A method for simulating rainfall landslide susceptibility based on binocular camera technology combined with deep learning algorithm and finite element analysis

Abstract

This invention discloses a method for simulating landslide susceptibility based on binocular camera technology combined with deep learning algorithms and finite element analysis. The method uses a UAV's binocular camera to collect high-precision 3D topographic data of landslide-prone areas, and combines this data with remote sensing to detect soil and vegetation physical properties to construct a 3D solid model. Coupled with real rainfall scenarios, a deep learning algorithm is used to optimize the rainfall intensity-soil cohesion correlation model, simulating various real rainfall conditions. Continuous displacement data of the landslide surface is extracted and imported into the finite element model, enabling dynamic simulation of the entire process of a landslide from local deformation to overall collapse and the formation of a new steady state. Simultaneously, model optimization and landslide susceptibility assessment are completed. This invention links existing monitoring technologies with mechanical simulation, solving the problem of inaccurately representing the dynamic evolution of landslides, and providing precise technical support for landslide disaster early warning and prevention.

Description

001. Technical Field

[0002] This invention falls under the technical category of landslide disaster mechanism simulation. Specifically, it is a dynamic modeling method for the entire landslide process that integrates UAV binocular camera visual measurement technology, soil and vegetation physical property parameter measurement, deep learning algorithms, and finite element analysis, along with a supporting implementation system. In short, it uses a UAV-mounted binocular camera to acquire high-precision 3D topographic data of landslide-prone areas. This data is combined with on-site soil sampling and remote sensing data, along with measured soil type, vegetation type and its coverage and age, and laboratory-measured soil cohesion, to construct a 3D solid model consistent with the actual landslide-prone area. By coupling with real rainfall scenarios, and utilizing deep learning algorithms and finite element analysis, it accurately simulates the dynamic correlation between different rainfall intensities and types and soil cohesion, as well as the impact of real-world conditions such as runoff direction, water lag, and root systems of different types and ages of vegetation on the on-site soil. Ultimately, it comprehensively predicts the entire process of a landslide, from the decrease in soil cohesion and local deformation under rainfall, to overall collapse, and finally to the formation of a new steady state after collapse, providing accurate simulation for landslide disasters. 002. Background Technology

[0004] Landslides fall under the category of sudden and highly destructive geological hazards. Accurately predicting whether a landslide will occur, under what conditions it is likely to happen, and the extent of its impact after it occurs has always been a challenging problem in disaster prevention and control. Currently used landslide monitoring and assessment methods can be broadly categorized into three types, each with its own significant drawbacks.

[0005] The first type relies on statistical methods using Geographic Information Systems (GIS). This method involves collecting predefined, highly relevant known data, such as topography, geological structure, groundwater, and land use from remote sensing data. It typically combines this with past landslide records and uses mathematical models or machine learning algorithms to determine which areas are prone to landslides. However, its drawbacks are significant: it only provides static results under given known conditions and cannot depict the real-time changes of a single landslide body under the influence of external factors such as dynamic rainfall, vegetation growth, soil surface runoff, and water retention. It is suitable for rough assessments over a large area, rather than for detailed dynamic susceptibility simulations of individual landslide bodies.

[0006] The second type is the fixed-point monitoring method, which involves installing sensors at the landslide site, such as GNSS receivers and inclinometers, to accurately measure the displacement of specific points. However, this technology suffers from the problem of "representing the whole by a single point," meaning it only knows the location of the sensor and cannot grasp the deformation of the entire landslide body. Displacement caused by the instability of the sensor itself will be directly misreported as a landslide. Furthermore, the equipment is expensive, and the deployment cost is extremely high in areas with complex terrain and many landslide-prone areas, and installation is quite difficult. Most importantly, it can only detect the displacement of the surface of the landslide-prone body when a landslide occurs or the soil becomes unstable, and it is of no use in simulating the susceptibility of rainfall-induced landslide disasters.

[0007] The third category involves monitoring using remote sensing technologies, such as InSAR and LiDAR. InSAR can monitor large-scale deformation to a certain extent; however, its satellite-based nature results in a relatively long monitoring period, and its effectiveness is somewhat reduced in areas with vegetation cover. It is commonly used for long-term assessments of landslides with minor displacements. LiDAR can acquire high-precision three-dimensional terrain data. However, rainfall-induced landslides are often accompanied by low visibility and strong convection. Airborne LiDAR faces challenges such as reduced accuracy due to water droplets scattering the laser beam, and the aircraft carrying LiDAR also faces safety hazards from severe convective weather. Therefore, it is generally used for long-term monitoring of minor displacements in densely vegetated landslides. Both InSAR and LiDAR can only monitor displacement and cannot simulate the landslide susceptibility under the influence of dynamically changing rainfall processes, vegetation, and surface runoff.

[0008] There is also a mechanical simulation technique called the "finite element method," which, theoretically, has the ability to accurately simulate the instability process of landslides. However, this method has a rather tricky problem: the reliability of the simulation results depends entirely on the accuracy of the input soil mechanical parameters. Parameters such as the growth of vegetation roots often require long-term, continuous field data collection at different seasonal points to achieve a certain level of accuracy.

[0009] The current technology faces a core challenge: there is a disconnect between a method for monitoring landslide deformation in a relatively complete and accurate manner and a model that can accurately simulate changes in landslide mechanics. Currently, there is no technology that can effectively utilize the real-time and comprehensive three-dimensional deformation data of the landslide surface in mechanical simulation. As a result, it is impossible to achieve accurate dynamic monitoring and early warning effects. 003. Summary of the Invention

[0011] This invention aims to address the problems of vague targets, low scene fit, insufficient parameter support, and imprecise implementation logic in existing landslide simulation technologies. It provides a precise and practical method for modeling the entire landslide process, along with a corresponding implementation system and related computer program storage media. Its core objective is to fully reproduce the entire landslide process (from changes in soil cohesion and local deformation under rainfall, to overall collapse, and then to the formation of a new steady state after collapse). By simulating the landslide "collapse," it focuses on the core closed loop of "rainfall → decreased cohesion → collapse → new steady state." The specific technical solution is as follows:

[0012] A landslide full-process modeling method based on UAV binocular cameras and finite element simulation is proposed. This method mainly covers four steps and can achieve a fully automated process of "data acquisition and model building → rainfall coupled simulation → landslide full-process simulation → model verification and optimization".

[0013] Step 1: The drone's binocular camera collects data and reconstructs a 3D model.

[0014] This step forms the basis of the entire simulation method. Its core is to acquire high-precision three-dimensional terrain data of the landslide area non-contactly and from all directions using a drone's binocular camera, and combine this data with relevant parameters measured by on-site soil sampling and remote sensing data to construct a three-dimensional solid model that is consistent with the real landslide-prone area.

[0015] In terms of equipment deployment, a drone equipped with a high-resolution binocular camera was selected. A scientifically planned cruise route was developed, taking into account the terrain features of the target landslide area. Through drone flight control, a 360° horizontal and 180° vertical full-range coverage of the landslide and its surrounding area was achieved, ensuring the binocular camera could completely and meticulously capture all terrain features of the landslide. During camera calibration, a high-precision calibration board and conventional camera calibration methods were used to accurately acquire the internal parameters of the binocular camera (including lens focal length, imaging center, etc.) and their relative positional relationships, effectively avoiding shooting errors and ensuring the accuracy of the acquired images. In the image acquisition phase, the drone flew along the preset cruise route, controlling the binocular camera through a synchronous triggering device. During the cruise, it continuously acquired digital image sequences of the entire landslide area, ensuring coverage of terrain features from different angles and in different areas, providing comprehensive data support for subsequent 3D reconstruction.

[0016] In the 3D reconstruction process, the acquired multi-view images are first optimized, including converting them to black and white images, removing blurred parts, and correcting shooting distortion. Then, based on the images acquired across the entire range, a region matching algorithm is used to accurately establish the pixel correspondence between images from different viewpoints. Finally, combined with the parameters obtained from camera calibration, the 3D coordinates of the corresponding points are calculated to generate dense 3D point cloud data covering the entire landslide area. Combined with parameters such as soil type, vegetation type and its coverage, and growth age determined by on-site soil sampling and remote sensing data, a 3D solid model of the landslide body is constructed.

[0017] Step 2: Rainfall Coupling Simulation (Close to Real Rainfall Scenario)

[0018] The core of this step is to establish a dynamic correlation between rainfall and soil properties and vegetation parameters, recreating the real working conditions under non-uniform rainfall. Deep learning algorithms are then used to optimize the fitting accuracy of the rainfall intensity-soil cohesion dynamic correlation model, providing precise load conditions for subsequent mechanical simulations. First, the core correlation logic is clarified: based on a literature-supported rainfall intensity-soil cohesion dynamic correlation model, combined with laboratory-measured soil cohesion parameters, the higher the rainfall intensity, the lower the soil cohesion, until it is insufficient to support the mountain's own weight. Simultaneously, the influence of vegetation root systems (different types and growth ages) on soil cohesion is considered.

[0019] The key operating condition simulation covers three aspects, fully reflecting real-world scenarios: First, soil water absorption characteristics simulation. When the soil is saturated with water during heavy rainfall, excess water forms surface runoff. The runoff direction is calculated through topographic slope analysis and hydrodynamic models, simulating the process of runoff converging into depressions and low-lying pits, while also considering the impact of vegetation cover on runoff. Second, water accumulation lag effect simulation. For water accumulation areas such as mountain pits, the simulation of rainwater accumulation process and the continuous impact of water accumulation on soil cohesion is performed, restoring the characteristics of delayed landslide occurrence. Third, non-uniform rainfall simulation. Avoiding the traditional assumption of uniform rainfall, the simulation sets spatial differences in rainfall intensity based on actual rainfall distribution patterns, reflecting the impact of real rainwater distribution and runoff path differences on soil stress.

[0020] The final output of this step is: rainfall coupling data (including dynamic cohesion, runoff force, water pressure, etc.).

[0021] Step 3: Extract the displacement data of the entire landslide surface.

[0022] The core task of this step is to transform the discrete three-dimensional point cloud data generated in step 1 into continuous displacement data that can be directly applied to finite element mechanical simulation, thus providing experimental basis for simulating the landslide deformation process.

[0023] Input for this step: Multi-phase 3D point cloud data generated in step 1; Output for this step: Continuous displacement data of the landslide surface.

[0024] The specific implementation process is as follows: First, point cloud registration is performed. The point cloud data collected in the first batch is selected as the reference point cloud. Conventional algorithms in the field of point cloud registration (such as the ICP algorithm) are used to adjust the point cloud data collected in subsequent batches to the same coordinate system as the reference point cloud, which facilitates subsequent displacement comparison and analysis. Second, displacement calculation is performed. By comparing the registered point clouds of each batch with the reference point cloud, the displacement change of each corresponding point in the X, Y, and Z directions is calculated. Finally, a continuous displacement field is generated. Since the measured point cloud is discretely distributed, a spatial interpolation method is used to fill the displacement data of discrete points into the preset mechanical simulation grid nodes, forming complete and directly usable continuous displacement data of the landslide surface for finite element simulation.

[0025] Step 4: Establish a finite element model and import displacement data

[0026] The core objective of this step is to leverage finite element analysis technology to achieve a complete dynamic simulation of the landslide's progression from "undeformed → localized deformation → overall collapse → new steady state," and to improve the realism and accuracy of the simulation by combining the rainfall coupling data from step 2 and the displacement data from step 3.

[0027] 1) Establish a geological model

[0028] Based on the field survey data (such as topographic maps, borehole data, etc.) and combined with the three-dimensional solid model of the landslide obtained in step 1, a three-dimensional geological model of the landslide is constructed to clarify the potential sliding surface, soil stratification and vegetation distribution. The three-dimensional solid model is then imported into mechanical simulation software (such as ABAQUS, COMSOL), the calculation mesh is divided, and the soil mechanics model (clarifying the soil stress and deformation law), initial boundary conditions and initial stress state are set to ensure that the model is consistent with the actual landslide geological conditions.

[0029] 2) Loading rainfall coupling data

[0030] The rainfall coupling data (including dynamic cohesion, runoff force, water pressure, etc.) output in step 2 are used as load conditions to load the finite element model. The rainfall intensity is gradually increased, and the changes in soil cohesion, vegetation root action and local deformation location are tracked in real time through the finite element model to reconstruct the initial deformation process of the landslide under the action of rainfall.

[0031] 3) Import displacement data

[0032] The continuous displacement data of the landslide surface obtained in step 3 is used as "known conditions" and loaded onto the surface nodes of the finite element model so that the simulation model can perform calculations according to the actual deformation of the landslide. The simulation continues until the mountain collapses as a whole, and the movement trajectory and accumulation process of the collapsed soil are tracked until the soil reaches a state of mechanical equilibrium and forms a new stable form, thus completing the closed-loop simulation of the entire landslide process.

[0033] Step 5: Optimize model parameters and evaluate landslide susceptibility

[0034] The core of this step is to calibrate and optimize the finite element model to improve its prediction accuracy, and to conduct a landslide susceptibility assessment based on the calibrated model to provide support for landslide disaster early warning. Specifically, it includes three stages: parameter optimization, model verification, and susceptibility assessment.

[0035] 1) Parameter optimization

[0036] With the optimization objective of "minimizing the difference between the simulated displacement and the actual measured displacement", an optimization algorithm is used to adjust key soil mechanical parameters (such as internal friction angle, soil cohesion, etc.) and vegetation influence parameters, so that the model simulation results are highly consistent with the actual landslide conditions, thereby improving the reliability and prediction accuracy of the model.

[0037] 2) Model Validation

[0038] Measured displacement data that were not involved in the parameter optimization process were selected to verify the adjusted model, test the model's simulation accuracy and stability, and ensure that the model has reliable landslide prediction capabilities, thus providing a guarantee for subsequent susceptibility assessment.

[0039] 3) Susceptibility assessment

[0040] Using a calibrated finite element model, the stability of landslides (i.e., the safety factor) is calculated. Through comparative analysis of the safety factors, the landslide susceptibility level and the danger zone where landslides are most likely to occur are determined. Geological monitoring equipment is prioritized in the danger zone, and the order of equipment deployment is rationally planned to give full play to the early warning role of geological monitoring equipment and provide precise guidance for landslide disaster prevention and control. 004. Description of the attached drawings

[0042] Figure 1 shows the complete process of the method of the present invention, which presents a series of complete steps from the acquisition of UAV binocular data, through the processing stage, to the model coupling stage, and finally to the determination of the most likely landslide location. 005. Detailed Implementation

[0044] To clearly illustrate the technical implementation process of this invention, the following detailed explanation is provided with reference to actual cases, focusing on the full process of landslide "no deformation → local deformation → overall collapse → new steady state":

[0045] A potential landslide area in a mountainous region was selected as the test subject. The elevation difference is about 300m, and it contains local depressions and pits. The surface layer is mainly loess mixed with gray soil, and the vegetation is deciduous shrubs with an average coverage of 40%. The goal is to reproduce the mechanical and morphological changes of the entire landslide process under rainfall and to verify the feasibility and accuracy of the method of this invention.

[0046] Step 1: Basic Data Acquisition and 3D Solid Model Construction

[0047] This step corresponds to the core requirements of the invention, completing the collection of 3D data, soil and vegetation parameters, and constructing a 3D solid model that closely resembles the real scene. A drone equipped with a high-resolution binocular camera was selected, employing a "spiral + reciprocating" cruising path to achieve full coverage of the landslide body and its surroundings at 360° horizontally and 80° vertically. Using a 2m×1.5m high-precision checkerboard calibration board and the Zhang Zhengming calibration method, parameters such as camera focal length and imaging center were obtained to eliminate shooting errors. The drone flew at an altitude of 100m and a speed of 5m / s, simultaneously acquiring thousands of frames of multi-view images. After preprocessing, a region matching algorithm was used to achieve 98.7% pixel matching, generating a 3D point cloud model with a density of 100 points / cm² and a terrain contour error ≤±2mm.

[0048] Sampling was conducted according to the principle of "grid-based sampling + focused density sampling." Soil types were determined based on the "Soil Texture Classification Standard." A ZJ-type direct shear tester was used to measure cohesion (28 kPa for loess vegetation areas, 22 kPa for unvegetated areas, and 35 kPa for mixed soils), while simultaneously acquiring parameters such as saturated water absorption rate. The vegetation index was calculated using the average number of roots at five 1-meter deep sampling points within 1 square meter (average 11 roots / m³), determining the root system's contribution to cohesion as 1.3-1.5. The point cloud model was integrated with soil and vegetation parameters to construct a 500m×300m×300m three-dimensional solid model, clearly defining key features such as pits and soil interfaces.

[0049] Step 2: Implementation of Rainfall Coupling Simulation

[0050] In accordance with the requirements of the invention, COMSOL finite element software coupled with a deep learning algorithm was used to set a gradient rainfall intensity (starting at 20 mm / h, increasing by 10 mm / h every 30 minutes, up to a maximum of 80 mm / h), and a non-uniform rainfall field was set based on historical data. Simulations were made of conditions such as runoff convergence after soil saturation and stagnant water accumulation in potholes (40 m³ of water soaking for 6 hours). The accuracy of the rainfall intensity-cohesion fitting was optimized using a deep learning algorithm, ultimately outputting dynamic data: after 2 hours of continuous 80 mm / h rainfall, the cohesion in the loess vegetation area decreased to 8 kPa, in the non-vegetated area to 5 kPa, and in the mixed soil to 12 kPa.

[0051] Step 3: Extraction of landslide surface displacement data

[0052] Based on the initial point cloud acquisition, a set of time-series point clouds is acquired every hour thereafter. The ICP registration algorithm (error ≤ ±1mm) is used to unify the coordinate system, calculate the three-dimensional displacement of each point in X, Y, and Z, and fill the grid nodes with discrete data using the Kriging interpolation method to generate continuous displacement field data that can be directly used for finite element simulation, with an interpolation error ≤ ±1.5mm.

[0053] Step 4: Finite element model construction and full-process simulation

[0054] The 3D solid model was imported into ABAQUS software, divided into tetrahedral elements, and an elastoplastic constitutive model was adopted. Fixed boundaries and initial conditions such as gravity and initial groundwater pressure were set. The rainfall coupling data from step 2 and the displacement data from step 3 were loaded, and the rainfall intensity was gradually increased: at 50 mm / h, a local deformation of 15 mm appeared on the surface of the loess around the pit; after 3 hours of continuous rainfall at 80 mm / h, the displacement accumulated to 50 mm, and the overall collapse simulation was initiated (lasting 12 minutes); 24 hours after the collapse, the soil formed a new stable slope, completing the closed-loop simulation of the entire process. Step 5: Model parameter optimization and verification and susceptibility evaluation.

[0055] With the goal of minimizing the mean square error between simulated and measured displacements, a genetic algorithm was used to optimize parameters, adjusting the internal friction angle from 25° to 28° and setting the adhesion correction coefficient to 0.95. After optimization, the mean square error decreased from 12.1% to 7.8%. Two sets of time-series point cloud data that were not involved in the optimization were selected for verification, with relative errors of 8.3% and 7.6%, respectively, both less than 10%, verifying the model's reliability. The safety factor was calculated using the calibrated model, with an overall value of 1.08 (high susceptibility). The area around the pit with no vegetation had the lowest safety factor (0.97), making it the most likely area to landslide. Monitoring equipment was prioritized in this area to support disaster early warning.

[0056] Summary of Implementation Results

[0057] This embodiment verifies the feasibility and accuracy of the method: millimeter-level terrain data acquisition and standardized parameter determination to construct a realistic model, rainfall coupling to restore actual working conditions, and finite element simulation to fully reproduce the entire landslide process with an error controlled within 10%. This effectively solves the pain points of traditional simulation and provides a precise modeling method for landslide mechanism research.

Claims

1. A method for simulating rainfall landslide susceptibility based on binocular camera technology and finite element analysis, characterized in that, The method comprises the following steps: Step 1: A multi-view digital image sequence of the target landslide area is collected by a UAV equipped with a binocular camera, and a three-dimensional entity model of the landslide body consistent with the real landslide area is constructed by combining the soil sampling on site, the soil parameters and vegetation parameters measured by remote sensing data; Step 2: Based on the rainfall intensity-soil cohesion dynamic correlation model, the fitting precision is optimized by combining the deep learning algorithm, the rainfall coupling simulation close to the real rainfall scene is carried out, and the rainfall coupling data is output; Step 3: Based on the three-dimensional point cloud data collected in multiple periods, the landslide surface continuous displacement data which can be directly used for finite element simulation is generated through point cloud registration, displacement calculation and spatial interpolation processing; Step 4: A finite element geological model of the landslide body is constructed, the rainfall coupling data obtained in step 2 and the landslide surface continuous displacement data obtained in step 3 are loaded, and the whole process dynamic simulation from the initial deformation of the landslide under the action of rainfall to the overall collapse and then to the new steady state after the collapse is completed; Step 5: The parameters of the finite element model are optimized and the model is verified, the landslide safety factor is calculated based on the calibrated finite element model, and the landslide susceptibility evaluation is carried out.

2. The method of claim 1, wherein the method is characterized by, The step 1 specifically comprises: The UAV cruising path is planned according to the topographic features of the target landslide area, the camera calibration of the binocular camera is completed through the high-precision calibration plate, and the internal parameters and relative position relationship of the camera are obtained; the UAV is controlled to fly according to the preset cruising path, the multi-view digital image sequence of the whole range of the landslide body is continuously collected through the synchronous triggering device controlling the binocular camera; the collected digital image sequence is preprocessed, the pixel correspondence relationship between different view images is established based on the region matching algorithm, the three-dimensional coordinates of the corresponding points are calculated combined with the camera calibration parameters, and the dense three-dimensional point cloud data of the whole landslide area is generated; the soil type, cohesion, and saturated water absorption rate parameters measured by soil sampling on site, the vegetation type, coverage, and growth period parameters measured by remote sensing data are combined, and the three-dimensional point cloud data is integrated to construct a three-dimensional entity model of the landslide body. 3.The method of claim 1, wherein, The rainfall coupling simulation of step 2 includes soil water absorption characteristic simulation, water accumulation hysteresis effect simulation and non-uniform rainfall simulation, and the rainfall coupling data includes dynamic soil cohesion, runoff force and water accumulation pressure data.

4. The method according to claim 3, wherein, The step 2 specifically comprises: based on the soil cohesion parameters measured in the laboratory, a rainfall intensity-soil cohesion dynamic correlation model is established, and the influence coefficient of different types and different growth periods of vegetation roots on soil cohesion is coupled; through terrain slope analysis and water flow dynamics model, the flow direction and convergence process of surface runoff after soil water absorption saturation are simulated, and the influence of vegetation coverage on runoff is coupled; for the water accumulation area of the landslide area, the rainwater convergence process and the continuous influence of water accumulation on soil cohesion are simulated, and the water accumulation hysteresis effect of the delayed occurrence of the landslide is restored; according to the actual rainfall distribution law, the spatial difference of rainfall intensity is set, the non-uniform rainfall simulation is carried out, and the influence of the difference between the real rainwater distribution and the runoff path on the soil stress is fitted; the fitting precision of the rainfall intensity-soil cohesion dynamic correlation model is optimized by a deep learning algorithm, and finally the rainfall coupling data is output.

5. The method of claim 1, wherein the method is characterized by, Step 3 specifically includes: selecting the initially acquired 3D point cloud data as the reference point cloud, and using a point cloud registration algorithm to register subsequent point cloud data. The point cloud data collected in each phase is adjusted to the same coordinate system as the reference point cloud; the registered point clouds of each phase are compared with the reference point cloud, and the displacement change of each corresponding point in the three directions of X-axis, Y-axis and Z-axis is calculated; the spatial interpolation method is used to fill the displacement data of discrete points into the preset mechanical simulation grid nodes to generate continuous displacement field data of landslide surface.

6. The method of claim 1, wherein the method is characterized by, Step 4 specifically includes: based on the on-site survey data and combined with the three-dimensional solid model of the landslide obtained in Step 1, constructing a three-dimensional geological model of the landslide to clarify the potential sliding surface, soil stratification, and vegetation distribution; importing the three-dimensional geological model into mechanical simulation software, dividing the calculation grid, and setting the soil mechanical model, initial boundary conditions, and initial stress state; loading the rainfall coupling data output in Step 2 as load conditions onto the finite element model, gradually increasing the rainfall intensity, and tracking changes in soil cohesion, vegetation root action, and local deformation locations in real time through the finite element model; loading the continuous displacement data of the landslide surface obtained in Step 3 onto the surface nodes of the finite element model, so that the model can perform calculations according to the actual deformation of the landslide; continuing the simulation until the entire mountain collapses, tracking the movement trajectory and accumulation process of the collapsed soil until the soil reaches a state of mechanical equilibrium and forms a new stable form, completing the closed-loop simulation of the entire landslide process.

7. The method according to claim 1, wherein, Step 5 specifically includes: taking the minimum difference between the simulated displacement and the actual measured displacement as the optimization objective, adjusting the key mechanical parameters of the soil and the vegetation influence parameters using an optimization algorithm to complete the model parameter optimization; selecting measured displacement data that did not participate in parameter optimization to verify the adjusted finite element model and detect the simulation accuracy and stability of the model; calculating the landslide safety factor using the calibrated finite element model, and determining the landslide susceptibility level and the danger zone where the landslide is most likely to occur through safety factor comparison analysis.

8. A rainfall-induced landslide susceptibility simulation system based on binocular camera technology and finite element analysis, characterized by... In the present application, It includes a drone data acquisition module, a 3D model construction module, a rainfall coupling simulation module, a displacement data extraction module, a finite element full-process simulation module, and a model optimization and susceptibility evaluation module. The UAV data acquisition module is communicatively connected to the 3D model construction module. The 3D model construction module is communicatively connected to the rainfall coupling simulation module and the finite element full-process simulation module. The rainfall coupling simulation module is communicatively connected to the finite element full-process simulation module. The displacement data extraction module is communicatively connected to the UAV data acquisition module and the finite element full-process simulation module. The finite element full-process simulation module is communicatively connected to the model optimization and susceptibility evaluation module.

9. The system according to claim 8, wherein, The UAV data acquisition module is equipped with a high-resolution binocular camera to acquire multi-view digital image sequences of the target landslide area; the 3D model construction module is used to preprocess the acquired images, generate 3D point clouds, and construct a 3D solid model of the landslide body by combining on-site sampling and remote sensing data; the rainfall coupling simulation module has a built-in dynamic correlation model of rainfall intensity-soil cohesion and a deep learning algorithm optimization unit to simulate soil water absorption characteristics, water accumulation lag effect, and non-uniform rainfall, and output rainfall coupling data. The displacement data extraction module is used to register, calculate displacement, and spatially interpolate multi-phase three-dimensional point cloud data to generate continuous displacement data of the landslide surface; the finite element full-process simulation module is used to construct a finite element geological model of the landslide body, load rainfall coupling data and continuous displacement data, and complete the dynamic simulation of the entire landslide process; the model optimization and susceptibility evaluation module is used to optimize and verify the parameters of the finite element model, calculate the landslide safety factor based on the calibrated model, and output the landslide susceptibility evaluation results.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the program is executed by the processor, it implements the rainfall-induced landslide susceptibility simulation method based on binocular camera technology and finite element analysis as described in any one of claims 1 to 7.

Images