Slope slip risk image classification and evaluation method in loess gully region

Abstract

The application discloses a loess gully area side slope sliding risk image grading evaluation method, relates to the technical field of remote sensing image processing, and comprises the following steps: acquiring images of multiple periods of aerial photography of a target side slope by an unmanned aerial vehicle, reconstructing a three-dimensional terrain, and rasterizing the three-dimensional terrain into elevation images of the multiple periods; calculating the slope and slope direction of each grid; determining a high and steep free surface area as a sliding prone part by circling the grid according to the slope exceeding a steep slope threshold and the slope direction pointing to the free surface; calculating the elevation time series differential of each grid after the multiple period elevation images are registered; separating a creep component according to the spatial coherence of the elevation time series differential along the direction domain in the direction domain determined by the slope direction, and obtaining a slope surface creep rate; normalizing the slope, free surface height difference and creep rate of the sliding prone part, weighting and scoring, mapping the sliding risk grade and outputting the sliding risk grade, and identifying the elevation time series differential submerged in reconstruction noise in the slope direction domain according to the spatial coherence, and separating the creep when the elevation micro-change amplitude does not exceed the reconstruction noise level.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to a method for image-based risk assessment of slope slippage in loess gully areas. Background Technology

[0002] The loess ridge and gully region is characterized by fragmented terrain, steep slopes, and deep gullies. Highways often extend along valleys and slope toes, resulting in dense and long stretches of slopes along these routes. Loess has a loose structure, well-developed vertical joints, and is prone to disintegration when exposed to water. Disturbances caused by rainfall infiltration, slope toe excavation, and vehicle vibrations pose a constant threat to road traffic and the safety of maintenance personnel. The key challenges in preventing landslides on these slopes are threefold: First, the slopes extend continuously for kilometers or even tens of kilometers along the highway, making it inefficient and incomplete for maintenance personnel to conduct visual inspections section by section on foot or by car. Furthermore, personnel cannot easily access steep sections. Second, loess landslides often undergo a long period of extremely slow slope creep before instability. During this stage, there are no visible cracks or faults on the slope, making early signs of landslides difficult to identify. Third, once significant displacement or obvious cracks appear on the slope, the slope has often entered a stage of accelerated deformation, nearing overall instability, leaving extremely limited time for warning and intervention. Therefore, how to reliably detect the precursors of landslides in the early stages when the slope is still in a state of slight creep and no obvious changes in appearance is the core challenge of disaster prevention for highway slopes in the Loess Gully Area.

[0003] With the development of UAV aerial photography and digital photogrammetry technology, the industry has begun to use UAVs to collect slope images and reconstruct three-dimensional terrain to obtain slope elevation information and deformation data. Patent application CN105783878A discloses a method for slope deformation detection and measurement based on remote sensing by a small UAV. This method involves setting up image control points and using a small UAV for aerial photography. Through digital photogrammetry, a three-dimensional point cloud and digital elevation model of the slope are generated. After registering the point clouds in two phases in a stable, unchanging area outside the slope, a grid calculator is used to calculate the difference between the two phases of the digital elevation model grid by grid. Areas with positive, negative, and zero differences are respectively identified as areas of increased terrain elevation, decreased terrain elevation, and areas of no change. Based on this, the deformation area and the increase or decrease in material volume of the slope are delineated and measured. However, before calculating the difference between the elevations of the two periods, this scheme explicitly requires that the root mean square error of the elevation results must be less than the estimated minimum slope deformation and at least one order of magnitude lower. In other words, the scheme only recognizes elevation changes as true deformations when the amplitude exceeds the noise level of a single-period 3D reconstruction, while filtering out minute elevation changes whose amplitudes are submerged in reconstruction noise as noise within the amplitude domain. This approach means that the scheme can only detect slope deformations that have developed to a significant amplitude, but it fails to identify the most valuable early signs of slippage, namely, minute creep. This is because the elevation changes of loess slopes in the creep stage are usually only on the order of millimeters and are completely masked by reconstruction noise. By the time the amplitude of the elevation change grows to the point where it can cross the noise threshold and be detected by the scheme, the slope is often already close to instability, which directly corresponds to the aforementioned pain point that significant displacement is already close to instability.

[0004] Furthermore, Chinese patent application CN117789064A discloses a highway slope UAV inspection and early warning system based on multi-source data. This system uses a UAV to periodically detect the degree of cracking, deformation, and displacement of highway slopes, and calculates a risk coefficient by weighting these parameters to determine the slope's risk status. However, the cracking, deformation, and displacement parameters relied upon by this system are also significant signs of slope development that can be clearly distinguished in images. It uses fixed weights for linear weighting of each parameter, failing to introduce topographical quantification elements such as slope aspect and airspace to locate the true slippery areas, and also failing to provide identification methods for early, minute creep submerged in reconstruction noise. Therefore, like CN105783878A, it cannot solve the core technical problem of reliably capturing early signs of slippage in the early stages when the slope appearance shows no obvious changes.

[0005] Therefore, existing slope monitoring schemes based on UAV aerial photography share a common limitation: they rely on the amplitude range approach of using whether the elevation change amplitude exceeds the reconstruction noise threshold as the criterion for true deformation, thus failing in the early stages of minute creep. Overcoming this amplitude range bottleneck and reliably separating creep precursors and quantifying and classifying slip risk under conditions where minute changes in slope elevation are still submerged in reconstruction noise is a pressing technical problem in this field. Summary of the Invention

[0006] To address the core bottleneck in existing technologies for disaster prevention of highway slopes in loess gullies, which relies on whether the elevation change amplitude exceeds the 3D reconstruction noise threshold as the criterion for true deformation, thus failing to identify precursors of landslides when the slope is still in the stage of slight creep, this invention provides an image-based risk classification and assessment method for slope landslides in loess gullies. By modifying the temporal differences in elevation submerged in reconstruction noise to be identified by spatial coherence within a directional domain determined by the slope aspect, the method separates the aspect-coherent slight elevation changes from the isotropic reconstruction noise into creep components. Under the constraint that the amplitude of slight elevation changes does not exceed the reconstruction noise level, and based on the physical mechanism that the slow sliding of the slope along the direction of gravity inevitably leads to spatially coherent changes in the slope elevation along the slope aspect, reliable identification and risk classification of early landslide precursors are achieved.

[0007] The technical solution of this invention is: a method for image-based classification and assessment of slope slip risk in loess gully areas, comprising the following steps: acquiring multi-period aerial images of a target slope taken by a UAV, reconstructing the three-dimensional topography of the slope and rasterizing it into elevation images for each period; calculating the slope and aspect grid by grid for the current elevation image, delineating high and steep free zones based on grids where the slope exceeds the steep slope threshold and the aspect points to the free surface, as slip-prone areas, where the free surface is the free boundary on one side of the slope toe; registering the multi-period elevation images to the same reference and then calculating the elevation time-series difference grid by grid, separating the creep component from the elevation time-series difference in the direction domain determined by the aspect based on the spatial coherence of the elevation time-series difference along the direction domain, and obtaining the slope creep rate; for the slip-prone areas, normalizing their slope, the free elevation difference relative to the free surface, and the creep rate, then weighting and scoring them to map them into a slip risk level and outputting the classification and assessment results.

[0008] Furthermore, when separating creep components, the slope aspect of each grid is used as the reference direction for the direction domain. The spatial coherence is obtained by statistically analyzing the consistency of the projection sign and the continuity of the amplitude of the temporal differences in elevation of the neighborhood along this direction. Components with a spatial coherence not less than the coherence threshold are identified as creep components. The direction of the principal eigenvector of the neighborhood elevation gradient covariance matrix is ​​used as the neighborhood principal slope aspect to replace the slope aspect as the reference direction for the direction domain. Furthermore, a creep extension zone is obtained by regional growth in the creep coherence intensity field using a high-steep free zone as a seed zone and incorporated into the slippery area. The slope, free zone elevation difference, and creep rate are quantized and adaptively weighted according to their respective correlation strength with slip instability. The confidence level of elevation reconstruction is obtained based on ground point density and vegetation cover index. Grids with insufficient confidence are gated out from the spatial coherence statistics and classification.

[0009] The beneficial effects of this invention are as follows:

[0010] Firstly, this invention modifies the temporal differences in elevation obtained grid-by-grid after multi-period elevation image registration by using spatial coherence within the directional domain determined by the slope aspect, rather than filtering them out using a noise threshold in the amplitude domain. This allows for the separation of creep components even when minute elevation changes are still submerged in reconstruction noise. When a slope is in the creep stage, it slowly slides down along the direction of gravity (slope aspect), inevitably causing coherent changes in slope elevation along the slope aspect that are consistent in direction and continuous in the neighborhood. In contrast, 3D reconstruction noise is an isotropic random disturbance. The slope aspect coherence criterion precisely utilizes the essential difference in their spatial structure to distinguish the structured weak creep signal from isotropic noise. Compared to only recognizing true deformation when the elevation change amplitude exceeds the reconstruction noise threshold, this invention advances the identifiable point of slip precursors from the obvious displacement stage to the minute creep stage, overcoming the deficiency of existing amplitude domain schemes that fail in the early creep stage.

[0011] Secondly, this invention first quantifies and delineates steep, high-altitude free-face areas as slippery locations by slope and aspect grid by grid. Slope aspect is used both to locate slippery locations and as the reference direction for separating creep components in the directional domain. One slope aspect element plays a role in two places. On loess slopes, the gravity sliding component is the largest and the constraint is the weakest in areas with steep slopes and high free-face areas, making them the most likely places for creep to initiate. The slope aspect at these locations is precisely the direction of slope sliding, which is also the coherent direction of creep elevation changes. By reusing the slope aspect as the directional domain reference, the topographic quantification and creep coherent separation are coupled on the same directional element. The two mutually corroborate each other, and their synergistic effect exceeds the sum of their individual effects. Slope aspect alone is only a static slippery factor, and elevation temporal differences alone are only random noise. Only by identifying elevation temporal differences in the directional domain defined by slope aspect can the originally unmeasurable micro-creep be transformed into a measurable creep rate.

[0012] Thirdly, this invention uses the principal eigenvector direction of the neighborhood elevation gradient covariance matrix as the neighborhood principal slope direction, replacing the grid-by-grid slope direction as the reference direction for the direction domain. In steep, high-altitude areas, the micro-topography at the top and rear edges is fragmented, and the single-grid slope direction is easily affected by local unevenness and rasterization errors, causing fluctuations. If the fluctuating single-grid slope direction is directly used as the reference for the direction domain, it will cause directional drift in coherence statistics, leading to missed creep detection. The neighborhood principal eigenvector direction, based on the dominant trend of the elevation gradient within the neighborhood, suppresses single-grid fluctuations, stabilizes the direction domain reference, and thus ensures the reliability of creep separation. Furthermore, the dispersion of the eigenvalues ​​of this covariance matrix simultaneously characterizes the fragmentation of the slope micro-topography, providing additional information for subsequent interpretation and achieving multiple uses from a single calculation.

[0013] Fourth, this invention uses a high-steep free zone as a seed region and grows a creep extension region within a creep coherence intensity field, which is then incorporated into the slippery area. The high-steep free zone defined by static topography reflects the current slope shape, while the coherence zone of a creeping slope often extends beyond the static high-steep free zone boundary along the slope direction. If only static boundary is used for evaluation, adjacent parts that have already begun to creep will be missed. By using dynamic creep coherence intensity to correct the boundary of the slippery area, the definition of the slippery area is upgraded from a purely static topography criterion to a joint criterion of static topography and dynamic creep. Compared with existing schemes that only rely on static stability coefficients or fixed thresholds for judgment, this avoids omissions caused by creep exceeding the boundary.

[0014] Fifth, this invention first performs quantile normalization on slope, free-of-sight elevation difference, and creep rate, and then adaptively weights them according to their respective correlation strength with slip instability. Slope, free-of-sight elevation difference, and creep rate have different dimensions and vastly different numerical distributions. If they are directly weighted linearly with fixed weights, the larger factors will overwhelm the smaller but more sensitive factors. Quantile normalization eliminates the differences in dimensions and distributions, and adaptive weighting gives higher weights to factors that contribute more. This results in a nonlinear risk leap when the three factors overlap in a high-altitude free-of-sight zone with steep slope, large free-of-sight elevation difference, and high creep, which is more effective in revealing truly high-risk areas than fixed-weighting.

[0015] Sixth, this invention calculates the confidence level of elevation reconstruction based on ground point density and vegetation cover index, and gates out questionable grids with insufficient confidence from spatial coherence statistics and risk classification. In vegetation-shaded areas, the true elevation of the slope is replaced by the canopy elevation, making the reconstruction unreliable. Subtle temporal differences in elevation may produce false directional coherence due to the canopy's wind-driven movement, leading to misclassification as creep. By using reconstruction confidence gating, questionable grids are excluded from coherence statistics and explicitly marked as questionable in the results, avoiding false creep misclassification induced by vegetation shading and ensuring the reliability of the classification results. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the method of the present invention;

[0017] Figure 2 This is a flowchart of the process of reconstructing three-dimensional terrain and rasterizing it into an elevation image in this invention;

[0018] Figure 3 This is a flowchart of the process for calculating slope and aspect and delineating steep free-range areas in this invention;

[0019] Figure 4 This is a flowchart of the separation of creep components based on spatial coherence in the directional domain in this invention;

[0020] Figure 5 This is a flowchart of the re-determination of the easy-slip region boundary and the normalized weighted classification of the creep coherence intensity field in this invention;

[0021] Figure 6 This is a flowchart of the elevation reconstruction confidence gating in this invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The embodiments described are only some, not all, of the embodiments of the present invention; all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, this embodiment takes a section of a secondary highway cutting slope extending along a gully in a loess ridge and gully area in Shaanxi Province as the object. The slope extends for about 1.2 km, with a slope height of 18 m to 34 m and a slope angle of 25° to 58°. The slope surface is partially covered with sparse grass and shrubs. The image classification assessment method for slope slippage risk in the loess gully area of ​​the present invention is implemented, including steps S1 to S5.

[0024] Step S1: Acquire aerial images of the target slope taken by the UAV in multiple phases, reconstruct the three-dimensional terrain of the slope and rasterize it into elevation images of each phase.

[0025] like Figure 2As shown, a multi-rotor UAV equipped with a positioning and attitude determination system was used to conduct phased aerial photography of the target slope along a grid route parallel to the slope's direction. The time interval between two consecutive phases of aerial photography was 30 days, and a total of 5 consecutive phases of imagery were collected. To ensure the accuracy of 3D reconstruction, the flight path was planned to ensure that the forward overlap rate of the aerial photography was no less than 75% and the lateral overlap rate was no less than 60%. The flight altitude was dynamically adjusted along the elevation at the slope toe and top to keep the relative flight altitude from the camera to the slope surface between 80m and 120m, and the ground sampling interval was controlled at 0.03m. Image control points were evenly distributed inside and outside the target slope, and the 3D coordinates of each image control point were obtained by real-time dynamic measurement using satellite positioning. For areas with steep slopes and inaccessible to personnel, a prism-free measurement method was used for supplementary control. Feature matching, aerial triangulation, and regional network adjustment were performed on the images from each phase of aerial photography. After adding image control point constraints, the 3D topography of the slope was reconstructed through dense matching to obtain the 3D point cloud of the slope. The 3D point cloud of the slope is denoised and ground point filtered to remove isolated points and points covered by slope cover. The remaining ground points are interpolated to generate a regular grid elevation image at a preset grid resolution. The same grid resolution is used for all elevation images in each period. In this embodiment, the grid resolution is 0.05m, meaning the grid resolution is between 0.02m and 0.10m. During the generation of elevation images for each period, it is necessary to ensure that the results of each period have a consistent absolute accuracy benchmark in terms of both planar position and elevation. Otherwise, the systematic deviation between multiple periods will overwhelm the millimeter-level creep signal that this invention aims to capture. Therefore, in this embodiment, the same subset of image control points is introduced into the adjustment of each period's aerial triangulation. After adjustment, another set of independent image control points is used as checkpoints to evaluate the absolute elevation error of each period's elevation image. The results of that period are accepted only if the absolute elevation error is no greater than 0.01m; otherwise, the subset of image control points used in the adjustment is readjusted and reconstructed until the accuracy requirements are met. Furthermore, to suppress the dense matching voids caused by changes in lighting and shadow on the loess slope, multi-view stereo matching was employed during the dense matching stage, and semi-global constraints were applied to areas with weak texture, ensuring the continuous and complete reconstruction of the slope elevation. After the above processing, the absolute elevation mean square errors of the five consecutive elevation images in this embodiment were 0.008m, 0.007m, 0.009m, 0.008m, and 0.007m, respectively, all meeting the requirement of not exceeding 0.01m, providing the necessary accuracy prerequisite for separating millimeter-level creep in a noisy background. Thus, the slope topography was upgraded from a planar photograph without elevation information to an elevation image with raster-by-raster elevation values, laying the data foundation for subsequent risk quantification based on terrain. Both the reconstruction of the three-dimensional slope terrain and the rasterization of the elevation images adopted mature photogrammetric processing procedures in this field, and their algorithm details will not be elaborated further.

[0026] During the reconstruction of the 3D terrain, the elevation reconstruction confidence score of each grid cell is calculated simultaneously for gating in subsequent step S5. Specifically, the number of ground points falling into each grid cell in the 3D point cloud of the slope is counted and divided by the grid cell area to obtain the ground point density of that grid cell, in pts / m²; simultaneously, the visible vegetation cover index of that grid cell is extracted from the orthophoto of the aerial photography. The elevation reconstruction confidence score is calculated using the following formula: ,in: Let be the confidence level for elevation reconstruction, be a scalar with a value ranging from 0 to 1, dimensionless, and calculated using this formula, representing the [missing value]. The elevation values ​​at each grid point in the image reflect the reliability of the actual slope surface; The raster number is a positive integer scalar with a value ranging from 1 to the total number of raster cells. It is dimensionless and used for raster-by-raster indexing. It is an S-shaped compression function, and its definition is: This monotonically compresses the real independent variable to the interval between 0 and 1, making it dimensionless. Its function is to normalize the weighted score into a confidence probability. It is a natural constant, approximately 2.718, and is dimensionless; For the first The ground point density of a grid is a scalar with a value ranging from 0 to 5000 and a unit of pts / m². It is obtained by dividing the number of ground points falling into the grid by the grid area and represents the sufficiency of effective ground points on which the grid reconstruction is based. The reference ground point density is a scalar with a value of 500 and a unit of pts / m². It is determined by the typical value of the ground point density in the exposed slope area. Its function is to make the ground point density dimensionless. If the value is too large, all grid confidences will be too low; if it is too small, the discrimination will be lost. For the first The vegetation cover index of each grid cell is a scalar with a value ranging from 0 to 1. It is dimensionless and is calculated from the red, green, and blue components of the orthophoto of that grid cell according to the visible light vegetation index. It represents the degree to which the grid cell is occluded by vegetation. is the ground point density gain coefficient, a scalar with a value of 3.0, dimensionless, obtained from sample calibration, and its function is to adjust the contribution intensity of ground point density to the confidence level; is the vegetation suppression coefficient, a scalar with a value of 4.0, dimensionless, obtained from sample calibration, and its function is to adjust the suppression strength of vegetation shading on confidence level; This is the bias term, a scalar with a value of 1.0. It is dimensionless and is obtained from sample calibration. Its function is to set the decision baseline for confidence level.

[0027] Step S2 involves calculating the slope and aspect grid-by-grid on the current elevation image. Grids with slopes exceeding the steepness threshold and aspects pointing towards the open air are designated as high-slope, easily slippery areas. For example... Figure 3As shown, the latest elevation image is taken as the current elevation image. The first-order difference of elevation in the east-west and north-south directions is calculated grid-by-grid for the current elevation image to obtain the elevation gradient of that grid. The slope of that grid is obtained by converting the magnitude of the elevation gradient; the slope is the steepness of the elevation descent along the elevation gradient direction, in degrees. The aspect of that grid is obtained by converting the azimuth of the elevation gradient; the aspect is the azimuth angle of the steepest descent direction of the slope, ranging from 0° to 360°. The above grid-by-grid calculation of slope and aspect is a mature topographic factor calculation method, and its calculation details will not be elaborated further.

[0028] After obtaining the slope and aspect of each grid cell, the free surface is determined. The free surface is the free boundary on the toe side of the slope, that is, the free interface formed by the slope extending downward to the road surface or the bottom of the valley, which is identified by the lowest elevation zone of the slope toe line in the elevation image. Then, grid cell by grid cell, if the slope of a grid cell exceeds the steep slope threshold and the slope aspect of the grid cell points to the free surface, the grid cell is included in the high-steep free area. In this embodiment, the steep slope threshold is 38°, that is, the steep slope threshold is in the range of 30° to 45°; the slope aspect pointing to the free surface means that the angle between the slope aspect of the grid cell and the direction in which the grid cell points to the free surface is less than 90°, indicating that the slope of the grid cell faces the free side, its gravity sliding component points to the free surface and lacks lateral constraint. Grid cells that meet the above two conditions are connected and aggregated to delineate the high-steep free area, and the high-steep free area is designated as a slippery area. The physical basis for determining the slope aspect pointing towards the free face is as follows: the instability of loess slopes is essentially the sliding of the slope body along the free face under its own weight. When the slope faces the free face, the downward component of gravity along the slope points towards the free face, lacking lateral soil and rock constraint support, making it most prone to creep and eventually sliding. Conversely, the part of the slope facing away from the free face is constrained by the pressure of the slope behind, resulting in relatively higher stability. Therefore, only grids with steep slopes and slope aspects pointing towards the free face are included in the high-steep free face zone. This not only conforms to the instability mechanism of loess slopes but also avoids misclassifying stable parts with steep slopes but facing away from the free face, narrowing the scope of subsequent creep separation and classification processing and improving targeting. During connectivity aggregation, grids that meet both the steep slope and free face orientation conditions are aggregated into several connected domains using the eight-connectivity criterion. Scattered grids with areas smaller than the preset minimum patch area are removed to suppress isolated misjudgments. In this embodiment, the minimum patch area is set to 0.25m². Therefore, risk assessment is transformed from the visual experience of maintenance personnel to a topographic criterion quantified grid by grid based on slope and aspect, solving the dilemma that it is impossible to judge the steepness of the slope or locate the slippery areas based solely on planar images. Simultaneously, for each grid within the slippery area, the difference between its elevation and the elevation of the free surface is taken as the free height difference of that grid relative to the free surface, in meters, for subsequent grading in step S4. The larger the free height difference, the higher the potential energy of the grid relative to the free surface at the toe of the slope; once instability occurs, the sliding kinetic energy and the resulting damage are greater. Therefore, the free height difference is an important topographic quantification factor characterizing slip risk.

[0029] Step S3: After registering multiple elevation images to the same reference, the elevation time-series difference is calculated grid by grid. Within the direction domain determined by the slope aspect, the creep component is separated from the elevation time-series difference based on the spatial coherence of the elevation time-series difference along the direction domain to obtain the slope creep rate.

[0030] like Figure 4 As shown, firstly, multiple periods of elevation images are registered to the same reference. An exposed bedrock area or road surface area outside the target slope is selected as a stable feature area. This stable feature area does not deform during the monitoring period. Using the consistency of the stable feature area's elevation in each period's elevation image as a constraint, registration parameters, including translation and rotation, are obtained for each period's elevation image relative to the reference period. These registration parameters are then used to register the elevation images of each period to the same reference, thereby eliminating the overall system deviation introduced by random errors in the positioning and attitude determination system during aerial photography. The selection of the stable feature area should ensure that its shape and position remain basically unchanged during the monitoring period. In this embodiment, a section of exposed bedrock cliff and a section of hardened road surface outside the slope are selected to form a stable feature area. These two sections are located on opposite sides of the slope toe, allowing for simultaneous constraint of translation and rotation registration parameters, avoiding under-constraint registration caused by selecting only a single area. During registration, the registration parameters were iteratively solved with the objective of minimizing the sum of squared elevation residuals of the corresponding rasters in the elevation images of stable feature areas across different periods. After convergence, the root mean square of the elevation residuals of stable feature areas decreased to 0.006 m, indicating that the elevation images of each period had been registered to a unified benchmark. After registration, the difference between adjacent registered elevation images was calculated grid by grid, i.e., the elevation of the previous period's rasters was subtracted from the elevation of the next period's rasters, to obtain the temporal difference of the elevation of that rasters, in meters. When the loess slope is in the early creep stage, this temporal difference of elevation is usually only on the order of millimeters, and its amplitude is submerged in the 3D reconstruction noise, making it difficult to distinguish from noise directly in the amplitude domain. Existing methods based on multi-period elevation difference can only detect significant deformations because their criteria are based on the amplitude of elevation changes. Only when the elevation change amplitude of a grid exceeds the reconstruction noise level is it considered a true deformation. However, the amplitude of elevation temporal differences during the creeping phase does not exceed the noise level and is therefore inevitably filtered out. This invention recognizes that the criteria should not be based on the amplitude of a single grid, but rather on the spatial structure of the elevation temporal differences between neighboring grids: during slope creep, the entire structure slides down the slope, and the elevation temporal differences between adjacent grids exhibit a spatially coherent structure with consistent direction and gradually changing magnitude along the slope; while 3D reconstruction noise originates from matching random errors, is isotropic in direction, and independent between adjacent grids. The amplitudes of the two can be comparable, but their spatial structures are drastically different. Therefore, this invention abandons filtering by noise thresholds in the amplitude domain and instead uses spatial coherence discrimination in the directional domain determined by the slope direction, allowing creep amplitudes submerged in noise to be identified by structural features.

[0031] To separate creep in a noisy environment, this invention does not use threshold filtering in the amplitude domain, but instead examines the spatial coherence of elevation temporal differences in the direction domain determined by the slope aspect. Before calculating the spatial coherence, a neighborhood window is taken centered on each raster in the current elevation image, and the covariance matrix of the elevation gradient within the neighborhood window is calculated to suppress single-raster slope aspect jitter. The covariance matrix is ​​calculated using the following formula: ,in: For the first The neighborhood elevation gradient covariance matrix of each grid cell is a 2x2 symmetric matrix with dimensionless elements, calculated using this formula, and represents the value of the elevation gradient covariance matrix of each grid cell. The directional distribution structure of the elevation gradient within the neighborhood of the center; The center grid number is defined in the same confidence formula as described above; For The set of grid cells contained in the neighborhood window centered on the grid cell. In this embodiment, a 7×7 grid window is used, and the number of its elements is constant. Let be the grid number within the neighborhood window, a positive integer scalar, and iterate through... ; For grid Relative center grid The Gaussian weight is a scalar with a value range of 0 to 1 and is dimensionless. It is calculated by the second formula in this paper and its function is to make the contribution of neighboring grids greater and the contribution of distant grids attenuated. For grid The elevation gradient is a 2D column vector whose components are the ratios of elevation differences to horizontal spacing. It is dimensionless and derived from the elevation image. The central difference at the point is used to characterize the descent direction and steepness of the grid slope; for transpose, superscript The transpose operation represents a vector or matrix; , They are grids , The planar position coordinates are 2D column vectors in meters, obtained by multiplying the raster row and column numbers by the raster resolution. The L2 norm of a vector; The Gaussian kernel scale parameter is a scalar with a value of 0.15 in meters. Determined by three times the raster resolution, it controls the spatial range of neighborhood weighting. A value that is too large will over-smooth and mask the true slope aspect changes, while a value that is too small will insufficiently suppress jitter. The covariance matrix is ​​eigenvalued, and the direction of the eigenvector corresponding to its largest eigenvalue is taken as the dominant slope aspect of the raster's neighborhood. The sign of this eigenvector is then assigned to the side pointing towards the elevation decrease, yielding the unit vector of the reference direction in the direction domain. Replacing the raster's slope aspect with the dominant slope aspect of the neighborhood as the reference direction in the direction domain effectively suppresses single-raster slope aspect jitter caused by micro-topographic fragmentation at the top and rear edges of steep, high-altitude areas. Simultaneously, the relative dispersion of the two eigenvalues ​​of the covariance matrix is ​​converted into slope micro-topographic fragmentation using the following formula and output: ,in: For the first The slope micro-topographic fragmentation of each grid is a scalar with a value ranging from 0 to 1. It is dimensionless and is calculated by this formula. It characterizes the degree of fragmentation and dispersion of the slope in the neighborhood of the grid. The closer its value is to 1, the more fragmented the slope is. , These are the covariance matrices. The largest and smallest eigenvalues ​​are both scalars, satisfying the following condition: Dimensionless, derived from the pair We obtain it by characteristic decomposition. Characterizing the elevation gradient energy in the dominant slope direction, Characterizes the elevation gradient energy in the sub-direction; To prevent extremely small positive numbers with a denominator of zero, let be a scalar and take the value . Dimensionless. When the slope is smooth, the elevation gradient is concentrated in the direction of elevation gradient. Much larger , Approaching 0; when the slope is fractured, the elevation gradient direction is scattered. near , Approaching 1.

[0032] After obtaining a stable reference direction for the direction domain, the main slope direction of the neighborhood of each grid is used as the reference direction for the direction domain. The spatial coherence of the elevation time series differences of the grid and its neighboring grids along the reference direction is examined. The sign consistency and amplitude continuity of the projection components are statistically analyzed to obtain the spatial coherence of the grid. The spatial coherence is calculated by the following formula: ,in: For the first The spatial coherence of a grid is a scalar with a value ranging from 0 to 1. It is dimensionless and is calculated by this formula. It characterizes the degree of coherence of the elevation time series difference in the neighborhood of the grid along the reference direction of the direction domain. The closer its value is to 1, the more coherent the elevation time series difference is along the reference direction of the direction domain and the more likely it is to originate from creep. , , , The definition is the same as the formula mentioned above; For grid The elevation time-series differential spatial gradient is a 2D column vector whose components are the ratio of the difference between adjacent grid cells to the horizontal spacing of the elevation time-series differential, and are dimensionless. This is derived from the elevation time-series differential field in... The central difference at the point is obtained, which characterizes the spatial direction and rate of change of the time-series elevation difference; For the first The reference direction unit vector of each grid is a 2-dimensional column vector with a magnitude of 1 and is dimensionless. It is obtained by taking the direction of the main slope of the neighborhood and pointing to the side of the elevation decrease, representing the direction of the slope gravity sliding down. Represents the dot product operation of vectors; This represents the projection vector obtained by projecting the spatial gradient of elevation time-series differentials onto the reference direction in the direction domain; This represents the L2 norm of the vector. When the spatial gradients of the elevation time-series differences of all grid cells in the neighborhood are consistent along the reference direction of the orientation domain, the sum of the numerator vectors does not cancel each other out and approaches the denominator. Approaching 1; when the spatial gradient direction of the elevation time-series differential is isotropic and random, the numerator vectors cancel each other out and approach 0, while the denominator remains positive. Approaching 0, thus distinguishing the slope-coherent creep signal from the isotropic reconstruction noise.

[0033] Furthermore, elevation time-series differences with spatial coherence not less than the coherence threshold are identified as creep components, while those with spatial coherence less than the coherence threshold are identified as noise and eliminated. The slope creep rate is obtained by dividing the elevation time-series difference identified as a creep component by the time interval between two adjacent aerial photographs. The conversion between the above identification and creep rate is performed using the following formula: ,in: For the first The slope creep rate of each grid is a scalar, ranging from 0 to 0.01, with units of m / d. It is calculated by this formula and characterizes the speed of creep of the grid slope along the reference direction of the directional domain. For the first The elevation time-series difference of each grid is a scalar with a value range of -0.05 to 0.05 and a unit of meters. It is obtained by calculating the difference of each grid in two adjacent registered elevation images and represents the small change in the elevation of the grid over time. The time interval between two adjacent aerial photography sessions is a scalar value, which is taken as 30 in this embodiment, and the unit is days (d). It is determined by the aerial photography plan. The definition is the same as the aforementioned formula for spatial coherence; The coherence threshold is a scalar with a value of 0.6, ranging from 0.5 to 0.8. It is dimensionless and is determined by the working characteristic curve of the sample subjects. Its function is to separate the creep component from noise. If the value is too large, weak creep will be missed; if the value is too small, noise will be mistakenly included. This is an indicator function, taking the value 1 when the condition within the parentheses is true and 0 when it is false. It is dimensionless and its function is to retain only the creep rate of the grid cells identified as creep components. Therefore, the identification of slip precursors is advanced from obvious displacement to minute creep, compensating for the shortcomings of visual inspection and amplitude domain schemes in detecting early creep.

[0034] Step S4: Redefine the boundary of the slippery part using the creep coherence intensity field, and normalize the slope, free height difference and creep rate of the slippery part and then weight it to score it, mapping it to the slip risk level.

[0035] like Figure 5 As shown, the spatial coherence of each grid obtained in step S3 is organized into a creep coherence intensity field according to the grid position. Considering that the coherence region of a creeping slope often extends beyond the statically defined high and steep free zone along the directional domain reference direction, the high and steep free zone is used as the initial seed region. Region growth is performed on grids in the creep coherence intensity field whose spatial coherence is not less than the coherence threshold, which are connected to the initial seed region and whose directional domain reference direction is compatible, to obtain the creep expansion region. The directional domain reference direction compatibility means that the cosine of the angle between the grid to be grown and its adjacent grids already in the region is not less than the directional compatibility threshold. In this embodiment, the directional compatibility threshold is taken as cos30°. The specific execution of the region growth is as follows: each grid in the initial seed region is placed into the growth queue. The first grid in the queue is taken out sequentially, and the grids in its eight neighbors that have not yet been included in the region are examined. If the spatial coherence of an adjacent grid is not less than the coherence threshold, and its directional domain reference direction is compatible with the directional domain reference direction of the first grid in the queue, then the adjacent grid is merged into the creep expansion region and added to the growth queue. This process is iterated until the growth queue is empty. The reason for introducing the directional domain reference direction compatibility constraint is that connecting only by spatial coherence can easily lead to the erroneous merging of two coherent regions with different directions. Adding the directional compatibility constraint can ensure that growth only extends along the creep region of the slope in the same downward direction, avoiding erroneous merging by crossing the watershed of abrupt slope aspect changes. The creep extension zone is incorporated into the slippery area, thereby updating the boundary of the slippery area from a purely static topographic delineation to a joint definition of static topography and dynamic creep, avoiding missed detections due to creep exceeding the boundary; conversely, if a static, steep, and open area has not detected coherent creep in multiple monitoring periods, its slippery area boundary will not be unnecessarily expanded due to static steepness, thus allowing the assessment to focus on the slope that is truly in an active state.

[0036] For each grid cell within the updated slippery area, its slope, relative elevation difference to the free surface, and creep rate are normalized and then weighted for scoring. Considering the different dimensions and significant numerical distributions of slope, elevation difference, and creep rate, quantile normalization is first performed on each cell according to its empirical distribution within the slippery area. Then, adaptive weights are determined based on their correlation strength with slip instability, and the weighted sum is used to obtain the grid slip risk score. The normalized weighted score is calculated using the following formula: ,in: For the first The grid slip risk score for each grid cell is a scalar with a value ranging from 0 to 1. It is dimensionless and is calculated by this formula, representing the overall slip risk level of the grid cell. The definition is the same as before; Use the factor identifier to traverse the set , representing the three factors: slope, elevation difference, and creep rate, respectively; Slope, in degrees; The height difference relative to the free surface is expressed in meters (m). The slope creep rate, as mentioned above The unit is m / d; For the first raster factor The original value, whose unit depends on the factor; As a factor The empirical distribution function in the slippery area has a value range of 0 to 1 and is dimensionless. It is obtained by dividing the rank of all grid values ​​of the factor in the slippery area by the total number of grids. Its function is to uniformly map factors of different dimensions to quantile normalization values ​​of 0 to 1, thereby eliminating the difference between dimensions and distribution. As a factor The adaptive weights are scalars, ranging from 0 to 1, and are dimensionless. They are calculated by the second formula of this paper, and the sum of the weights of the three factors is 1. Their function is to give higher weights to factors that are more strongly correlated with slip instability. As a factor The correlation strength with slip instability is a scalar, ranging from -1 to 1, dimensionless, and is obtained from the correlation coefficient between the value of this factor and historical slip instability labels; Indicates taking the absolute value; To sum using factors, iterate through the set. Quantile normalization allows slope, free-of-sight elevation difference, and creep rate to participate in weighting on the same 0 to 1 scale, preventing large-scale factors from drowning out smaller but more sensitive factors. When the quantile normalized values ​​of the three factors in the same grid are simultaneously high, i.e., when steep slope, large free-of-sight elevation difference, and high creep rate are superimposed, the weighted score produces a nonlinear jump, thus highlighting the truly high-risk areas.

[0037] Finally, the grid slip risk scores are mapped to slip risk levels according to preset grading thresholds: when the grid slip risk score is less than the first grading threshold, it is determined to be low risk; when it is not less than the first grading threshold and less than the second grading threshold, it is determined to be medium risk; and when it is not less than the second grading threshold, it is determined to be high risk. In this embodiment, the first grading threshold is set to 0.4 and the second grading threshold is set to 0.7. The slip risk levels of each grid are organized according to the grid position and output as grading assessment results, thus obtaining the image grading assessment results of slope slip risk in the Loess Gully Area.

[0038] As supporting evidence, grayscale dual-threshold edge extraction is performed on the orthophoto corresponding to the current elevation image. Pixels with grayscale values ​​not less than the high threshold are marked as strong edges, and pixels with grayscale values ​​between the high and low thresholds and connected to the strong edges are marked as weak edges. Slope cracks are obtained by connecting the strong and weak edges, and the crack length per unit area in the slippery area is counted as the crack density. When the crack density exceeds the crack density threshold, the slip risk level of the slippery area is upgraded by one level, thereby making the detection of slope cracks an independent piece of evidence for slip risk, and upgrading the risk from a single-factor judgment to a multi-factor comprehensive classification.

[0039] Step S5: Gating of questionable graticules based on elevation reconstruction confidence level. For example... Figure 6As shown, the confidence level of the elevation reconstruction calculated synchronously in step S1 is taken, and the gratings with an elevation reconstruction confidence level less than the confidence threshold are marked as questionable gratings. In this embodiment, the confidence threshold is 0.5. The questionable gratings are not included in the spatial coherence statistics in step S3. That is, when calculating the spatial coherence of a certain grating, the questionable gratings in its neighborhood window are not included in the summation, so as to prevent the vegetation canopy from swaying with the wind and generating false directional coherence at the questionable gratings, which would be misjudged as creep. At the same time, the questionable gratings are marked as questionable in the graded evaluation results output in step S4, and the slip risk level is not output, so as to explicitly indicate that the elevation reconstruction of this section is unreliable due to vegetation shading and its risk conclusion is questionable. It is particularly important to note that in the Loess gully region, pseudo-creep induced by vegetation shading is highly deceptive: the swaying of the grass and shrub canopy under wind forces causes multiple elevation images to exhibit similar directional and continuous elevation temporal differences within the same shading area. Their spatial structure closely resembles true creep, and without differentiation, they can easily be mistakenly classified as creep by spatial coherence criteria, leading to falsely reported high risks during classification. This invention uses elevation reconstruction confidence as a pre-gating mechanism, eliminating questionable grids precisely before spatial coherence statistics. This cuts off the path for pseudo-creep to enter creep separation and classification at the source. This is precisely the value of coupling elevation reconstruction confidence gating with directional domain coherence separation; the former ensures that the elevation temporal differences processed by the latter truly originate from a reliably reconstructed real slope, enabling the latter to robustly separate true creep from reliable data. Therefore, sections where vegetation obstruction makes elevation reconstruction unreliable are marked and removed from relevant statistics and classifications, resulting in an assessment that allows for terrain reconstruction, easy slippage and location, creep tracking and risk classification, with clear doubtful markings for unreliable sections.

[0040] In summary, steps S1 to S5 of this invention form a closed-loop process with sequential coupling and mutual feedback between static and dynamic elements: Step S1 reconstructs the three-dimensional terrain and rasterizes the elevation image, and simultaneously calculates the elevation reconstruction confidence level; Step S2 calculates the slope and aspect from the elevation image, delineates the high and steep free zone and the free elevation difference, wherein the slope aspect not only locates the slippery parts but also provides the reference direction for the direction domain to Step S3; Step S3 separates the creep component within the direction domain determined by the slope aspect, obtains the creep rate, and organizes the spatial coherence into a creep coherence intensity field; Step S4 uses the creep coherence intensity field to re-define the boundary of the slippery parts, and then combines the slope and free elevation difference into a normalized weighted classification; Step S5 uses the elevation reconstruction confidence level to gate the coherence statistics of Step S3 and the classification output of Step S4. In the above process, the slope aspect is consumed in two places, in steps S2 and S3; the creep coherence intensity field is generated in step S3 and consumed in step S4; and the elevation reconstruction confidence is generated in step S1 and consumed in steps S3 and S4. Each data object has a clear generation and consumption link, and there is no suspended data, thus ensuring the complete closure of the entire evaluation process.

[0041] In this embodiment, the elevation images from five consecutive aerial photographs are processed according to steps S1 to S5 as described above. After coherent separation in the directional domain in step S3, a coherent creep zone with a slope creep rate of approximately 0.6 mm / d is detected in a steep, high-altitude section with a slope height of approximately 27 m and a slope angle of approximately 46°. The elevation time-series micro-difference amplitude in this zone is approximately 0.5 mm, far lower than the approximately 3 mm elevation noise level in this 3D reconstruction. If only significant deformation exceeding the noise threshold is detected, it will be filtered out as noise. After region growing in step S4, this coherent creep zone extends approximately 2.5 m further along the directional domain reference direction from the static steep, high-altitude section and merges into a slippery area. After normalized weighted scoring, this area is classified as high-risk due to the superposition of slope, elevation difference, and creep rate (3 highs combined) reaching a score of 0.82. Furthermore, the crack density in this area exceeds the threshold, maintaining the risk level at high risk. Two small sections with sparse grass and shrub cover on the slope are marked as questionable and no rating is output because the elevation reconstruction confidence level is below 0.5. In subsequent monitoring phases 6 and 7, visible continuous tension cracks appeared in the aforementioned high-risk areas of the slope, verifying the effectiveness of the present invention in reliably capturing and correctly classifying early signs of slippage during the early creep stage when the slope appearance does not change significantly.

[0042] To further illustrate the effectiveness of this invention, this embodiment is compared with a comparative processing method that filters out noise only within the amplitude range. For the same set of five consecutive elevation images, the comparative processing method subtracts the values ​​of adjacent elevation images grid by grid, retaining only grids where the elevation change amplitude exceeds the reconstruction noise level by approximately 3 mm as deformation zones, and filtering out the rest as noise. In the processing of phases 1 to 5, the comparative processing method did not detect any deformation zones in the aforementioned steep, open sections until phase 7, when the cumulative elevation change in that area increased to approximately 3.4 mm, exceeding the noise level, at which point continuous tension cracks had already appeared on the slope, and the slope had entered an accelerated deformation stage. In contrast, this embodiment detects the coherent creep zone in phase 5 with a creep rate of approximately 0.6 mm / d and an elevation time-series micro-difference amplitude of approximately 0.5 mm, detecting the precursor to slippage approximately two monitoring cycles (60 days) earlier than the comparative processing method, thus gaining a valuable time window for early warning and response. This comparison intuitively demonstrates that shifting the criterion from the amplitude domain to the direction domain determined by the slope is the key to the present invention's ability to capture precursors in the early stages of creep.

[0043] Furthermore, this embodiment also verified the effectiveness of replacing the single-grid slope aspect with the neighborhood main slope aspect. At the aforementioned steep, open-air section's trailing edge micro-topographical fracture, the single-grid slope aspect can fluctuate by more than 20° between adjacent grids. If the single-grid slope aspect is directly used as the reference direction for the directional domain, the reference directions for each grid within the same creep zone will be inconsistent, leading to artificially suppressed spatial coherence and some creep grids being missed due to insufficient spatial coherence. By using the principal eigenvector direction of the elevation gradient covariance matrix of a 7×7 neighborhood window as the neighborhood main slope aspect, the fluctuation of the reference direction in this area is reduced to less than 5°, the overall spatial coherence of the creep zone is improved, and missed detections are significantly reduced. Simultaneously, the output slope micro-topographical fracture degree is significantly higher at this trailing edge, consistent with the actual situation of slope fracture and fissure development on-site, providing useful additional information for interpretation.

[0044] It should be noted that the parameters such as steep slope threshold, coherence threshold, directional compatibility threshold, grading threshold, confidence threshold, and crack density threshold in each step of this invention can be appropriately adjusted within the range of values ​​given in this invention, based on the slope soil and rock properties, aerial photography accuracy, and engineering safety requirements. Different values ​​only affect the rigor of the identification without changing the essence of this invention. For slopes with nearly uniform aspect and relatively smooth micro-topography, the grid-by-grid aspect can be directly used as the reference direction for the directional domain, omitting the calculation of the main aspect of the neighboring domain. For slopes with extremely sparse vegetation cover, the confidence level of elevation reconstruction is generally high, with very few questionable grids, and the gating step has minimal impact on the results. All of these are routine adjustments of this invention under different working conditions and fall within the protection scope of this invention.

[0045] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for image-based classification and assessment of slope slippage risk in loess gully areas, characterized in that, include: Acquire aerial images of the target slope taken by UAVs in multiple phases, reconstruct the three-dimensional topography of the slope and rasterize it into elevation images of each phase; The slope and aspect are calculated grid by grid in the current elevation image. Based on the grid where the slope exceeds the steep slope threshold and the aspect points to the free face, the high and steep free face area is delineated as the slippery part. The free face is the free boundary on the side of the slope toe. After registering multiple elevation images to the same reference, the elevation temporal difference is calculated grid by grid. Within the direction domain determined by the slope aspect, the creep component is separated from the elevation temporal difference based on the spatial coherence of the elevation temporal difference along the direction domain to obtain the slope creep rate. For the slippery parts, their slope, the difference in height relative to the free surface, and the creep rate are normalized and weighted to form a slip risk level, and the graded assessment result is output.

2. The image-based classification and assessment method for slope slippage risk in loess gully areas according to claim 1, characterized in that, The step of separating the creep component from the elevation time series difference within the direction domain determined by the slope aspect based on the spatial coherence of the elevation time series difference along the direction domain includes: Using the slope direction of each grid as the reference direction of the grid's direction domain, the elevation time-series differential of the grid and its neighboring grids is projected along the reference direction of the direction domain, and the sign consistency and amplitude continuity of the projected components are statistically analyzed to obtain the spatial coherence of the grid. The elevation time-series micro-differences with spatial coherence not less than the coherence threshold are identified as creep components, and the elevation time-series micro-differences with spatial coherence less than the coherence threshold are identified as noise and eliminated.

3. The image-based classification and assessment method for slope slippage risk in the loess gully region according to claim 2, characterized in that, Before calculating the spatial coherence, a neighborhood window is taken for each grid cell in the current elevation image. The covariance matrix of the elevation gradient within the neighborhood window is calculated. The direction of the eigenvector corresponding to the largest eigenvalue of the covariance matrix is ​​taken as the main slope direction of the neighborhood of the grid cell. The main slope direction of the neighborhood is used to replace the slope direction of the grid cell as the reference direction of the direction domain. The directional dispersion of the elevation gradient within the neighborhood window is also output as the slope micro-topographic fragmentation.

4. The image-based classification and assessment method for slope slippage risk in the loess gully region according to claim 3, characterized in that, The spatial coherence of each grid is organized into a creep coherence intensity field according to the grid position. Taking the high and steep free zone as the initial seed region, a region growth is performed on the grids in the creep coherence intensity field whose spatial coherence is not less than the coherence threshold, which are connected to the initial seed region and whose directional domain reference direction is compatible, to obtain the creep expansion region. The creep expansion region is then incorporated into the slippery part to update the boundary of the slippery part.

5. The image-based classification and assessment method for slope slippage risk in the loess gully region according to claim 4, characterized in that, The weighted score, which is normalized and calculated based on the slope, the height difference relative to the free surface, and the creep rate, includes: The slope, the height difference, and the creep rate of each grid in the slippery area are normalized according to their respective empirical distributions in the slippery area to obtain normalized values ​​for slope, height difference, and creep rate. An adaptive weight is determined based on the correlation strength between the slope normalization value, the free height difference normalization value, and the creep rate normalization value and slip instability. The grid slip risk score is obtained by weighted summation of the slope normalization value, the free height difference normalization value, and the creep rate normalization value, and the grid slip risk score is mapped to the slip risk level.

6. The image-based classification and assessment method for slope slippage risk in loess gully areas according to claim 5, characterized in that, When reconstructing the three-dimensional terrain, the elevation reconstruction confidence of each grid is obtained based on the ground point density of each grid in the point cloud obtained from the image reconstruction and the vegetation cover index of that grid. Grids with elevation reconstruction confidence less than the confidence threshold are marked as questionable grids. The questionable grids are not included in the spatial coherence statistics, and the questionable grids are marked as questionable in the classification assessment results without outputting the slip risk level.

7. The image-based classification and assessment method for slope slippage risk in loess gully areas according to claim 1, characterized in that, When acquiring aerial images of the target slope taken by the UAV in multiple phases, the forward overlap rate of the aerial photography shall not be less than 75%, and the lateral overlap rate shall not be less than 60%; when rasterizing into elevation images of each phase, the elevation images of each phase shall use the same raster resolution, which is 0.02m to 0.10m.

8. The image-based classification and assessment method for slope slippage risk in loess gully areas according to claim 1, characterized in that, The steep slope threshold is between 30° and 45°; the slope direction is towards the free surface, meaning the angle between the slope direction of the grid and the direction of the grid pointing towards the free surface is less than 90°.

9. The image-based classification and assessment method for slope slippage risk in loess gully areas according to claim 1, characterized in that, The process of registering multiple elevation images to the same reference includes: selecting an exposed bedrock area or road surface area outside the target slope as a stable feature area; obtaining registration parameters with the consistency of the elevation of the stable feature area in each period of elevation images as a constraint; and registering the elevation images of each period to the same reference using the registration parameters.

10. The image-based classification and assessment method for slope slippage risk in loess gully areas according to claim 1, characterized in that, Gray-scale double-threshold edge extraction is performed on the orthophoto corresponding to the current elevation image to obtain slope cracks and the crack density in the slippery area is counted; when the crack density exceeds the crack density threshold, the slip risk level of the slippery area is increased by one level.

Images