Slope verification method and related equipment based on UAV vision recognition

Abstract

This invention discloses a slope verification method and related equipment based on UAV visual recognition. This invention relates to the field of image processing technology. First, it collects data from multiple sources using a UAV and fuses it to generate a multi-scale fused image, serving as a unified analysis basis. Second, it extracts regional features from the image, performs intelligent partitioning, and identifies key element features, mapping them to each partition to establish a region-element association model. Next, it introduces a time-series dimension, comparing the current partition data with historical slope structure data, and innovatively employs an element reduction analysis method to dynamically correct mechanical parameters based on the development and evolution of elements, thereby calculating the slope change and structural deformation reflecting the true damage state. Finally, it generates a slope heat map based on these quantitative indicators, achieving visualized risk output.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a slope verification method and related equipment based on UAV visual recognition. Background Technology

[0002] In the field of landform observation and analysis, especially in environments with significant human intervention and complex dynamic changes, continuous and accurate verification of the three-dimensional landform is a core element in ensuring operational safety and optimizing planning and design. Topographic stability assessment, deformation monitoring and early warning, and engineering remediation design all heavily rely on the precise depiction of the regional landform's three-dimensional morphology. Currently, with the widespread adoption of remote sensing technologies such as UAV aerial surveying and aerial photogrammetry, automatically generating digital elevation models and topographic contour lines based on visible light or remote sensing imagery through techniques such as stereo matching and 3D reconstruction has become the mainstream technical approach for rapidly acquiring large-scale landform data. Compared to traditional single-point measurement methods, this method has significant advantages in efficiency, cost, and coverage, and theoretically can establish an efficient and dynamic topographic update and deformation monitoring system for relevant areas.

[0003] However, the Earth's surface, shaped by intense human activity, is a continuously and rapidly changing unnatural geomorphological system. Its topographic features differ fundamentally from natural landforms, exhibiting a highly discontinuous and heterogeneous complex spatial structure. For example, multi-level platforms, steep slopes, subsidence and deformation of loose deposits, and abrupt, rugged edges formed by mechanical operations or external forces collectively constitute a unique scene filled with sharp geometric features and low-texture areas. Simultaneously, these areas are highly dynamic, with continuously changing work surfaces, interference from moving objects, changes in surface cover, localized water accumulation, and variable lighting and weather conditions, all posing significant challenges to the quality of image data acquisition. These unique working conditions and environmental factors make these areas typical application scenarios requiring extremely high robustness and accuracy in terrain modeling techniques.

[0004] Existing image-based automatic terrain reconstruction technologies typically optimize their algorithms for the continuous, smooth transitions of natural landforms. However, when directly applied to scenarios with strong human intervention, a significant mismatch arises between the technological principles and objective reality. In key technical aspects, stereo matching algorithms are prone to matching failures or errors in low-texture or high-contrast areas such as step edges and shadow zones. This results in voids, noise, or smoothing distortions in the generated point cloud data at critical terrain abrupt changes. Consequently, the resulting digital elevation models and contour lines exhibit systematic biases in representing surface geometric parameters (such as platform width, slope angle, and step height). The results often manifest as: unreasonably dense or sparse contour lines at steep slopes, missing or artificially closed contour lines in shadowed and waterlogged areas, and "pseudo-change" noise in the terrain under dynamic disturbances. These distortions are not random errors but rather systematic biases with spatial regularity strongly correlated with regional terrain structure and dynamic activities.

[0005] In summary, the core flaw of existing technologies lies in the inherent limitations of the automated image-based terrain data generation process under special and harsh environments and dynamic disturbances. This leads to systematic and structural distortions in the resulting basic spatial data. This distortion directly impacts the terrain verification and analysis processes that rely solely or primarily on this data. Consequently, all subsequent geometric parameter extraction, stability calculations, deformation comparisons, and monitoring and early warning systems are built upon unreliable "input errors," severely affecting the accuracy and reliability of security risk assessments in relevant areas. Therefore, overcoming the application bottlenecks of existing automated terrain modeling technologies in such complex scenarios and obtaining reliable terrain data that accurately reflects complex surface features under human disturbance has become a critical technical issue that urgently needs to be addressed to ensure the effectiveness of relevant digital management and control systems. Summary of the Invention

[0006] Based on the problems mentioned above, the purpose of this invention is to provide a slope verification method and related equipment based on UAV visual recognition, which solves the problem that the existing technology is difficult to adapt to dynamic scenes with drastic human changes and complex geometric features, resulting in systematic distortion of terrain data automatically generated from images, and thus causing large errors in slope verification.

[0007] This invention is achieved through the following technical solution:

[0008] The first aspect of this invention provides a slope verification method based on UAV visual recognition, comprising the following steps:

[0009] Step S1: Collect target slope data using a drone equipped with multi-source acquisition devices to obtain multi-source raw data for slope verification; fuse the multi-source raw data to generate a multi-scale fused image;

[0010] Step S2: Extract the regional features related to slope stability verification from the multi-scale fused image, and divide the target slope into several slope structure regions suitable for slope verification and judgment based on the regional features.

[0011] Step S3: Identify the feature elements that affect slope stability in the multi-scale fused image, and map each feature element to the corresponding slope structure region according to the spatial position relationship to obtain several slope structure regions containing feature elements.

[0012] Step S4: Obtain historical slope structure data from the previous period. Based on regional division, compare the structure of several slope structure regions containing feature characteristics with historical slope structure data. Based on the comparison results, use the feature reduction analysis method to calculate the slope change and structural deformation.

[0013] Step S5: Based on the slope change and structural deformation of each region, generate a target slope heat map for slope verification, and complete the data visualization output of slope verification.

[0014] In the above technical solution, firstly, data is collected by multi-source devices using UAVs and fused to generate multi-scale fused images, which serve as a unified analysis basis. Secondly, regional features are extracted from these images for intelligent partitioning, and key element features are identified and mapped to each partition to establish a region-element correlation model. Next, a time-series dimension is introduced, comparing the current partition data with historical slope structure data. An innovative element reduction analysis method is used to dynamically correct mechanical parameters based on the development and evolution of elements, thereby calculating the slope change and structural deformation that reflect the true damage state. Finally, a slope heat map is generated based on these quantitative indicators to achieve visualized risk output.

[0015] In one optional embodiment, extracting regional features related to slope stability verification from the multi-scale fused image includes the following steps:

[0016] The multi-scale fused image is input into the feature extraction model, and the feature extraction model is used to perform multi-scale parallel extraction to generate multiple sets of feature maps containing detailed and global information.

[0017] The importance weights of features at different scales in the multiple sets of feature maps for identifying boundary characteristics and structural surface characteristics at each spatial location are determined, and engineering semantic fusion is performed on the multiple sets of feature maps based on the importance weights to obtain a multi-scale engineering semantic feature map.

[0018] A geological prior attention map is constructed, and the geological prior attention map is used as a gating signal to perform element-wise multiplication with the multi-scale engineering semantic feature map to obtain a primary slope feature map containing regional features.

[0019] In one optional embodiment, the feature extraction model includes a high-resolution detail module; wherein, multi-scale parallel extraction using the high-resolution detail module includes the following steps:

[0020] Guided fusion of multi-scale fused images is performed to generate a multimodal guided fusion input tensor;

[0021] The multi-directional Sobel gradient operator is applied to the multimodal guided fusion input tensor for convolution, and the intensity and directional gradient maps in multiple directions are extracted to generate multi-directional gradient feature maps.

[0022] The multimodal guided fusion input tensor and the multidirectional gradient feature map are concatenated in the channel dimension to form a primary enhanced feature map;

[0023] After smoothing and activating the primary enhanced feature map using a 5*5 depth separable convolutional layer, a deformable convolutional layer is then connected to extract non-rigid feature contours, generating a geometrically adaptive feature map.

[0024] The geometric adaptive feature map is sequentially passed through two 3*3 depth separable convolutional layers to extract texture / grain features and local pattern features. The extracted texture / grain features and local pattern features are then horizontally connected with the geometric adaptive feature map to generate a multi-scale detail feature map.

[0025] Calculate the spatial attention map and channel attention vector of the multi-scale detail feature map, and use the spatial attention map and channel attention vector to reweight the multi-scale detail feature map to generate an attention-reweighted feature map;

[0026] The attention-reweighted feature map is subjected to engineering semantic interaction by small dilated convolutional blocks to generate a local engineering context feature map.

[0027] In one optional embodiment, the target slope is divided into several slope structure regions for slope verification and determination based on the region characteristics, including the following steps:

[0028] Align the primary slope feature map with the three-dimensional point cloud data of the target slope, and quantify the aligned primary slope feature map into an engineering geological parameter map.

[0029] Based on the quantified engineering geological parameter map, a clustering algorithm is used to group the slope pixels, and areas with similar lithological combinations and structural surface development characteristics are merged into the same mean unit, outputting a primary slope mechanical property zoning map.

[0030] The morphological verification of the primary slope mechanical property zoning map is performed to generate a slope structure region division map.

[0031] In one optional embodiment, identifying key feature elements affecting slope stability in the multi-scale fused image, and mapping each feature element to the corresponding slope structure region according to its spatial location, includes:

[0032] Based on a deep learning-based linear feature segmentation model and Hough transform, linear features are identified from high-resolution multi-band orthophotos in the multi-scale fused image. Based on the DEM model, the linear features are verified by slope aspect consistency analysis and surface roughness calculation, and a vectorized structural surface line segment map is generated.

[0033] By combining the lithological and geomorphological spectral library, spectral angle mapping is used to perform lithological analysis on the high-resolution multi-band orthophoto image, generating a lithological classification map and a spatial probability distribution map of weak interlayers.

[0034] The high-resolution multi-band orthophoto image is used to identify micro-topographic deformation by combining terrain parameter threshold segmentation technology with image texture, and a micro-topographic deformation marker distribution map is generated.

[0035] The vectorized structural surface line segment map, the lithology classification map, the spatial probability distribution map of the weak interlayer, and the distribution map of the micro-geomorphic deformation markers are subjected to spatiotemporal correlation analysis, and the results of the spatiotemporal correlation analysis are mapped to the corresponding slope structural regions to generate slope element feature maps.

[0036] In one optional embodiment, historical slope structure data from the previous period is obtained, and a structural comparison is performed between several slope structure regions containing feature characteristics and the historical slope structure data based on regional division, including the following steps:

[0037] Spatially align several slope structure regions containing feature characteristics with historical slope structure data to generate a spatially aligned image;

[0038] Calculate the morphological changes and feature evolution variables of the spatially aligned image in each region.

[0039] In one optional embodiment, the slope change and structural deformation are calculated using the element reduction analysis method based on the comparison results, including the following steps:

[0040] Based on rock mechanics theory, a vector of basic strength reduction coefficients for each element is constructed, and an element type-basic reduction coefficient lookup table is generated.

[0041] The spatial overlay analysis method is used to quantify the characteristics of each element within each slope structure area, and generate a slope structure area-element characteristic association attribute table.

[0042] By combining the comprehensive element type-basic reduction coefficient lookup table and the slope structure area-element feature association attribute table, parameter reduction corrections are made for the morphological change and the element feature evolution variables.

[0043] Based on the morphological changes after parameter reduction and the evolution variables of the aforementioned element characteristics, the slope change and structural deformation are calculated.

[0044] A second aspect of the present invention provides a slope verification system based on UAV visual recognition, comprising:

[0045] The data preprocessing module is used to collect target slope data through a drone equipped with a multi-source acquisition device, obtain multi-source raw data for slope verification, and fuse the multi-source raw data to generate a multi-scale fused image.

[0046] The region division module is used to extract regional features related to slope stability verification from the multi-scale fused image, and divide the target slope into several slope structure regions that are suitable for slope verification and judgment based on these regional features.

[0047] The feature recognition module is used to identify the feature elements that affect slope stability in the multi-scale fused image, and to map each feature element to the corresponding slope structure region according to the spatial position relationship, thereby obtaining several slope structure regions containing feature elements.

[0048] The variable calculation module is used to obtain historical slope structure data from the previous period. Based on the regional division, it performs structural comparison between several slope structure regions containing feature characteristics and historical slope structure data, and calculates the slope change and structural deformation based on the comparison results using the feature reduction analysis method.

[0049] The alarm module is used to generate a target slope heat map for slope verification based on the slope change and structural deformation of each area, and to complete the data visualization output of slope verification.

[0050] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a slope verification method based on UAV visual recognition.

[0051] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a slope verification method based on UAV visual recognition.

[0052] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0053] This invention achieves a transformation from static homogeneous assessment to dynamic zonal damage quantification and early warning by dynamically embedding the spatiotemporal evolution of geological elements into a mechanical parameter reduction model, significantly improving the accuracy and predictability of slope condition analysis and risk warning. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0055] Figure 1 This is a flowchart illustrating the slope verification method based on UAV visual recognition provided in Embodiment 1 of the present invention.

[0056] Figure 2 This is a schematic diagram of the slope verification system based on UAV visual recognition provided in Embodiment 2 of the present invention;

[0057] Figure 3 This is a schematic diagram of the structure of an electronic device provided in Embodiment 2 of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0059] Embodiment 1 of this invention provides a slope verification method based on UAV visual recognition. Figure 1 This is a flowchart illustrating the slope verification method based on UAV visual recognition provided in Embodiment 1 of the present invention, as shown below. Figure 1 As shown, the slope verification method based on UAV visual recognition includes the following steps:

[0060] Step S1: Collect target slope data using a drone equipped with a multi-source acquisition device to obtain multi-source raw data for slope verification; fuse the multi-source raw data to generate a multi-scale fused image;

[0061] Step S2: Extract the regional features related to slope stability verification from the multi-scale fused image, and divide the target slope into several slope structure regions suitable for slope verification and judgment based on the regional features.

[0062] Step S3: Identify the feature elements that affect slope stability in the multi-scale fused image, and map each feature element to the corresponding slope structure region according to the spatial position relationship to obtain several slope structure regions containing feature elements.

[0063] Step S4: Obtain historical slope structure data from the previous period. Based on regional division, compare the structure of several slope structure regions containing feature characteristics with historical slope structure data. Based on the comparison results, use the feature reduction analysis method to calculate the slope change and structural deformation.

[0064] Step S5: Based on the slope change and structural deformation of each region, generate a target slope heat map for slope verification, and complete the data visualization output of slope verification.

[0065] It should be noted that, firstly, data is collected from multiple sources using drones and fused to generate multi-scale fused images, which serve as a unified analysis basis; secondly, regional features are extracted from these images for intelligent partitioning, and key element features are identified and mapped to each partition to establish a region-element correlation model; nextly, a time-series dimension is introduced, comparing the current partition data with historical slope structure data, and innovatively employing an element reduction analysis method to dynamically correct mechanical parameters based on the development and evolution of elements, thereby calculating the slope change and structural deformation that reflect the true damage state; finally, a slope heat map is generated based on these quantitative indicators to achieve visualized risk output.

[0066] The multi-source raw data includes: multi-view oblique images, multispectral images, and 3D point cloud data; specifically, the fusion of the multi-source raw data includes:

[0067] Acquire multi-view oblique images of the target area, ensuring the images cover the entire target area with an overlap of at least 70% and a resolution of at least 0.1m. Based on the multi-view oblique images, use the Structure for Motion Recovery (SFM) algorithm to extract and match feature points, constructing the relative pose relationships between images. Then, use a dense matching algorithm to generate a high-precision optical 3D point cloud of the target area, with a point density of at least 100 points / m².

[0068] Based on the generated optical 3D point cloud, a digital surface model (DSM) is generated through a triangular irregular network (TIN) construction algorithm. At the same time, texture mapping technology is used to map the texture information of the multi-view tilted image onto the surface of the digital surface model to generate a true-color orthophoto. The resolution of the true-color orthophoto is consistent with that of the input multi-view tilted image.

[0069] The lidar point cloud data of the target area is collected, and the lidar point cloud and the optical 3D point cloud are accurately registered using an improved algorithm based on iterative nearest point (ICP).

[0070] After registration is completed, data optimization is performed: the LiDAR point cloud is selected as the geometric skeleton for the subsequent construction of the Digital Surface Model (DSM) and Digital Elevation Model (DEM) to ensure elevation accuracy; the optical 3D point cloud and its texture information are selected as the surface spectral information source for the subsequent image spectral information supplementation.

[0071] Multispectral images of the target area are acquired, containing at least RGB and near-infrared bands, with a resolution of no less than 0.2m. A pixel-level fusion algorithm is used to fuse the multispectral images with true-color orthophotos to generate a high-resolution multi-band orthophoto. This high-resolution multi-band orthophoto combines the high resolution of true-color orthophotos with the rich spectral information of multispectral images.

[0072] Based on the registered point cloud data, a rasterization method is used to generate digital elevation models (DEM) and digital surface models (DSM) of different resolutions.

[0073] High-resolution multi-band orthophotos were aligned with DEMs / DSMs of different resolutions to ensure that all data strictly adhered to the WGS84 coordinate system. The aligned multi-band orthophotos were then integrated with DEMs / DSMs of various resolutions to form a multi-scale fused image.

[0074] It should be noted that the algorithms and models mentioned above are all existing technologies, and the use of algorithms and models for preprocessing multi-source raw data is also an existing technology. This method does not improve upon them.

[0075] In one optional embodiment, extracting regional features related to slope stability verification from the multi-scale fused image includes the following steps:

[0076] The multi-scale fused image is input into the feature extraction model, and the feature extraction model is used to perform multi-scale parallel extraction to generate multiple sets of feature maps containing detailed and global information.

[0077] The importance weights of features at different scales in the multiple sets of feature maps for identifying boundary characteristics and structural surface characteristics at each spatial location are determined, and engineering semantic fusion is performed on the multiple sets of feature maps based on the importance weights to obtain a multi-scale engineering semantic feature map.

[0078] A geological prior attention map is constructed, and the geological prior attention map is used as a gating signal to perform element-wise multiplication with the multi-scale engineering semantic feature map to obtain a primary slope feature map containing regional features.

[0079] The feature extraction model is a neural network with a multi-branch encoder. This neural network processes the boundary details of the target slope and the global features reflecting the regional background in parallel through the hollow spatial pyramid pooling module and the high-resolution detail module, respectively. This generates multiple sets of feature maps containing detailed and global information, providing an accurate basis for subsequent partitioning.

[0080] Specifically, neural networks with multi-branch encoders include:

[0081] The input layer receives multi-scale fused images, normalizes them, and inputs them in parallel into the subsequent hollow spatial pyramid pooling module and high-resolution detail module.

[0082] The void space pyramid pooling module is used to extract global contextual engineering semantic features covering different ranges, such as rock layer distribution and large-scale structural outlines. This module includes four parallel convolutional branches, each of which uses 3*3 void convolution with different void ratios and a global average pooling branch, thereby expanding the receptive field at multiple scales. In this embodiment, the void ratios of the four convolutional branches are 1, 6, 12, and 18, respectively.

[0083] Furthermore, in this embodiment, the regional features related to slope stability verification include: platform region features, slope surface region features, slope crest line region features, and slope bottom line region features.

[0084] The platform area, serving as a buffer and load-bearing structure for the slope, directly affects the local stability and drainage capacity of the slope due to its width, shape, and compaction. Currently, when using traditional neural networks for identification, these features are easily overlooked due to insufficient resolution or reliance solely on elevation. Therefore, in this embodiment, the construction of high-resolution detail modules considers micro-topographic transition lines and abrupt changes in surface texture. While maintaining high spatial resolution, these subtle surface textures and spectral gradients are enhanced layer by layer, thereby accurately locating the ambiguous but engineering-significant transition boundary between the platform and the slope.

[0085] The slope gradient, slope shape, soil and rock properties, and protection status of a slope area are key to assessing overall stability and potential landslide risk. Currently, these details are difficult to extract coherently in conventional images due to their small scale, low contrast, and frequent obscuring by shadows or vegetation patches. Therefore, in this embodiment, when constructing the high-resolution detail module, the fine patterns of the original joint network on the rock surface, shallow erosion grooves, and local spalling areas are considered to identify the slope area features of slope integrity, weathering degree, and potential damage modes.

[0086] The crest region often bears tensile stress (the reaction force of an object to an external force that tends to stretch the object). Crack development, settlement, or dislocation in this area are precursor signals of slope instability. Damage begins with millimeter-centimeter-level microcracks and particle-level transport (particle-level transport refers to the movement of particles in physical space, usually including several basic types such as diffusion, settlement, convection, and weathering). These appear in images as extremely fine dark linear patterns or subtle increases in local texture roughness. These features are highly blurred at the pixel scale and are easily confused with shadows and weeds. Therefore, in this embodiment, when constructing high-resolution detail modules, the local contrast between subpixel linear features and the background is considered, and the complex patterns of real microcracks and imaging noise are distinguished.

[0087] As a region of significant stress concentration and shear stress, the deformation, bulging, or seepage of the slope baseline directly reflects the stability of the slope foundation. Deformation of the slope baseline often manifests as localized disorder in the orientation of surface soil and rock (change in texture direction) or irregular color infiltration contours caused by water seepage. These features have irregular boundaries, low contrast, and are often covered by colluvial deposits. Therefore, when constructing high-resolution detail modules in this embodiment, it is necessary to consider how to identify the subtle edges of surface material and moisture distribution caused by internal deformation in the absence of significant elevation changes.

[0088] Based on the aforementioned engineering identification requirements, this embodiment constructs a high-resolution detail module. Systematic examination and comprehensive analysis of these regional features can accurately identify weak points, potential failure modes, and evolution trends of slopes.

[0089] In one optional embodiment, the feature extraction model includes a high-resolution detail module; wherein, multi-scale parallel extraction using the high-resolution detail module includes the following steps:

[0090] Guided fusion of multi-scale fused images is performed to generate a multimodal guided fusion input tensor;

[0091] The multi-directional Sobel gradient operator is applied to the multimodal guided fusion input tensor for convolution, and the intensity and directional gradient maps in multiple directions are extracted to generate multi-directional gradient feature maps.

[0092] The multimodal guided fusion input tensor and the multidirectional gradient feature map are concatenated in the channel dimension to form a primary enhanced feature map;

[0093] After smoothing and activating the primary enhanced feature map using a 5*5 depth separable convolutional layer, a deformable convolutional layer is then connected to extract non-rigid feature contours, generating a geometrically adaptive feature map.

[0094] The geometric adaptive feature map is sequentially passed through two 3*3 depth separable convolutional layers to extract texture / grain features and local pattern features. The extracted texture / grain features and local pattern features are then horizontally connected with the geometric adaptive feature map to generate a multi-scale detail feature map.

[0095] Calculate the spatial attention map and channel attention vector of the multi-scale detail feature map, and use the spatial attention map and channel attention vector to reweight the multi-scale detail feature map to generate an attention-reweighted feature map;

[0096] The attention-reweighted feature map is subjected to engineering semantic interaction by small dilated convolutional blocks to generate a local engineering context feature map.

[0097] It should be noted that the high-resolution detail module, in structural order, specifically includes:

[0098] The preprocessing layer guides the fusion of the multi-scale fused images processed by the input layer. Specifically, it calculates in real time the slope map reflecting the steepness of the terrain and the profile curvature map reflecting the undulation of the terrain from the highest precision DEM. The original RGB bands, slope map, profile curvature map and NDVI map are stitched together in the channel dimension to generate a multimodal guided fusion input tensor. This tensor deeply fuses the high-dimensional input of the original spectrum, derived terrain features and phenological features, directly guiding the network to focus on elements related to slope stability. The professional knowledge of the slope engineering field is processed in the input stage of the neural network training (the preprocessing layer here) in the form of computable feature maps, realizing the forward embedding of domain knowledge and reducing the learning cost of the neural network.

[0099] A set of multi-directional Sobel gradient operators are used to convolve the multimodal guided fusion input tensor to extract multi-directional intensity and directional gradient maps, resulting in a multi-directional gradient feature map. The multi-directional gradient feature map is then concatenated with the original multimodal guided fusion input tensor in the channel dimension to form a primary enhancement feature map, which strengthens directional structural features such as cracks at the top of the slope and platform edges, thus achieving direction-aware primary edge enhancement.

[0100] The initial enhanced feature map undergoes initial smoothing and feature activation through a first-layer 5x5 depthwise separable convolutional layer, followed by a deformable convolutional layer. This deformable convolutional layer learns a two-dimensional offset field, enabling the convolutional kernel sampling points to adaptively align with the contours of non-rigid features such as irregular micro-cracks and wetted feather boundaries. The deformable convolutional layer drives adaptive detail capture, where the geometry of micro-features is encoded more accurately, providing accurate data for the identification of features in the slope crest, slope bottom, and slope surface regions.

[0101] A multi-scale detail feature pyramid is constructed, and geometric adaptive feature extraction is performed on the geometric adaptive feature map. Specifically, it is passed through two 3*3 depthwise separable convolutional layers in sequence, with group normalization and SiLU activation following each separable convolutional layer. The first depthwise separable convolutional layer is used to capture texture / grain features, and the second depthwise separable convolutional layer is used to capture local pattern features. The captured texture / grain features, local pattern features and geometric adaptive feature map are aligned and horizontally connected through a 1*1 convolutional layer to generate a multi-scale detail feature map, which simultaneously contains edge, texture and local shape information.

[0102] A collaborative attention module is constructed to calculate the spatial attention map of the multi-scale detail feature map to identify where it is important (e.g., the spatial zone of the top and bottom of the slope). The channel attention vector of the multi-scale detail feature map is calculated to identify which features are important (e.g., the feature channel that is sensitive to humidity). The two are combined to reweight the multi-scale detail feature map to generate an attention-reweighted feature map, in which the responses of key regions and key feature channels are significantly enhanced, while irrelevant background information is suppressed.

[0103] By using small dilated convolutional blocks (dilation rate = 2), spatially related points on the attention-reweighted feature map (such as pixels belonging to the same discontinuous crack) are made to interact with engineering semantics, and local context is aggregated; a local engineering context feature map with local context is generated, and fragmented weak features (such as discontinuous cracks) are initially associated with engineering semantics through the local engineering context feature map.

[0104] Finally, channel adjustment and fusion are performed through the output layer (1*1 convolutional layer) to output a high-resolution detail feature map with the same spatial resolution as the original input.

[0105] A 1*1 convolutional layer is used to perform feature dimensionality reduction and regularization on the engineering semantic context feature map output by the hole spatial pyramid pooling module and the high-resolution detail feature map output by the high-resolution detail module, so as to unify and reduce the feature channel dimension. At the same time, nonlinear fusion preparation is performed to reduce the amount of computation and improve the efficiency of subsequent fusion.

[0106] The cross-branch feature alignment and fusion layer upsamples the high-resolution detail feature map to the same spatial size as the engineering semantic context feature map through bilinear interpolation. Then, the two are added element by element. Finally, a 3*3 convolutional layer is used to smooth and refine the fused features, generating a fused feature map that has both fine local details and rich global semantics.

[0107] Furthermore, the feature extraction model also includes a multi-scale output layer, which takes intermediate features from the high-resolution detail module and intermediate features with different void ratios from the void space pyramid pooling module as bypass outputs, and together forms multiple sets of feature maps containing detailed and global information for weight fusion in the 3*3 convolutional layer of the cross-branch feature alignment and fusion layer.

[0108] In one optional embodiment, the target slope is divided into several slope structure regions for slope verification and determination based on the region characteristics, including the following steps:

[0109] Align the primary slope feature map with the three-dimensional point cloud data of the target slope, and quantify the aligned primary slope feature map into an engineering geological parameter map.

[0110] Based on the quantified engineering geological parameter map, a clustering algorithm is used to group the slope pixels, and areas with similar lithological combinations and structural surface development characteristics are merged into the same mean unit, outputting a primary slope mechanical property zoning map.

[0111] The morphological verification of the primary slope mechanical property zoning map is performed to generate a slope structure region division map.

[0112] It should be noted that the slope structure zoning map not only satisfies the morphological characteristics of the platform area, slope surface area, slope crest line area, and slope bottom line area, but also satisfies the lithological combination and structural surface development characteristics. After such division, the slope structure zoning map provides accurate morphological and developmental characteristics for subsequent calculation of slope change and structural deformation using the element reduction analysis method, so that the analysis of the expected change process of each element in each area can be satisfied when performing element reduction.

[0113] In one optional embodiment, identifying key feature elements affecting slope stability in the multi-scale fused image, and mapping each feature element to the corresponding slope structure region according to its spatial location, includes:

[0114] Based on a deep learning-based linear feature segmentation model and Hough transform, linear features are identified from high-resolution multi-band orthophotos in the multi-scale fused image. Based on the DEM model, the linear features are verified by slope aspect consistency analysis and surface roughness calculation, and a vectorized structural surface line segment map is generated.

[0115] By combining the lithological and geomorphological spectral library, spectral angle mapping is used to perform lithological analysis on the high-resolution multi-band orthophoto image, generating a lithological classification map and a spatial probability distribution map of weak interlayers.

[0116] The high-resolution multi-band orthophoto image is used to identify micro-topographic deformation by combining terrain parameter threshold segmentation technology with image texture, and a micro-topographic deformation marker distribution map is generated.

[0117] The vectorized structural surface line segment map, the lithology classification map, the spatial probability distribution map of the weak interlayer, and the distribution map of the micro-geomorphic deformation markers are subjected to spatiotemporal correlation analysis, and the results of the spatiotemporal correlation analysis are mapped to the corresponding slope structural regions to generate slope element feature maps.

[0118] It should be noted that high-resolution multi-band orthophotos and DEM models were used to extract rock mass structural surface information through deep learning-based linear feature segmentation models and geometric analysis techniques. An improved UNet linear feature segmentation model was employed to initially identify linear features from the images, followed by vectorization using post-processing techniques such as Hough transform. Simultaneously, slope consistency and surface roughness were calculated based on the DEM to verify the linear features and to determine their geometric attitude (the attitude of planar bodies and lines, primarily referring to their spatial extension and orientation, including three elements: strike, dip, and dip angle). Finally, a vectorized structural surface line segment map and its attribute table were generated, containing key parameters such as the spatial location, length, dip, dip angle, and confidence level of each line segment.

[0119] Using multispectral / hyperspectral and thermal infrared imagery data, lithological boundaries and weak interlayers are identified through spectral analysis and classification techniques. Preliminary lithological classification is performed using methods such as spectral angle mapping or support vector machines, combined with a known lithological spectral library. Specifically for weak interlayers, thermal inertia analysis of thermal infrared data or the absorption spectral characteristics of specific minerals are used to identify strata with high water content and easy weathering. Finally, a lithological classification map and a spatial probability distribution map of weak interlayers are generated, clearly marking the location and extent of different lithological units and weak zones.

[0120] Micro-topographic deformation identification includes: segmenting slope maps and planar curvature maps using topographic parameter thresholding technology (the top 5% and bottom 5% quantiles of the overall statistical topographic parameter distribution), extracting "extremely steep areas" and "extremely high convexity / concavity areas" respectively, and then performing a logical OR operation to generate a binary image of topographic anomalies, initially delineating all pixel areas that may contain steep slopes, cracks, bulges, or depressions; within the regions of the binary image, using the gray-level co-occurrence matrix to calculate the contrast, homogeneity, and entropy of each pixel's neighborhood, generating a multi-channel texture feature map, where high contrast, low homogeneity, and high entropy areas indicate disordered arrangement of surface materials (such as collapsed accumulations and crack zones), while smooth areas may be intact rock surfaces; constructing a decision tree or simple rule model: classifying "high slope and..." Regions with "high texture contrast" are classified as "potential steep slopes or cracks"; regions with "high planar curvature (positive) and medium texture entropy" are classified as "potential bulges"; and regions with "high planar curvature (negative) and low texture homogeneity" are classified as "potential subsidence or depressions." This generates a preliminary micro-topographical marker map, where each labeled cell is assigned a preliminary type label and confidence score. Morphological closing operations are applied to connect neighboring fragmented cells of the same type. Then, connected component analysis is used to transform raster patches into independent objects. Finally, based on the active contour model, the boundaries of each object are smoothed and optimized to generate a vectorized micro-topographical object layer. Each polygonal object has geometric and attribute information such as type, area, perimeter, average slope, average curvature, and average texture entropy.

[0121] In one optional embodiment, historical slope structure data from the previous period is obtained, and a structural comparison is performed between several slope structure regions containing feature characteristics and the historical slope structure data based on regional division, including the following steps:

[0122] Spatially align several slope structure regions containing feature characteristics with historical slope structure data to generate a spatially aligned image;

[0123] Calculate the morphological changes and feature evolution variables of the spatially aligned image in each region.

[0124] It should be noted that the morphological changes include the changes in area of ​​each region; the evolution of element characteristics includes the changes in steep slopes (steep terrain with a slope of more than 70 degrees), cracks, bulges or depressions.

[0125] In one optional embodiment, the slope change and structural deformation are calculated using the element reduction analysis method based on the comparison results, including the following steps:

[0126] Based on rock mechanics theory, a vector of basic strength reduction coefficients for each element is constructed, and an element type-basic reduction coefficient lookup table is generated.

[0127] The spatial overlay analysis method is used to quantify the characteristics of each element within each slope structure area, and generate a slope structure area-element characteristic association attribute table.

[0128] By combining the comprehensive element type-basic reduction coefficient lookup table and the slope structure area-element feature association attribute table, parameter reduction corrections are made for the morphological change and the element feature evolution variables.

[0129] Based on the morphological changes and element characteristic evolution variables after parameter reduction and correction, calculate the slope change and structural deformation.

[0130] It should be noted that, by sorting out rock mechanics theory and localized test data, a quantitative influence relationship between unit development intensity and rock and soil strength is established for each type of key geological element, namely the foundation strength reduction coefficient vector; the qualitative geological description is transformed into calculable mathematical parameters, forming an element type-foundation reduction coefficient lookup table, which provides a physical basis for subsequent analysis.

[0131] Spatial overlay analysis technology is used to perform precise spatial calculations on the identified feature vector map and the divided slope structure area vector map. This includes calculating spatial statistical indicators such as the quantity, density, area, and attitude of each feature in each area, generating a slope structure area-feature feature association attribute table. This establishes a data correlation between the macroscopic region and the microscopic geological defects, realizing the regional quantitative collection of geological information.

[0132] The integrated element type-basic reduction coefficient lookup table and the slope structure region-element feature association attribute table combine the development intensity index of various elements in the region with their basic reduction coefficients using a preset nonlinear coupling reduction model. This dynamically calculates the comprehensive reduction rate of the rock and soil strength in the region, i.e., the variable reduction rate of each element feature in each specific region. For example, the evolution variables of cracks are different in the platform region, slope region, slope crest line region, and slope bottom line region. Moreover, their evolution variables are also different under different rock mass conditions. This step takes into account the evolution of element features in different regions and comprehensively calculates the slope change and structural deformation, rather than simply superimposing the historical slope structure data of the previous period with the previously identified regional elements. This step is more consistent with the actual geological conditions of slope change and structural deformation compared to existing technologies.

[0133] Among them, the slope change is calculated by using the grid difference method to calculate the change in local surface slope based on the corrected morphological change; the structural deformation is calculated by combining the corrected morphological change and the evolution variables of element characteristics to calculate displacement, strain, etc.

[0134] Embodiment 2 of the present invention provides a slope verification system based on UAV visual recognition. Figure 2This is a schematic diagram of the slope verification system based on UAV visual recognition provided in Embodiment 2 of the present invention, as shown below. Figure 2 As shown, the slope verification system based on UAV visual recognition includes:

[0135] The data preprocessing module is used to collect target slope data through a drone equipped with a multi-source acquisition device, obtain multi-source raw data for slope verification, and fuse the multi-source raw data to generate a multi-scale fused image.

[0136] The region division module is used to extract regional features related to slope stability verification from the multi-scale fused image, and divide the target slope into several slope structure regions that are suitable for slope verification and judgment based on these regional features.

[0137] The feature recognition module is used to identify the feature elements that affect slope stability in the multi-scale fused image, and to map each feature element to the corresponding slope structure region according to the spatial position relationship, thereby obtaining several slope structure regions containing feature elements.

[0138] The variable calculation module is used to obtain historical slope structure data from the previous period. Based on the regional division, it performs structural comparison between several slope structure regions containing feature characteristics and historical slope structure data, and calculates the slope change and structural deformation based on the comparison results using the feature reduction analysis method.

[0139] The alarm module is used to generate a target slope heat map for slope verification based on the slope change and structural deformation of each area, and to complete the data visualization output of slope verification.

[0140] Embodiment 3 of the present invention provides an electronic device, Figure 3 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention, as shown below. Figure 3 As shown, the electronic device includes a processor 21, a memory 22, an input device 23, and an output device 24; the number of processors 21 in the computer device can be one or more. Figure 3 Taking a processor 21 as an example; the processor 21, memory 22, input device 23, and output device 24 in an electronic device can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0141] The memory 22, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor 21 executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory 22, thereby implementing the slope verification method based on UAV visual recognition in Embodiment 1.

[0142] The memory 22 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory 22 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory 22 may further include memory remotely located relative to the processor 21, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0143] Input device 23 can be used to receive user input such as ID and password. Output device 24 is used to output the network configuration page.

[0144] Embodiment 4 of the present invention also provides a computer-readable storage medium, wherein the computer-executable instructions, when executed by a computer processor, are used to implement the slope verification method based on UAV visual recognition as provided in Embodiment 1.

[0145] The storage medium containing computer-executable instructions provided in the embodiments of the present invention is not limited to the method operation provided in Embodiment 1, but can also perform related operations in the slope verification method based on UAV visual recognition provided in any embodiment of the present invention.

[0146] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A slope verification method based on UAV visual recognition, characterized in that, Includes the following steps: Step S1: Collect target slope data using a drone equipped with multi-source acquisition devices to obtain multi-source raw data for slope verification; fuse the multi-source raw data to generate a multi-scale fused image; Step S2: Extract the regional features related to slope stability verification from the multi-scale fused image, and divide the target slope into several slope structure regions suitable for slope verification and judgment based on the regional features. Step S3: Identify the feature elements that affect slope stability in the multi-scale fused image, and map each feature element to the corresponding slope structure region according to the spatial position relationship to obtain several slope structure regions containing feature elements. Step S4: Obtain historical slope structure data from the previous period. Based on regional division, compare the structure of several slope structure regions containing feature characteristics with historical slope structure data. Based on the comparison results, use the feature reduction analysis method to calculate the slope change and structural deformation. Step S5: Based on the slope change and structural deformation of each region, generate a target slope heat map for slope verification, and complete the data visualization output of slope verification.

2. The slope verification method based on UAV visual recognition according to claim 1, characterized in that, Extracting regional features related to slope stability verification from the multi-scale fused image includes the following steps: The multi-scale fused image is input into the feature extraction model, and the feature extraction model is used to perform multi-scale parallel extraction to generate multiple sets of feature maps containing detailed and global information. The importance weights of features at different scales in the multiple sets of feature maps for identifying boundary characteristics and structural surface characteristics at each spatial location are determined, and engineering semantic fusion is performed on the multiple sets of feature maps based on the importance weights to obtain a multi-scale engineering semantic feature map. A geological prior attention map is constructed, and the geological prior attention map is used as a gating signal to perform element-wise multiplication with the multi-scale engineering semantic feature map to obtain a primary slope feature map containing regional features.

3. The slope verification method based on UAV visual recognition according to claim 2, characterized in that, The feature extraction model includes a high-resolution detail module; wherein, multi-scale parallel extraction using the high-resolution detail module includes the following steps: Guided fusion of multi-scale fused images is performed to generate a multimodal guided fusion input tensor; The multi-directional Sobel gradient operator is applied to the multimodal guided fusion input tensor for convolution, and the intensity and directional gradient maps in multiple directions are extracted to generate multi-directional gradient feature maps. The multimodal guided fusion input tensor and the multidirectional gradient feature map are concatenated in the channel dimension to form a primary enhanced feature map; After smoothing and activating the primary enhanced feature map using a 5*5 depth separable convolutional layer, a deformable convolutional layer is then connected to extract non-rigid feature contours, generating a geometrically adaptive feature map. The geometric adaptive feature map is sequentially passed through two 3*3 depth separable convolutional layers to extract texture / grain features and local pattern features. The extracted texture / grain features and local pattern features are then horizontally connected with the geometric adaptive feature map to generate a multi-scale detail feature map. Calculate the spatial attention map and channel attention vector of the multi-scale detail feature map, and use the spatial attention map and channel attention vector to reweight the multi-scale detail feature map to generate an attention-reweighted feature map; The attention-reweighted feature map is subjected to engineering semantic interaction by small dilated convolutional blocks to generate a local engineering context feature map.

4. The slope verification method based on UAV visual recognition according to claim 2, characterized in that, Based on the characteristics of the region, the target slope is divided into several slope structure regions for adaptive slope verification and judgment, including the following steps: Align the primary slope feature map with the three-dimensional point cloud data of the target slope, and quantify the aligned primary slope feature map into an engineering geological parameter map. Based on the quantified engineering geological parameter map, a clustering algorithm is used to group the slope pixels, and areas with similar lithological combinations and structural surface development characteristics are merged into the same mean unit, outputting a primary slope mechanical property zoning map. The morphological verification of the primary slope mechanical property zoning map is performed to generate a slope structure region division map.

5. The slope verification method based on UAV visual recognition according to claim 1, characterized in that, Identifying the features affecting slope stability in the multi-scale fused image, and mapping each feature to the corresponding slope structure region according to its spatial location, including: Based on a deep learning-based linear feature segmentation model and Hough transform, linear features are identified from high-resolution multi-band orthophotos in the multi-scale fused image. Based on the DEM model, the linear features are verified by slope aspect consistency analysis and surface roughness calculation, and a vectorized structural surface line segment map is generated. By combining the lithological and geomorphological spectral library, spectral angle mapping is used to perform lithological analysis on the high-resolution multi-band orthophoto image, generating a lithological classification map and a spatial probability distribution map of weak interlayers. The high-resolution multi-band orthophoto image is used to identify micro-topographic deformation by combining terrain parameter threshold segmentation technology with image texture, and a micro-topographic deformation marker distribution map is generated. The vectorized structural surface line segment map, the lithology classification map, the spatial probability distribution map of the weak interlayer, and the distribution map of the micro-geomorphic deformation markers are subjected to spatiotemporal correlation analysis, and the results of the spatiotemporal correlation analysis are mapped to the corresponding slope structural regions to generate slope element feature maps.

6. The slope verification method based on UAV visual recognition according to claim 1, characterized in that, To obtain historical slope structure data from the previous period, and based on regional division, to perform structural comparison between several slope structure regions containing feature characteristics and historical slope structure data, the following steps are included: Spatially align several slope structure regions containing feature characteristics with historical slope structure data to generate a spatially aligned image; Calculate the morphological changes and feature evolution variables of the spatially aligned image in each region.

7. The slope verification method based on UAV visual recognition according to claim 6, characterized in that, Based on the comparison results, the slope change and structural deformation are calculated using the element reduction analysis method, including the following steps: Based on rock mechanics theory, a vector of basic strength reduction coefficients for each element is constructed, and an element type-basic reduction coefficient lookup table is generated. The spatial overlay analysis method is used to quantify the characteristics of each element within each slope structure area, and generate a slope structure area-element characteristic association attribute table. By combining the comprehensive element type-basic reduction coefficient lookup table and the slope structure area-element feature association attribute table, parameter reduction corrections are made for the morphological change and the element feature evolution variables. Based on the morphological changes and element characteristic evolution variables after parameter reduction and correction, calculate the slope change and structural deformation.

8. A slope verification system based on UAV visual recognition, used to implement the slope verification method based on UAV visual recognition as described in any one of claims 1 to 7, characterized in that, include: The data preprocessing module is used to collect target slope data by using a drone equipped with multi-source acquisition devices to obtain multi-source raw data for slope verification. The multi-source raw data are fused to generate a multi-scale fused image; The region division module is used to extract regional features related to slope stability verification from the multi-scale fused image, and divide the target slope into several slope structure regions that are suitable for slope verification and judgment based on these regional features. The feature recognition module is used to identify the feature elements that affect slope stability in the multi-scale fused image, and to map each feature element to the corresponding slope structure region according to the spatial position relationship, thereby obtaining several slope structure regions containing feature elements. The variable calculation module is used to obtain historical slope structure data from the previous period. Based on the regional division, it performs structural comparison between several slope structure regions containing feature characteristics and historical slope structure data, and calculates the slope change and structural deformation based on the comparison results using the feature reduction analysis method. The alarm module is used to generate a target slope heat map for slope verification based on the slope change and structural deformation of each area, and to complete the data visualization output of slope verification.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the slope verification method based on UAV visual recognition as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the slope verification method based on UAV visual recognition as described in any one of claims 1 to 7.

Images