes proposed in this invention. Detailed Implementation

[0145] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0146] Please see Figure 1 This invention provides a technical solution: a method for modeling highway slopes based on UAV inspection routes. Specifically, the following method for modeling highway slopes based on UAV inspection routes is provided; please refer to [link / reference]. Figure 1 The method includes the following steps:

[0147] Step S1: Measure and conduct on-site survey of the road slope to be measured to obtain the length, topographic features and geological complexity of the road slope to be measured. Based on the length, topographic features and geological complexity of the road slope to be measured, select the type of UAV and sensor combination.

[0148] Preliminary measurements were taken of the roadside slope to be measured.

[0149] The length of highway slopes is measured using ground surveying methods. First, the start and end points of the slope need to be determined, usually based on the actual site conditions. Traditional surveying tools are used for direct measurement. During the measurement process, the start and end points, as well as key points, need to be marked on the ground for recording and subsequent analysis. Measurement points are determined as follows: control points are set up every 20-50m along the slope direction; in areas with drastic terrain changes, the density is increased to 5-10m. Key points are determined through on-site surveys combined with the geological survey report from the highway design phase, including the slope start / end points, geological structural inflection points, areas with dense surface cracks, and the location of drainage facilities. These key points serve as the benchmarks for the surveying network, controlling the overall measurement accuracy. The area covered by the measurement points is used to construct a continuous terrain model through interpolation algorithms. During the measurement process, detailed data for each measurement point needs to be recorded, including the measured length and location. Considering topographical factors, in sloping sections, the horizontal distance from the side stake to the center stake varies with the ground slope. Therefore, the measurement method needs to be adjusted according to the topographical changes (for example, laying out the distance from the toe of the slope to the center stake along the cross-section to determine the roadbed side stakes. In the actual measurement process, the point-by-point approach method is used, and calibration is performed on-site while measuring). The length data of all measurement points are accumulated to obtain the total length of the slope.

[0150] By combining on-site investigation and geological survey reports from the highway design phase, the general topographic and geomorphological characteristics of the highway slope to be measured are obtained. On-site investigation refers to surveyors directly observing the highway slope to obtain intuitive geological information, including the distribution of strata, rock properties, and surface morphology. During the highway design phase, detailed geological survey reports are typically prepared. These reports include slope topography and geomorphology, stratigraphic structure, physical and mechanical properties of soil and rock, rock mass structure types and structural surface characteristics, groundwater and surface water development and corrosivity, and the development of adverse geological conditions and special soil and rock types. A detailed analysis of the collected geological survey reports is conducted to extract information on relevant topographic and geomorphological features. By comprehensively utilizing the information from the on-site investigation and geological survey reports, the topographic and geomorphological characteristics of the highway slope to be measured are obtained.

[0151] Expected geological complexity: The topographic features are analyzed by professional geologists and classified according to the geological complexity grading standard. The geological complexity level is divided into four levels: A, B, C, and D, which correspond to complex, relatively complex, moderately complex, and simple, respectively.

[0152] Based on the length of the roadside slope to be measured, the general topographic features, and the expected geological complexity, the type of drone and the combination of sensors to be carried out were initially selected by precisely matching the equipment performance with the requirements of the scenario, achieving a balance between cost and efficiency while meeting technical specifications.

[0153] Combination 1: Long-endurance fixed-wing UAV + medium-to-high precision visible light camera + lightweight ground-penetrating radar:

[0154] Applicable scenarios: slope length 1000 meters, relatively regular terrain (e.g., long, gentle slopes, few gullies); Geological conditions: C / D level of geological complexity (moderately complex / simple, requires detection of underground faults and cavities).

[0155] Drone type and sensor configuration:

[0156] Drones: Utilizing long-endurance fixed-wing aircraft, with a 24-hour flight time, providing efficient coverage over a wide area;

[0157] sensor:

[0158] Medium-to-high resolution visible light camera: Choose a medium-to-high resolution visible light camera with 16-20 megapixels (such as the low-resolution mode of the Nikon D850). This can reduce the amount of data while maintaining a certain resolution.

[0159] Integrated lightweight ground-penetrating radar (GPR): Selecting a lightweight ground-penetrating radar with a frequency of 800-1000MHz (such as the GSSISIR-3000 system with an 800MHz antenna, which typically weighs around 1.2-1.5kg and has a detection depth of 2.5-4.5 meters) can detect shallow (2.5-4.5 meters deep) aquifers and loose overburden, meeting the basic requirements for shallow geological structure detection.

[0160] Combination 2: Large multi-rotor UAV + high-resolution visible light camera + ultra-lightweight ground-penetrating radar:

[0161] Applicable scenarios: Slope length: 500~800 meters, complex terrain (such as steep slopes, dense gullies, high and steep retaining walls); Geological conditions: Geological complexity level B / C (relatively complex / moderately complex, focusing on cracks and loose layers 3-5 meters below the surface).

[0162] Drone and sensor configuration:

[0163] Drones: Select hexacopter drones (such as DJI Matrice 600 Pro, XAG P1000), with hovering accuracy of ±0.1 meters, supporting low-altitude slope-hugging flight (5-20 meters from the slope).

[0164] Sensors: Visible light camera: Equipped with an aerial survey camera of 20 megapixels or higher (such as a modified Sony A7R4), in conjunction with the POS system, it generates a DOM with centimeter-level resolution to accurately identify surface cracks and weathering.

[0165] Ultralight Ground Penetrating Radar (GPR): Integrates a 1000MHz high-frequency antenna (such as the MALACX10 miniature radar), weighs less than 1kg, and has a detection depth of 3-5 meters, making it suitable for shallow water-bearing areas and loose overburden detection.

[0166] Combination 3: Vertical Take-Off and Landing Compound Wing UAV + High-Resolution Visible Light Camera + Small Ground Penetrating Radar (GPR):

[0167] Applicable scenarios: slope length < 800 meters, mixed terrain (open area + steep gully / high cliff); geological conditions: geological complexity level A / B / C (full scene coverage, taking into account long distance and complex terrain detection).

[0168] Drone and sensor configuration:

[0169] Unmanned aerial vehicles (UAVs) are equipped with vertical take-off and landing composite wings (such as Aerospace Rainbow CH-817 and Guandian Defense GD-100), which combine the endurance of fixed wings (35 hours) with the take-off and landing flexibility of multi-rotor UAVs to adapt to complex airspace in mountainous areas.

[0170] Sensors: Visible light camera is the same as that in combination two, ensuring three-dimensional accuracy of the ground surface; Small GPR: Integrated 10MHz and below low frequency antenna, detection depth 5~8 meters.

[0171] First-level matching rule: Prioritization of slope length

[0172] Prioritize selecting the combination that best covers the slope length and is closest to it (to avoid efficiency loss due to over-range detection):

[0173] For slope lengths < 500 meters: prioritize combination 3. The coverage radius of a single drone in combination 3 is ≤ 500 meters, enabling precise detection with "no blind spots and no redundancy". The applicable scenarios are closest to the slope length.

[0174] For slope lengths of 500-800 meters: prioritize combination 2 or combination 3, as the applicable scenarios for combinations 2 and 3 not only fully cover the slope length, but also cover a range of scenarios close to the 500-800 meter slope length.

[0175] For slope lengths of 800-1000 meters: prioritize combination 2 or combination 1, as the applicable scenarios for combinations 2 and 1 not only fully cover the slope length, but also cover a range of applicable scenarios close to the 800-1000 meter slope length.

[0176] slope length 1000 meters: Multiple drones were used for zoned data collection; zoning principle: the slope was divided into independent sub-zones according to the boundary of geological unit or abrupt topographic change zone (such as gully separation), and the slope length of each sub-zone was <1000 meters; the combination selection in each sub-zone was selected using the first-level matching rule and the second-level matching rule; data stitching was achieved through GNSS synchronization to avoid the limitations of single device endurance and coverage.

[0177] Secondary matching rules: Terrain / geology subdivision adaptation:

[0178] Secondary matching is enabled when matching multiple combinations of slope lengths, allowing for refined selection based on terrain complexity and geological exploration requirements:

[0179] Terrain complexity:

[0180] Regular terrain (long, gentle slopes, few gullies) → Combination 1 (fixed-wing long-endurance, suitable for large, regular areas);

[0181] Complex terrain (steep slopes, dense gullies, high and steep retaining walls) → Combination 2 (multi-rotor low-altitude slope-hugging, suitable for detailed exploration);

[0182] Mixed terrain (open areas + steep gullies / high cliffs) → Combination 3 (compound wing vertical take-off and landing, taking into account both long distances and complex terrain).

[0183] Geological exploration needs:

[0184] Requires deep exploration (faults, cavities, C / D level) → Combination 1 (lightweight radar covering deep layers);

[0185] Requires shallow, fine-grained detection (cracks, loose layers, B / C level) → Combination 2 (ultra-lightweight, high-resolution radar);

[0186] Full-scene coverage (A / B / C level) → Combination 3 (small radar balances depth and efficiency).

[0187] In the secondary matching, a priority determination is added: terrain complexity > geological exploration requirements. For example, if the terrain is regular but shallow fine exploration is required (such as crack monitoring), according to the priority determination, where regular terrain corresponds to combination 1 and shallow fine exploration is required to correspond to combination 2, combination 1 is selected first in the secondary matching.

[0188] Alternative solution:

[0189] When the detection requirements exceed the capabilities of a single sensor configuration (e.g., the detection requirements simultaneously need shallow fracture monitoring and deep ground-penetrating radar, and the capabilities of any combination of sensors are insufficient to meet the detection needs), the selection should be based on the principle of "performance coverage + lowest cost":

[0190] Performance coverage requirements:

[0191] Slope length: Select the minimum combination that covers the target length;

[0192] Terrain / Geology: Select the simplest configuration of sensors that can cover the requirements. For example, if only shallow detection is required, do not select deep radar.

[0193] Lowest cost basis:

[0194] Combination 3 (compound wing) > Combination 2 (multirotor) > Combination 1 (fixed wing). Due to the high technical complexity of compound wings, multirotors are next, while fixed wings have low operating costs. Priority should be given to combinations with low platform costs, such as multirotors replacing compound wings and fixed wings replacing multirotors.

[0195] Step S2: Design a spatial matching scheme for the UAV surface information inspection route and the shallow geological exploration scanning route. Based on the spatial matching scheme and sensor combination, the UAV simultaneously acquires high-resolution image data and ground-penetrating radar data, wherein the ground-penetrating radar data includes the coordinates collected by the ground-penetrating radar.

[0196] Spatial matching scheme:

[0197] Surface information inspection route:

[0198] Principles of direction:

[0199] If the geological structure of the slope is known, such as if the on-site investigation report clearly states that the slope direction is north-south, the drone equipped with a visible light camera is parallel to the slope direction, with the direction being south to north, to ensure continuous coverage of the three-dimensional morphology of the slope.

[0200] If the geological structure of the slope is unknown and the on-site survey report cannot clearly define the geological structure of the slope, the direction of the surface information inspection route can be determined based on the geological survey report during the highway design stage. For example, if the slope runs east-west along the highway, the surface information inspection route is designed to run back and forth in an east-west direction.

[0201] Key parameters:

[0202] Flight altitude: The UAV equipped with a visible light camera is 100~200m above the slope to ensure image resolution ≤5cm / pixel; ensure sufficient image overlap, with lateral overlap ≥60% and longitudinal overlap ≥80% (to ensure the integrity of 3D reconstruction);

[0203] Shallow geological exploration scanning route:

[0204] Principles of direction:

[0205] If the geological structure of the slope is known, the UAV equipped with ground-penetrating radar (GPR) is perpendicular to the geological structure (such as fault joint surfaces) to make the GPR reflection wave group clearer. The reflection signal is strongest when the structure is cut perpendicularly.

[0206] If the geological structure orientation is unknown, it is assumed to be at a 30° angle to the surface information inspection route. The surface information inspection route is the surface information inspection route with the unknown geological structure orientation of the slope mentioned above, and it intersects the surface information inspection route at a 30° angle (the 30° intersection can increase the oblique coverage of the survey line and the slope orientation within a limited width, while shortening the length of a single survey line (compared to a vertical intersection), avoiding overlapping flight trajectories or data redundancy caused by the narrow space of the UAV), and is densified laterally along the slope.

[0207] Key parameters: The spacing of the shallow geological exploration scanning route is 5-10 meters in sections with geological complexity A / B, 10-20 meters in sections with medium geological complexity C, and 20-30 meters in sections with simple geological complexity D.

[0208] The flight altitude is 10-20cm above the slope (GPR is for near-surface detection of low-altitude flight to improve signal penetration);

[0209] Trigger mode synchronizes GNSS positioning, and records a set of GPR data every 1-2cm to ensure the spatial resolution of underground structures.

[0210] Synchronous data acquisition: Through integrated collaborative inspection path planning, near-synchronous acquisition of surface information and shallow geological information is achieved, reducing data inconsistency caused by time differences. This ensures that while the UAV acquires surface morphology data (high-resolution image data), GPR can simultaneously acquire ground radar data (dielectric property data) of the shallow slope (a few meters to tens of meters below the surface), which can indirectly reflect underground structural characteristics such as soil and rock stratification, water content, potential cavities, or loose areas.

[0211] Simultaneously, data from both the sensor itself and the acquired data are also recorded: the focal length of the visible light camera, the internal and external orientation elements of the image, including the camera's position and attitude (such as yaw angle, pitch angle, roll angle), and the camera's intrinsic parameters (such as focal length and principal point coordinates); and the coordinates of the detection point of the radar. These coordinates, along with the depth and angle of the GPR during the probe, are obtained via the GPS satellite positioning system.

[0212] Step S3: Process the high-resolution image data using photogrammetry to generate a digital elevation model (DEM) and an orthophoto map of the highway slope. Combine the DEM and the orthophoto map (DOM) using 3D modeling software to obtain a 3D model of the highway slope surface.

[0213] Image screening and preprocessing:

[0214] When processing the acquired high-resolution images, a rigorous screening process is essential. Staff must review each image individually, eliminating those that are blurry. For example, images where the drone's movement during flight, such as shaking or inaccurate focusing, results in unclear footage. Overexposed images must also be removed; these are often taken in excessively bright light, resulting in a bright, washed-out image that loses much detail. Furthermore, underexposed images must be excluded; these are typically taken in low-light conditions, making the overall image dark and obscuring ground features.

[0215] After image screening, preprocessing is performed on the selected images. Due to the inherent optical characteristics of UAV lenses, the captured images may exhibit geometric distortions, such as barrel distortion or pincushion distortion. To address this, the team uses software to analyze the lens imaging model, determine the distortion parameters, and then reposition and correct each pixel in the image to obtain a processed high-resolution image. This high-resolution image accurately reflects the shape and location of actual ground features, effectively improving image quality and laying a solid foundation for subsequent feature extraction, matching, and 3D modeling.

[0216] Feature point extraction:

[0217] Feature points are automatically extracted from each processed high-resolution image using photogrammetric SfM technology:

[0218] First, the input high-resolution drone image is processed by convolving the processed high-resolution image with Gaussian kernels of different scales to obtain a series of images at different scales. Let the original image be... Gaussian kernel is ,in If it is a scale factor, then at a certain scale The image below It can be represented as By changing To obtain images at different scales, among which It is a scale The scale factor.

[0219] Detecting extreme points:

[0220] In the constructed scale space, extreme points are found by comparing the difference of Gaussians (DoG) between each point and its surrounding neighbors (the neighborhood of a 3D cube naturally includes 26 neighboring points in all directions). If a point's value is greater than or less than the values ​​of all its neighbors, then that point is considered a feature point.

[0221] Difference of Gaussians (DoG) is the difference between two adjacent Gaussian blurred images. and If the images are two Gaussian blurred images of different scales, then the DoG is:

[0222] in, and These are two Gaussian blurred images at different scales. It is a processed high-resolution image. is a Gaussian kernel for two images at different scales, used to blur the image, where k is the scaling factor between adjacent scales.

[0223] Generate feature point descriptors:

[0224] For a given feature point, take a local region centered on the feature point and divide the region into several sub-regions. Use a histogram in 8 directions. Each sub-region has 4x4 pixel blocks. Thus, a feature point can be described by a 128-dimensional (4x4x8) vector, which is the descriptor of the feature point.

[0225] Image matching:

[0226] By comparing feature points in different images, we can find points with the same name (i.e., feature points that correspond to the same actual ground feature in different images):

[0227] For two processed high-resolution images A and B, feature points and their descriptors are extracted respectively, and corresponding points are found by comparing the distance between the descriptors.

[0228] Let the feature points in image A be... The descriptor is Feature points in image B The descriptor is Then the Euclidean distance between them

[0229]

[0230] Where d is the Euclidean distance between descriptors in images A and B, and n is the dimension of the descriptor, such as SIFT being 128-dimensional and SURF being 64-dimensional. .

[0231] Set a threshold If the Euclidean distance between two points Less than If they are, then they are considered to be points with the same name.

[0232] Ratio test:

[0233] To further improve the accuracy of matching, a ratio test is used. For a feature point in image A, the Euclidean distance between it and the descriptors of all feature points in image B is calculated.

[0234] To improve matching accuracy, the ratio is calculated by taking the nearest neighbor distance between feature points in image A and all feature point descriptors in image B. Distance to next nearest neighbor If the ratio of the two If the value is less than the set threshold, the corresponding point is determined to be a point with the same name.

[0235] Based on points with the same name, connections between images are constructed to form an image network.

[0236] Sparse point cloud generation:

[0237] Based on the correspondence of corresponding points, relative orientation is performed to determine the relative position and orientation relationship between adjacent images.

[0238] Bundle adjustment is used to optimize the entire image network as a whole. The least squares method is used to adjust the exterior orientation elements (position and orientation) of the image to minimize the reprojection error and obtain a high-precision sparse point cloud.

[0239] Relative orientation:

[0240] Based on pairs of corresponding points, the relative position and pose relationships (relative orientation elements) between adjacent images are determined using the coplanarity condition equation. The expression for the coplanarity condition equation F is:

[0241]

[0242] in, , , ) and the left and right ray vectors are coplanar. and Let be the coordinates of the corresponding point in the image space coordinate system. This refers to the camera's focal length.

[0243] By iteratively solving the equation, the relative position and orientation (relative orientation element) of adjacent images can be determined.

[0244] Using the relative orientation elements obtained from relative orientation, biangular repetition is performed on corresponding points to calculate their one-dimensional spatial coordinates, forming an initial sparse point cloud. This point cloud contains only the three-dimensional coordinates of a small number of feature points, which are used to construct spatial connectivity relationships between images.

[0245] Constraint Adjustment and Optimization:

[0246] Global adjustment model construction: Based on the initial sparse point cloud, a constrained adjustment model is established, with the relative orientation elements (position, pose) of all images and the 3D coordinates of the point cloud as unknowns, and collinearity equations as constraints.

[0247]

[0248] in, For image point coordinates, Let the principal point coordinates be... Coordinates of the photography center , These are the coordinates of the object point.

[0249] The above equations are solved using the least squares method. By iteratively adjusting the image exterior orientation elements and point cloud coordinates, the reprojection error (the deviation between actual and theoretical image points) is minimized, ultimately yielding a high-precision sparse point cloud. This point cloud possesses a unified three-dimensional coordinate system and is uniformly distributed, effectively representing the spatial location of features.

[0250] Dense point cloud generation

[0251] Multi-view stereo matching:

[0252] Based on high-precision sparse point clouds, dense stereo matching is performed between multiple images. By searching for corresponding points on the images and calculating the three-dimensional coordinates of each point, a dense point cloud is generated.

[0253] Digital Surface Model (DSM) Generation:

[0254] Point cloud interpolation and meshing:

[0255] First, interpolation is performed on the dense point cloud to generate a regular grid. The purpose of interpolation is to generate continuous data between these discrete points, and the grid size is on the order of centimeters.

[0256] Then, an elevation value is assigned to each grid point to form a digital surface model (DSM).

[0257] These elevation values ​​are calculated and assigned based on the height information in the point cloud data.

[0258] Suppose there are three points in the point cloud data. , and For grid points located within the triangle formed by these three points Its elevation value It can be calculated using the area weighting method:

[0259] Calculate the area of ​​the triangle: Total area of ​​the triangle

[0260] |

[0261] in, It is the total area of ​​the triangle. , , and , , These are the x-axis and y-axis coordinates of three points.

[0262] Calculate the area of ​​the sub-triangle: Let the area of ​​the sub-triangle formed by P, B, and C be... The area formed by A and C is The area formed by A and B is .

[0263] Weighting: Weight , , ;

[0264] Elevation calculation:

[0265]

[0266] in, It is an elevation value. , and Yes, the elevation values ​​of the three points in the point cloud data. , and These are the corresponding weighting coefficients.

[0267] The DSM contains height information for all features on the Earth's surface, such as vegetation and buildings. This means that the DSM not only reflects the height changes of the terrain itself, but also includes the heights of other objects on the surface, thus comprehensively depicting the actual shape of the Earth's surface.

[0268] Digital Elevation Model (DEM) Generation:

[0269] Initialize the cloth mesh:

[0270] The ground points and non-ground points are separated from the DSM. This process usually uses filtering algorithms. This invention uses cloth simulation filtering to remove ground features such as vegetation and buildings, resulting in a point cloud that only contains the terrain surface.

[0271] Define a virtual cloth grid on the DSM, and set the initial elevation of the grid points to a fixed value higher than the highest point of the DSM to ensure that the cloth is initially above the DSM.

[0272] Define physical parameters:

[0273] Define the physical properties of the fabric, including virtual mass (m) and elastic modulus (k). These parameters will affect the fabric's motion behavior during the iteration process.

[0274] Iterative calculation:

[0275] Calculate the resultant force:

[0276] For each point in the cloth mesh, calculate the gravity (Fg) acting on it and the tension (Ft) from the surrounding points.

[0277] The formula for calculating gravity is:

[0278]

[0279] in, It is gravity, m is virtual mass, and g is gravitational acceleration.

[0280] Calculating tension is complex, as it depends on the relative positions and distances between adjacent grid points. Hooke's law is typically used to approximate the calculation of tension.

[0281]

[0282] in, It is the tension from surrounding points, and k is the elastic coefficient. It represents the change in distance between adjacent points.

[0283] Calculate acceleration, velocity, and position:

[0284] According to Newton's second law Calculate the acceleration (a) at each point using the following formula:

[0285]

[0286] in, It is the acceleration at each point. It is gravity. It is the tension from the surrounding points, and m is the virtual mass.

[0287] The velocity (v) of each point is updated based on the acceleration. The velocity update formula is as follows:

[0288]

[0289] in, The updated speed It is the acceleration at each point. The velocity before the update, where n is the number of iterations and Δt is the time step.

[0290] The position (x) of each point is updated based on the velocity. The position update formula is:

[0291]

[0292] in, It represents the updated position of each point, where Δt is the time step. This is the position before the update. That's the updated speed.

[0293] Convergence criteria:

[0294] After each iteration, check the position changes of the cloth mesh points. Check if the position changes of all points are less than a certain threshold. If the position changes of all points are less than the threshold, the iteration is considered to have converged.

[0295] Extract ground points

[0296] Once the iteration converges, the points where the cloth adheres to the terrain surface are the ground points. These ground points are extracted from the original DSM to form a point cloud containing only the terrain surface.

[0297] Interpolation and meshing:

[0298] The extracted ground point cloud is interpolated to generate a regular grid.

[0299] Each grid point is assigned an elevation value (the elevation value calculation process is the same as that of DSM) to form a digital elevation model (DEM), which can reflect the true elevation changes of the terrain.

[0300] Orthophoto map (DOM) generation:

[0301] Orthographic projection and image mosaic:

[0302] Based on the generated digital elevation model (DEM) and the internal and external orientation elements of the imagery, the internal and external orientation elements of the imagery include the camera's position, attitude (such as heading angle, pitch angle, roll angle) and camera intrinsic parameters (such as focal length, principal point coordinates).

[0303] Each image is transformed by orthophoto projection to eliminate distortions caused by terrain undulations and image tilt.

[0304] The orthophoto images are mosaicked and stitched together, and the tonal differences at the stitching seams are processed to generate an orthophoto map (DOM), which is an orthophoto image with geographic coordinates.

[0305] It should be noted that when mosaicking the orthophoto-projected images, tonal differences may exist between adjacent images, requiring processing using a histogram-based matching method. For example, for two adjacent images... and Calculate their histograms and Then by adjusting The grayscale value makes and Match as closely as possible to minimize color differences.

[0306] Construction of a 3D model of the roadside slope surface:

[0307] Elevation grid foundation establishment:

[0308] Using a DEM as the terrain framework, a three-dimensional terrain surface is formed. The DEM contains only elevation information of the terrain, removing the height influence of features such as vegetation and buildings. This makes it more accurate when performing pure terrain analysis.

[0309] Mapping and Visual Rendering:

[0310] By using the DOM as a texture layer and registering it with geographic coordinates, it is overlaid onto the elevation grid of the DEM, giving the 3D model the color and texture features of a realistic surface and improving the visualization accuracy of the model.

[0311] Geometric Reconstruction and Detail Integration:

[0312] In 3D modeling software, elevation data from the DEM is used to create a mesh (such as a triangular mesh), which is then combined with texture information from the DOM for surface rendering. For key features (such as retaining walls and slope protection), point cloud density information from the DSM can be used to enhance detail rendering.

[0313] Step S4: Use geophysical inversion method to obtain the reflection characteristics in the ground penetrating radar data. Project the reflection characteristics onto the three-dimensional surface model of the highway slope according to the coordinates collected by the ground penetrating radar for spatial registration and mutual verification to obtain the spatial logical connection between the surface and the shallow interior.

[0314] The process of preprocessing the synchronously acquired ground-penetrating radar (GPR) data by filtering, gain adjustment, and offset is as follows:

[0315] Filtering: Bandpass filtering is used. Bandpass filtering removes noise by setting a frequency range, allowing signals of a specific frequency band to pass through while blocking high-frequency and low-frequency noise, and retaining the frequency range of effective reflected signals, thereby achieving noise removal.

[0316] Gain adjustment:

[0317] It compensates for the energy attenuation of electromagnetic waves as they propagate in the medium, enhances the recognizability of weak reflected signals, and dynamically adjusts the gain parameters according to the electromagnetic wave propagation time (i.e., detection depth) to apply higher gain to deep reflected signals, so as to keep the reflection intensity of shallow and deep parts balanced and facilitate overall interpretation.

[0318] Offset correction:

[0319] By repositioning the tilted or curved reflective interface to its true spatial location, image distortion caused by the detection angle is eliminated. Based on the principle of ray tracing, each reflection point is projected back along the electromagnetic wave propagation path to its true location, correcting the hyperbolic reflection characteristics and making the horizontal interface appear as a straight line, which facilitates the accurate tracking of geological interfaces.

[0320] Through the above steps, the signal-to-noise ratio and geometric fidelity of GPR data are significantly improved, laying the foundation for subsequent geophysical inversion.

[0321] Using geophysical inversion methods, effective reflection wave groups, in-phase axes, and anomalous reflection regions in GPR are identified and extracted.

[0322] Input data: GPR data after filtering, gain, and offset preprocessing, including information such as the intensity, phase, frequency, and propagation time of the reflected wave signal.

[0323] The core of the physical model: Based on the propagation laws of electromagnetic waves in different media, a correlation model of "dielectric constant difference - reflection characteristics" is established.

[0324] For example: the interface between soil and rock layers generates reflected wave groups due to a sudden change in dielectric constant; water has a high dielectric constant (about 80), and water-rich areas exhibit strong reflection; cavities have a dielectric constant close to 1 (air), resulting in a lack of reflection or weak reflection.

[0325] Technical implementation process of the inversion algorithm:

[0326] First, the underground space is divided into... Each grid cell is a small grid cell. The dielectric constant of each grid cell is unknown, and is set to... ( ).

[0327] Establish a system of equations:

[0328] According to the theory of GPR wave propagation in underground media, for each pair of transmit-receive antenna positions, the wave propagation time is... ( Indicates the transmitting antenna. (Representing the receiving antenna) can be represented as:

[0329]

[0330] in, It is a coefficient that represents the wave originating from the transmitting antenna. After grid cells Reaching the receiving antenna The path coefficient (1 if the wave passes through the cell, 0 otherwise). Is it a wave in the grid cell? The propagation path length in Is it a wave in the grid cell? The speed of propagation in It is the speed of light in a vacuum.

[0331] By rearranging the above equations, we can obtain information about the dielectric constant. The equations are given. For multiple transmit-receive antenna pairs, a system of equations can be obtained.

[0332] Equations are often nonlinear, and to facilitate solving them, methods such as Taylor series expansion are used for linearization. For example, for the function... (here It includes all (the vector), in the initial guess value Taylor expansion nearby:

[0333]

[0334] in, Is the function at the initial guess value? The value at that location, Is the function in The gradient (partial derivative vector) at that point. It is the difference vector between the current value and the initial guessed value.

[0335] For each nonlinear equation, this Taylor series expansion method is used for linearization.

[0336] After linearization, all equations can be written as a system of linear equations. ,in It is a coefficient matrix. It is an unknown vector (here, the dielectric constant vector). ), It is a vector of constant terms.

[0337] This linearization method allows the approximate solution dielectric constant vector to be obtained by solving the originally complex nonlinear equations using the method of solving linear equations. The dielectric constant vector ε plays a crucial role in geophysical inversion. It is not merely a simple numerical vector, but also contains information about the physical properties of the subsurface medium.

[0338] Location of abnormal reflection areas:

[0339] Identification is performed using the dielectric constant anomaly vector obtained from the inversion, combined with the dielectric constant difference-reflection feature correlation model:

[0340] High dielectric constant region ( >30): Water-rich areas or saturated soil and rock masses; low dielectric constant regions ( <5): Cavities, dry sand or loose deposits; areas with drastic fluctuations in dielectric constant: uneven density of soil and rock or fracture zones.

[0341] Effective reflected wave group extraction:

[0342] When the dielectric constant vector is in a state of abrupt change, that is, when the dielectric constant vector of the grid point region suddenly increases or decreases from the next grid point region, it corresponds to the interface between soil and rock layers, such as the interface between sandstone and clay, the top surface of bedrock, etc. (the specific layering is determined by the dielectric constant vector).

[0343] In-phase axis tracking:

[0344] In-phase axis refers to the trajectory formed by connecting points with the same phase in the electromagnetic wave reflection signal recorded by ground penetrating radar. In simpler terms, it is the line formed by points that reach the same vibration state (such as wave crest or trough) at the same moment in a series of continuous reflected wave signals.

[0345] By performing three-dimensional spatial extension analysis on the continuous reflection wave sequence, the orientation and variation trend of the phase axis are obtained. Combined with prior knowledge of geological structure, the orientation and variation trend of the phase axis are constrained to determine the spatial distribution of groundwater level, continuous weak interlayers, etc., and their burial depth and dip are inverted.

[0346] By analyzing effective reflection wave groups, in-phase axes, and anomalous reflection areas, reflection characteristics are obtained. These reflection characteristics can indicate shallow geological structure information such as the interfaces of different soil and rock layers, groundwater levels, density changes, potential cavities, or water-rich areas.

[0347] The reflection features obtained from the inversion (such as soil layer interfaces and groundwater levels) are projected onto the three-dimensional surface model of the highway slope based on the coordinates of the trajectory collected by GPR. Spatial registration and mutual verification are performed with surface geomorphological features (such as cracks, drainage outlets, and catchment areas) to establish a spatial logical connection between the surface and the shallow internal structure.

[0348] Spatial registration: When collecting data, GPR devices record their own location information (such as longitude, latitude, and elevation). Assuming the coordinates of a GPR detection point are... These coordinates are obtained through the GPS satellite positioning system.

[0349] Simultaneously, based on the detection depth and angle of the GPR, the relative coordinates of the reflection feature with respect to the GPR detection point are calculated. .

[0350] The specific calculation process is as follows:

[0351]

[0352]

[0353]

[0354] in,( ) are the coordinates of the GPR detection point, ( x, y, z) are the relative coordinates of the geological reflection features with respect to the GPR detection point. θ is the detection angle of the GPR, and d is the detection depth of the GPR.

[0355] Elevation assignment of 3D surface model for highway slope: The 3D surface model provides elevation information of the surface. *For each GPR detection point, the corresponding geological reflection feature, its final three-dimensional coordinates The calculation is as follows:

[0356]

[0357]

[0358]

[0359] in, The final three-dimensional coordinates of geological reflection features, ( x, y, z) is the relative coordinate of the geological reflection feature with respect to the GPR detection point. () are the coordinates of the GPR detection point. It is the surface height value obtained from the three-dimensional model of the roadside slope, used to determine the location of geological reflection features in the vertical direction.

[0360] The inverted reflection features are then registered with the 3D model of the highway slope surface according to the final 3D coordinates. This can be achieved by finding corresponding points in the data (i.e., points with obvious features that can be identified in both the reflection feature data and the 3D model of the highway slope surface). For example, the reflection feature points of surface cracks in the GPR data and the actual location points of the cracks in the 3D model of the highway slope surface can be selected as corresponding points.

[0361] Mutual verification:

[0362] Verification of cracks and water-rich areas:

[0363] If a strong GPR reflection anomaly is present beneath a surface fissure, it may indicate that the fissure extends into a water-rich area.

[0364] In specific analysis, after identifying cracks in the 3D model of the highway slope surface, the corresponding GPR reflection characteristic data are examined. If a region with a high dielectric constant (manifested as strong reflection in GPR data) appears within a certain depth below the crack, and the dielectric constant is greater than 30 (usually considered a characteristic value of water-rich areas), it can be determined that the crack may be connected to a water-rich area.

[0365] Verification of catchment area and groundwater level:

[0366] If there are continuous phase axes beneath the surface catchment area, it may correspond to a rise in the groundwater level.

[0367] After determining the location of the catchment area in the 3D model of the highway slope surface, observe the corresponding GPR reflection characteristics. If a continuous in-phase axis is found below the catchment area (the in-phase axis represents a continuous geological interface, and there will be a clear in-phase axis at the groundwater level), it can be inferred that the groundwater level has risen in this area.

[0368] By spatially registering and cross-validating the reflection features obtained through inversion with the three-dimensional model of the highway slope surface, a preliminary spatial logical connection between the surface and the shallow internal structure is established.

[0369] Step S5: Based on the spatial logical connection between the surface and the shallow interior, the reflection features are digitized and geometrically reconstructed using 3D modeling software to obtain 3D surface elements representing the reflection features. These 3D surface elements representing the reflection features are then embedded into the 3D surface model of the highway slope to form a 3D geological model of the highway slope.

[0370] The specific process of generating a 3D geological model of a highway slope:

[0371] Based on the spatial logical connection (spatial matching) between the surface and shallow interior, digital and geometric reconstruction is carried out in a 3D modeling software environment to form 3D surface elements representing these geological structures.

[0372] For reflection features, they are discretized into a series of points, and the final three-dimensional coordinates of the geological reflection features are obtained based on these points. Three-dimensional surface elements are constructed using a triangulation interpolation algorithm. For example, for three points... , and This allows the construction of a triangular element.

[0373] The shallow geological reflection characteristics are accurately embedded into the three-dimensional surface model of the highway slope to form a three-dimensional geological model of the highway slope.

[0374] Ensure the embedding process conforms to geometric constraints:

[0375] Layer continuity: For adjacent soil-rock layer boundaries, during the embedding process, it is necessary to ensure the continuity of the interface between adjacent layers. Assuming the distance between two adjacent soil-rock layer boundaries in a certain direction is *l*, the position of the geological reflection features is adjusted to ensure... Approaching 0 ensures continuity at different levels.

[0376] Geological attributes and visualization parameters are assigned as follows:

[0377] Assigning different properties to soil and rock units: Based on soil and rock types (such as sand, clay, rock, etc.), assign corresponding physical properties (such as density) to different geological reflection characteristics. Elastic modulus Poisson's ratio These attribute values ​​are derived from geological manuals or field test data.

[0378] Visualization parameter settings: Set colors for geological reflection characteristics based on soil and rock types. Texture and transparency Visual parameters, such as sand, can be set to light yellow. Clay can be set to brown. wait.

[0379] The final generated 3D geological model of the highway slope allows users to "see through" the interior of the slope by adjusting the transparency of different geological units, making arbitrary cuts, or partially hiding certain layers. This provides a deeper understanding of the spatial distribution of the main soil and rock layers, the location and shape of potential weak structural surfaces, and provides a cognitive depth for slope stability analysis that far exceeds that of traditional two-dimensional or simple surface models.

[0380] Step S6: By introducing the time dimension, high-resolution image data and ground-penetrating radar data are repeatedly acquired to update the three-dimensional geological model of the highway slope and construct a dynamic evolution model of the highway slope.

[0381] The process of constructing a slope dynamic evolution model:

[0382] Duplicate data acquisition and model update

[0383] Data Acquisition:

[0384] Following steps S1 to S5, repeatedly acquire the following data: high-resolution imagery data and ground-penetrating radar (GPR) data collected by the UAV.

[0385] Historical data series:

[0386] The three-dimensional geological model is updated based on the acquired data, and high-precision registration technology is used to accurately map data from different periods to a unified coordinate system.

[0387] The updated 3D geological models are organized chronologically, and key features (slope volume changes, crack propagation, and soil density) are extracted to form a historical data sequence. By introducing a time dimension, data is collected at fixed time intervals (e.g., quarterly, annually), with each collected 3D geological model serving as a time slice, thus constructing a dataset containing the time dimension, i.e., historical time series data, to provide historical information for model input.

[0388] The change in slope volume was calculated using a geometric method based on a three-dimensional geological model. In this model, the slope can be considered as being composed of multiple irregular triangular prisms, tetrahedrons, and other basic geometric units. The change in slope volume can be calculated by measuring the volume differences between the three-dimensional slope models at different times.

[0389] Basic geometric unit volume formula:

[0390] Volume of a triangular prism:

[0391] in, S is the area of ​​the base of the triangle, which can be determined using Heron's formula. ( It is half the perimeter. , , Z is the length of the three sides of the triangle (these data are all calculated based on the mesh division of the three-dimensional geological model), and Z is the height of the triangular prism, which is based on the elevation value of the vertex coordinates of the triangular prism.

[0392] Tetrahedral volume:

[0393] ,

[0394] in, , , These are three edge vectors originating from the same vertex in a three-dimensional geological model.

[0395] Calculation of total slope volume: The three-dimensional geological model is divided into multiple basic geometric units. The volume of each unit is calculated separately and then summed.

[0396]

[0397] in, It is the total volume of the slope. The number of geometric units. For the first Unit volume.

[0398] The propagation of cracks, including changes in length, width, and depth, is measured by identifying and analyzing crack features in a three-dimensional geological model.

[0399] Crack length: formula for distance between two points

[0400] If the crack can be approximated as a straight line in three-dimensional space, and the coordinates of the two ends of the crack are known to be... and Then the crack length The calculation formula is:

[0401]

[0402] in, It is the crack length. and These are the coordinates of the two ends of the known crack.

[0403] Crack width: measured perpendicular to the crack direction.

[0404] Within a certain cross-section of the crack, select two corresponding points on both sides of the crack, with coordinates as follows: and The crack width The calculation formula is:

[0405]

[0406] in, It is the width of the crack. and These are the coordinates corresponding to the two sides of the crack.

[0407] Crack depth: based on surface reference point:

[0408] Assume the coordinates of the deepest point of the crack are The coordinates of the corresponding reference point on the Earth's surface are Then the crack depth The calculation formula is: .

[0409] The change in the density of soil and rock is analyzed and judged based on the change in the physical parameter of dielectric constant.

[0410] Data preprocessing

[0411] Data cleaning: Remove outliers and missing values, and fill in missing data using regression prediction methods; Standardization: Standardize attribute data of different dimensions to ensure data consistency and comparability, and improve model training performance.

[0412] Model architecture design:

[0413] Input layer: The input consists of future time period information (such as the next 3 months, 6 months, etc.) and historical time series data (containing the features of a 3D geological body model at multiple time steps). Key features of the 3D geological body model (such as the thickness of each rock layer, groundwater level, fracture length, etc.) are extracted and arranged in chronological order as input features.

[0414] LSTM layer: Multiple LSTM cells are set up to capture the time dependencies and complex evolution patterns in historical data. LSTM cells can effectively process long-term time series data and remember the state changes of slope geological bodies at different times.

[0415] Intermediate layer: Add a fully connected layer to further extract and fuse the features output by the LSTM layer, in preparation for outputting a future 3D geological model.

[0416] Output layer: Outputs a 3D geological model for a specified future time period, including information such as the spatial distribution of each rock layer, changes in groundwater level, and fracture development.

[0417] Model training:

[0418] Training data partitioning: The preprocessed historical data is divided into training set, validation set and test set, generally in a ratio of 7:1:2.

[0419] Loss function and optimizer: Mean squared error (MSE) is used as the loss function to measure the difference between the predicted future 3D geological body model and the actual situation; the Adam optimizer is used to update the model parameters and accelerate the training convergence speed.

[0420] Training process: Through continuous iterative training, the model parameters are adjusted to minimize the loss of the model on the validation set, thereby improving the model's prediction accuracy and generalization ability.

[0421] The current 3D geological model is compared with the predicted future 3D geological model, and the differences between the two are calculated (changes in slope volume, crack propagation, and changes in soil and rock density).

[0422] Slope volume differences:

[0423] ,

[0424] in, For the slope volume of the future three-dimensional geological model, This represents the slope volume in the current 3D geological model.

[0425] Crack Differences: Statistical analysis of differences in crack length, width, and depth.

[0426] Crack length difference :

[0427]

[0428] in, It's a difference in crack length. This represents the current crack length. For future crack length

[0429] Crack width difference :

[0430]

[0431] in, It is the difference in crack width The current crack width, For the future crack width.

[0432] Crack depth difference :

[0433]

[0434] in, It's due to the difference in crack depth. It is the future depth of the crack. This represents the current crack width.

[0435] The difference in density between soil and rock masses is analyzed based on the difference in their dielectric constant physical parameters.

[0436] Differences in the density of soil and rock masses: =

[0437] in, It is the difference in density of the soil and rock mass (the difference in dielectric constant). It is the future density of the soil and rock mass. It represents the current density of the soil and rock mass.

[0438] First, the differences in slope volume, crack length, crack width, crack depth, and soil density are standardized (Z-score standardization) to unify the dimensions.

[0439] Based on the calculation results of model differences, an instability risk assessment index is established:

[0440] The weights of each evaluation indicator were determined using the Analytic Hierarchy Process (AHP). The problem is decomposed into a target layer (slope instability risk assessment), a criterion layer (slope volume difference, crack difference, soil and rock density) and an indicator layer.

[0441] Experts in the field of geological engineering were invited to construct a judgment matrix based on the importance of three categories of indicators in the criterion layer: geometric difference, soil and rock density, and crack difference, as well as the specific indicators in each indicator layer, using a 1-9 scale. For example, if the crack difference indicator is considered slightly more important than the geometric difference indicator, a value can be assigned to the corresponding position in the judgment matrix.

[0442]

[0443] in These are the original values ​​for each difference. and We use the mean and standard deviation of the corresponding indicators to ensure that all indicators are under the same dimension before weighting and summing them.

[0444] Based on the calculated range of the instability risk assessment index R, different risk levels are classified, providing an intuitive basis for slope management decisions.

[0445]

[0446] By fusing surface imagery data acquired by drones with shallow geological data obtained through geophysical resonant imaging (GPR), a more comprehensive 3D geological model of highway slopes can be constructed. Projecting GPR data onto the 3D surface model allows for the spatial correlation between the surface and shallow geological structures, providing a more intuitive view of the slope's internal structure. Through periodic data acquisition and model updates, a four-dimensional dynamic evolution model of the highway slope can be built. This model can track the dynamic changes in the slope's surface and internal structure. Combined with time series analysis and machine learning algorithms, it can predict slope stability trends, providing early warnings and decision support for highway management departments.

[0447] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0448] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

Claims

1. A highway slope modeling method based on a UAV inspection route, characterized in that, The method comprises the following steps: Step S1: measuring and on-site investigating the to-be-measured highway slope to obtain the length, topographic and geomorphic features and geological complexity of the to-be-measured highway slope, and selecting a type of unmanned aerial vehicle and a sensor combination based on the length, topographic and geomorphic features and geological complexity of the to-be-measured highway slope; Step S2: designing a space matching scheme of the unmanned aerial vehicle surface information inspection route and the shallow geological exploration scanning route, and synchronously acquiring high-resolution image data and ground penetrating radar data by the unmanned aerial vehicle based on the space matching scheme and the sensor combination, wherein the ground penetrating radar data comprises ground penetrating radar acquisition coordinates; Step S3: processing the high-resolution image data by using a photogrammetry technique to generate a highway slope digital elevation model and an orthographic image, and combining the digital elevation model and the orthographic image by using a three-dimensional modeling software to obtain a highway slope surface three-dimensional model; Step S4: obtaining reflection features in the ground penetrating radar by using a geophysical inversion method, projecting the reflection features to the highway slope surface three-dimensional model according to the ground penetrating radar acquisition coordinates to perform space registration and mutual verification, and obtaining a space logical relationship between the surface and the shallow internal part; Step S5: based on the space logical relationship between the surface and the shallow internal part, obtaining a three-dimensional surface element representing the reflection features by digitizing and geometrically reconstructing the reflection features by using the three-dimensional modeling software, and embedding the three-dimensional surface element representing the reflection features into the highway slope surface three-dimensional model to form a highway slope three-dimensional geological body model; Step S6: repeatedly acquiring the high-resolution image data and the ground penetrating radar data by introducing a time dimension, updating the highway slope three-dimensional geological body model, and constructing a highway slope dynamic evolution model. 2.The highway slope modeling method based on the route inspection of the unmanned aerial vehicle according to claim 1, wherein, The selecting the type of unmanned aerial vehicle and the sensor combination based on the length, topographic and geomorphic features and geological complexity of the to-be-measured highway slope comprises: Selecting the type of unmanned aerial vehicle and the sensor combination: Combination 1: long-endurance fixed-wing unmanned aerial vehicle + medium-high-precision visible light camera + light ground penetrating radar; Applicable scene: slope length < 1000 meters, relatively regular terrain; geological condition: geological complexity C / D level; Unmanned aerial vehicle and sensor configuration: Unmanned aerial vehicle: long-endurance fixed-wing, endurance 24 hours; Sensor: medium-high-precision visible light camera: select a medium-high-precision visible light camera with 16-20 million pixels; integrated light ground penetrating radar: select a light ground penetrating radar with 800-1000 MHz, which can detect shallow water-bearing areas and loose overburden, and meet the basic shallow geological structure detection requirements; Combination 2: large multi-rotor unmanned aerial vehicle + high-resolution visible light camera + ultra-light ground penetrating radar: Applicable scene: slope length: 500-1000 meters, complex terrain; geological condition: geological complexity B / C level; Unmanned aerial vehicle and sensor configuration: Unmanned aerial vehicle: select a six-rotor unmanned aerial vehicle with hovering accuracy ± 0.1 meters, and support low-altitude slope flying; Sensor: visible light camera: carry a 20 million pixel or above aerial survey camera, cooperate with a POS system to generate centimeter-level resolution DOM, and accurately identify surface cracks and weathering and peeling; Ultra-light ground penetrating radar (GPR): integrated 1000MHz high-frequency antenna, weight <1kg, detection depth 3-5 meters, suitable for shallow water-bearing area, loose overburden detection; Combination 3: vertical take-off and landing compound wing UAV + high-resolution visible light camera + small ground penetrating radar: Applicable scenarios: sideline length <800 meters, mixed terrain; geological conditions: geological complexity level A / B / C; UAV and sensor configuration: UAV: choose vertical take-off and landing compound wing, combine fixed-wing endurance and multi-rotor landing flexibility, adapt to complex airspace in mountainous areas; Sensor: visible light camera and combination two are the same, ensure the three-dimensional accuracy of the ground surface; small GPR: integrated 500MHz antenna, detection depth 5~8 meters. 3.The highway slope modeling method based on the route inspection of the UAV according to claim 2, characterized in that, Based on the spatial matching scheme and sensor combination, the UAV synchronously acquires high-resolution image data and ground penetrating radar data, including: Spatial matching scheme: Surface information inspection route: Principle of strike: If the strike of the slope geological structure is known, the on-site investigation report clearly indicates that the strike of the slope is north-south, and the UAV carries a visible light camera parallel to the strike of the slope, the strike is north-south, which ensures the continuous coverage of the three-dimensional form of the slope surface; If the strike of the slope geological structure is unknown, the on-site investigation report cannot clearly determine the strike of the slope geological structure, and the surface information inspection route can be determined according to the geological investigation report in the highway design stage. If the slope along the highway is east-west, the surface information inspection route is designed to be east-west back and forth; Key parameters: Flight height: the UAV carrying a visible light camera is 100~200m away from the slope surface, which ensures that the image resolution is ≤5cm / pixel; ensures sufficient image overlap, lateral overlap ≥60%, longitudinal overlap ≥80%; Shallow geological exploration scanning route: Principle of strike: If the strike of the slope geological structure is known, the UAV carries a ground penetrating radar (GPR) perpendicular to the strike of the geological structure, which makes the GPR reflection wave group more clearly vertical to the structure, and the reflection signal is the strongest when cutting the structure vertically; If the structure strike is unknown, it is assumed to be 30° with the surface information inspection route, which is the surface information inspection route described above for the unknown slope geological structure, and is 30° with the surface information inspection route, and is transversely encrypted along the slope; Key parameters: shallow geological exploration scanning route spacing: 5~10 meters in geological complexity level A / B; 10~20 meters in geological complexity level C; 20~30 meters in geological complexity level D; flight height 10~20cm away from the slope surface; Trigger mode: synchronize GNSS positioning, record a set of GPR data every 1-2cm to ensure the spatial resolution of underground structures; Data synchronous acquisition: through integrated collaborative inspection path planning, nearly synchronous acquisition of surface information and shallow geological information is realized, which reduces the inconsistency of data caused by time difference, and ensures that the UAV can acquire surface form data while the GPR can synchronously acquire shallow ground penetrating radar data of the slope, which can indirectly reflect the characteristics of underground structures such as rock-soil layering, water-bearing condition, potential cavity or loose area. 4.The highway slope modeling method based on the UAV inspection route of claim 3, wherein, The generation of digital elevation model and orthophoto map of highway slope includes: Digital elevation model generation: Initialization cloth network: Separate ground points and non-ground points from the DSM, which adopts cloth simulation filtering to remove vegetation and building points, and obtain point cloud containing only terrain surface; Define a virtual cloth grid on the DSM, and set the initial height of the grid points to a fixed value higher than the highest point of the DSM to ensure that the initial height of the virtual cloth grid is above the DSM; Define physical parameters: Define the physical properties of the cloth, including virtual mass and elastic coefficient, which will affect the motion behavior of the cloth in the iteration process; Iterative calculation: Calculate the resultant force: For each point in the cloth grid, calculate the gravity and tension from the surrounding points; The gravity calculation formula is: ； wherein is the gravitational force, m is the virtual mass, and g is the gravitational acceleration; The calculation of tension is more complex, which depends on the relative position and distance between adjacent grid points. The Hook's law is used to approximate the calculation of tension, that is ； wherein, is the tension from the surrounding points, k is the elastic coefficient, is the distance change between adjacent points; Calculate acceleration, velocity and position: According to Newton's second law , the acceleration of each point is calculated, formula: ； where, is the acceleration of each point, is the gravity, is the tension from surrounding points, m is the virtual mass; Update the velocity of each point according to the acceleration, and the velocity update formula is: ； wherein, the updated velocity, is the acceleration of each point, the velocity before the update, n is the iteration number, and Δt is the time step; Update the position of each point according to the velocity, and the position update formula is: ； wherein, is the updated position of each point, Δt is the time step, is the position before the update, is the updated velocity; Convergence judgment: After each iteration, check the change in position of the cloth mesh points whether the change in position of all points is less than a certain threshold value, if the change in position of all points is less than the threshold value, then the iteration is considered to have converged; Extract ground points: After the iteration converges, the points where the cloth adheres to the terrain surface are the ground points, which are extracted from the original DSM to form a point cloud containing only the terrain surface; Interpolation and gridding processing: Perform interpolation processing on the extracted ground point cloud to generate regular grids; Assign elevation values to each grid point to form a digital elevation model (DEM) that can reflect the true elevation changes of the terrain; Orthographic image generation: According to the generated digital elevation model (DEM) and the interior and exterior orientation elements of the image, including the position and attitude of the camera and the interior parameters of the camera; Perform orthographic projection transformation on each image to eliminate the deformation caused by terrain undulations and image tilt; Mosaic the orthographic projected images to handle the color difference at the seam, and generate an orthographic image map that has geographic coordinates. 5.The highway slope modeling method based on the UAV inspection route according to claim 4, characterized in that, The combination of the digital elevation model (DEM) and the orthographic image map (DOM) through a three-dimensional modeling software to obtain a highway slope surface three-dimensional model, including: Highway slope surface three-dimensional model construction: Elevation grid foundation establishment: Use DEM as the terrain framework to form a terrain surface in three-dimensional space; DEM only contains elevation information of the terrain, and removes the height influence of vegetation and building objects; this makes it more accurate in pure terrain analysis; Mapping and visual rendering: Use DOM as a texture layer and register it through geographic coordinates to superimpose it on the elevation grid of DEM, so that the three-dimensional model has real surface color and texture characteristics, and improves the visualization accuracy of the model; Geometric reconstruction and detail fusion: In the three-dimensional modeling software, divide the grid based on the elevation data of DEM, and perform surface rendering combined with the texture information of DOM. For key objects, use the point cloud density information in DSM to enhance the detail description.

6. The highway slope modeling method based on the route inspection of the unmanned aerial vehicle according to claim 5, characterized in that, The geophysical inversion method is used to obtain the reflection characteristics in the ground penetrating radar (GPR), including: Through the geophysical inversion method, identify and extract the effective reflection wave group, phase axis, and abnormal reflection area in GPR: Input data: GPR data after filtering, gain, and offset preprocessing, containing intensity, phase, frequency, and propagation time information of reflected wave signals; Technical implementation process of inversion algorithm: First, the underground space is divided into small grid cells, and the dielectric constant of each grid cell is unknown and set to , ; Equation system is established: According to the propagation theory of GPR waves in the underground medium, for each pair of transmitting-receiving antenna positions, the propagation time of the wave can be expressed as: ； wherein is a coefficient representing the path coefficient of the wave from the transmitting antenna through the grid element to the receiving antenna , is the propagation path length of the wave in the grid element , is the propagation speed of the wave in the grid element ; Rearranging the above equation, one obtains an equation for the dielectric constant ε, and for multiple transmit-receive antenna pairs, one obtains a system of equations; The equation is nonlinear. In order to facilitate solving, Taylor series expansion method is used for linearization. Each nonlinear equation is linearized by this Taylor series expansion method; After linearization, all equations can be written in the form of linear equations where is the coefficient matrix, is the unknown vector, is the constant vector; This linearization method makes it possible to obtain an approximate solution of the originally complex nonlinear equation set by solving a linear equation set The permittivity vector ε plays a key role in geophysical inversion; Abnormal reflection area positioning: Using the dielectric constant anomaly vector obtained by inversion, combined with geological experience: High dielectric constant area: water-rich area or saturated rock-soil body; Low dielectric constant area: cavity, dry sand or loose accumulation; Dielectric constant fluctuation area: uneven rock-soil body density or fracture zone; Effective reflected wave group extraction: When the dielectric constant vector is in a sudden change state, that is, the dielectric constant vector of this grid point area suddenly increases or decreases to the dielectric constant vector of the next grid point area, the corresponding rock-soil layer interface, such as the interface between sandstone and clay, the top surface of bedrock; Tracing of the same phase axis: The same phase axis refers to the track formed by the points with the same phase in the electromagnetic wave reflection signal recorded by the ground penetrating radar, that is, the line formed by those points reaching the same vibration state at the same time in a series of continuous reflection wave signals. Through three-dimensional space extension analysis of continuous reflection wave sequence, the trend of the same phase axis is obtained, and the trend of the same phase axis is constrained combined with prior knowledge of geological structure to determine the spatial distribution of underground water level and continuous weak interlayer, and to invert the buried depth and inclination. Through analysis of effective reflection wave group, same phase axis, and abnormal reflection area, reflection characteristics are obtained, which can indicate the interface between different rock-soil layers, underground water level, density change, potential cavity or water-rich area, and shallow geological structure information.

7. The highway slope modeling method based on the route inspection of the unmanned aerial vehicle according to claim 6, characterized in that, After digitizing and geometric reconstruction of the reflection characteristics by the three-dimensional modeling software, a three-dimensional surface element representing the reflection characteristics is obtained, including: For the reflection feature, it is discretized into a series of geological reflection feature points, and according to the final three-dimensional coordinates of the geological reflection feature points , a three-dimensional facet is constructed by a triangulation interpolation algorithm. For three points , and , a three-dimensional facet can be constructed. 8.The highway slope modeling method based on the UAV inspection route according to claim 7, characterized in that, Embedding the three-dimensional surface element representing the reflection characteristics into the three-dimensional model of the highway slope surface to form a three-dimensional geological body model of the highway slope, including: Embedding the shallow geological reflection characteristics into the three-dimensional model of the highway slope surface to form a three-dimensional geological body model of the highway slope; Ensure that the embedding process meets the geometric constraints: Layer surface continuity: for the adjacent rock-soil layer interface, in the embedding process, it is necessary to ensure that the interface between adjacent layers is continuous; the distance of the adjacent two rock-soil layer interfaces in a certain direction is l, and by adjusting the position of the geological reflection characteristics, the distance tends to 0, ensuring the layer surface continuity tends to 0, ensuring the layer surface continuity Geological properties and visualization parameters are given: Different rock-soil unit attribute assignment: according to the rock-soil type, different geological reflection characteristics are given corresponding physical properties; Visual parameter settings: color, texture, and transparency are set for geological reflectance characteristics according to rock type visualization parameters;​​ The final generated three-dimensional geological body model of the highway slope allows users to adjust the transparency of different geological units, make arbitrary cuts, or hide certain layers locally, so as to "see through" the inside of the slope and intuitively understand the spatial distribution of the main rock-soil layers, the location and shape of the potential weak structure plane, providing a much deeper understanding of the slope stability than traditional two-dimensional or simple surface models. 9.The highway slope modeling method based on the UAV inspection route of claim 8, wherein, Through the introduction of time dimension, repeated acquisition of high-resolution image data and ground penetrating radar data, the three-dimensional geological body model of the highway slope is updated, including: Repeated data acquisition and model updating: Data acquisition: The following data is repeatedly acquired: high-resolution image data and ground penetrating radar data collected by a drone; Historical data sequence: According to the obtained data, update the three-dimensional geological body model, and accurately correspond the data of different periods to a unified coordinate system through high-precision registration technology; Arrange the updated three-dimensional geological body model in chronological order, extract key features, form a historical data sequence, and collect data at fixed time intervals by introducing a time dimension. The three-dimensional geological body model collected each time serves as a time slice, and a data set containing a time dimension, i.e., historical time series data, is constructed to provide historical information for model input. 10.The highway slope modeling method based on the UAV inspection route according to claim 9, characterized in that, The constructed highway slope dynamic evolution model includes: Model architecture design: Input layer: The input is future time period information and historical time series data. Key features of the three-dimensional geological body model are extracted and arranged in chronological order as input features. LSTM layer: Multiple LSTM units are set to capture the time dependence and complex evolution rules in historical data. LSTM units can effectively process long-term time series data and remember the state changes of the slope geological body at different times. Intermediate layer: Add a fully connected layer to further extract and fuse the features output by the LSTM layer, and prepare for outputting the future three-dimensional geological body model. Output layer: Output the three-dimensional geological body model for a specified period of time, including the spatial distribution of each rock layer, groundwater level changes, and crack development information. Model training: Training data division: Divide the preprocessed historical data into training set, validation set and test set in the ratio of 7:1:

2. Loss function and optimizer: Use mean square error as the loss function to measure the difference between the predicted future three-dimensional geological body model and the actual situation. Use the Adam optimizer to update the model parameters and speed up the training convergence speed. Training process: Through continuous iteration training, adjust the model parameters to make the loss of the model on the validation set reach the minimum, and improve the prediction accuracy and generalization ability of the model.

Images