A landslide recognition method based on multi-directional edge-gradient information enhancement

Abstract

The application discloses a landslide identification method based on multi-directional edge-slope information enhancement, comprising the following steps: step 1, obtaining a reservoir area image and pre-processing; step 2, constructing a landslide identification model based on a deep learning framework; step 3, training and testing the landslide identification model; step 4, using the trained landslide identification model to identify landslides; the landslide identification model in step 2 is a three-stage architecture of Backbone-Neck-Head. The method aims to realize high-precision positioning of small landslide targets, efficient modeling of complex landform structures and lightweight deployment on platforms with limited computing resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of reservoir landslide identification, and particularly relates to a landslide identification method based on multi-directional edge-slope information enhancement. BACKGROUND

[0002] In recent years, influenced by the frequent occurrence of extreme climate events and the continuous enhancement of human activity intensity, geological disasters such as reservoir landslides have shown a high and frequent trend, which seriously threatens the regional ecological safety, the stability of water conservancy projects and the safety of the lives and property of surrounding residents. Especially in typical reservoirs such as mountain valley reservoirs and large-scale water storage power stations, influenced by multiple factors such as geological structure, water level fluctuation, vegetation change and engineering disturbance, landslide disasters usually have the characteristics of strong suddenness, high concealment and large destructiveness, and their dynamic monitoring and risk warning have become the key and difficulty of current geological disaster prevention and control work.

[0003] Traditional landslide identification methods, such as manual patrol and fixed monitoring station layout, although have certain accuracy and reliability, generally have limitations such as insufficient real-time performance, limited spatial coverage and large consumption of human resources, and are difficult to adapt to the high-frequency, large-scale and automated information acquisition needs of modern disaster monitoring.

[0004] Therefore, automatic landslide identification technology based on unmanned aerial vehicle remote sensing images has emerged as the times require, and has gradually become an important development direction of intelligent geological disaster prevention and control. At present, the methods for realizing accurate identification of landslide areas in unmanned aerial vehicle remote sensing images mainly include the following three kinds: landslide identification method based on traditional image processing and geometric features, landslide identification method based on deep learning, and landslide identification method based on geological and multi-source remote sensing data fusion. Although these methods have achieved certain accuracy improvement on the public landslide remote sensing data set, there are still various problems in real complex scenes such as reservoirs.

[0005] Landslide identification method based on traditional image processing and geometric features: this kind of method mainly relies on color, texture, shape and terrain information in remote sensing images, and adopts classical image processing techniques such as image segmentation, edge detection, texture analysis and threshold segmentation to extract targets. This kind of method is relatively simple to implement on resource-limited platforms, has small computational overhead, and has certain practicality and deployment advantages. However, since it highly depends on low-level visual features, it is easily affected by external disturbances such as light changes, terrain obstructions and vegetation cover, resulting in high misidentification rate and poor stability in small-scale areas with complex landslide morphology and weak texture features, and is difficult to perform accurate identification tasks in real environments.

[0006] Landslide recognition methods based on deep learning: This type of method is usually based on an end-to-end neural network model, and representative architectures include semantic segmentation models centered on U-Net, object detection models represented by the YOLO series, and context understanding models based on the Transformer model. U-Net extracts high-level semantic features through downsampling and restores spatial details using skip connections, making it suitable for pixel-level segmentation tasks. However, it often suffers from information loss when dealing with small-scale targets and fuzzy boundary areas, and has problems such as large parameter quantity and low inference efficiency. The YOLO series has good detection speed and target positioning ability, but still lacks the ability to model irregular landslide boundaries and recognize small landslides. Transformer-based models perform well in capturing long-range context dependencies and resisting complex background interference, but their high computational complexity and resource requirements limit their deployment suitability on edge platforms, especially unmanned aerial systems. Overall, this type of method has limitations such as insufficient boundary fitting accuracy, limited fine-grained structure modeling capability, and high model redundancy in actual reservoir area remote sensing applications.

[0007] Landslide recognition methods based on the fusion of geology and multi-source remote sensing data: This type of method uses optical images, synthetic aperture radar (SAR), laser radar, and historical geological survey data to realize collaborative modeling of multi-source information and comprehensive discrimination of landslide areas. SAR has strong penetration ability and can provide continuous observation in adverse weather conditions such as clouds and rain. Laser radar can output high-precision three-dimensional terrain structure, and geological data provides regional landslide evolution background and geological stability information. Although multi-source fusion methods have strong detection robustness and information complementarity advantages in theory, they still face challenges in practical engineering applications, including high data acquisition and processing costs, spatial resolution and time frequency differences between different sensors, making data registration and feature alignment difficult, and complex feature fusion paths, making it difficult to build an efficient and unified recognition framework. The above factors significantly restrict the promotion and practical deployment of this type of method in real-time unmanned aerial monitoring tasks.

[0008] In summary, in the context of unmanned aerial remote sensing applications for reservoir landslide recognition, existing deep learning-based methods have improved the detection efficiency of landslide areas to some extent, but there are still problems to be solved: (1) insufficient small-scale landslide target recognition capability in complex terrain backgrounds; (2) insufficient landslide target geometric structure modeling capability; (3) high-performance model deployment challenges under limited computing resources.

[0009] In view of this, the present invention is proposed. SUMMARY

[0010] The purpose of the present application is to solve the problems existing in the prior art, and to propose a landslide recognition method based on multi-directional edge-slope information enhancement, aiming to realize high-precision positioning of small landslide targets, efficient modeling of complex landform structures, and lightweight deployment on platforms with limited computing resources.

[0011] In order to achieve the above-mentioned purpose, the present application also adopts the following technical solutions:

[0012] A landslide recognition method based on multi-directional edge-slope information enhancement, comprising the following steps:

[0013] Step 1: Obtain the image of the reservoir area and perform preprocessing;

[0014] Step 2: Based on the deep learning framework, construct a landslide recognition model;

[0015] Step 3: Train and test the landslide recognition model;

[0016] Step 4: Use the trained landslide recognition model to recognize landslides;

[0017] The landslide recognition model in step 2 is a three-stage architecture of Backbone-Neck-Head;

[0018] The Backbone part includes two CBS modules (Convolution + BatchNormalization + SiLU Activation, CBS) connected in turn, which are used to extract basic texture and edge features of the image and realize preliminary down-sampling; two CSP modules (CrossStage Partial Network, CSP) are used to realize multi-stage feature fusion; MVM module (Multi-scale Variable Module, MVM) is used to extract deep features and enhance boundary and slope information expression; MSAVM module (Multi-Scale Attention Variable Module, MSAVM) is used to realize global semantic modeling and feature alignment at 32 resolution, and is used to extract bottom features;

[0019] ​The Neck part comprises: three BHFM modules (Bi-branch Hybrid Fusion Module, BHFM) respectively connected with two CSP modules and the MVM module of the Backbone part, for decomposing features into high-frequency and low-frequency parts and performing weighted fusion; three MVM modules respectively connected with the three BHFM modules, the MVM module of the first layer connected with the MVM module of the second layer, and the MVM module of the second layer connected with the MVM module of the third layer, for realizing multi-scale semantic enhancement; two CCGM modules (Contextual Cue Guidance Module, CCGM) connected with the MVM module of the first layer and the GGM module of the lower layer, and the CCGM of the lower layer connected with the MVM modules of the second layer and the third layer, for supplementing contextual information and fine-grained feature expression; an SPPF module (Spatial Pyramid Pooling - Fast, SPPF) connected with the MSAVM module of the Backbone part, for completing multi-receptive field fusion; and a MSAVM module connected with the SPPF module, for realizing feature alignment and enhancement, and connected with the MVM module of the third layer through a residual path to keep the main information flowing.

[0020] The MVM module and the MSAVM module are embedded with DEHB modules (Direction-Enhanced Hybrid Block, DEHB), and the MSAVM module fuses a multi-head self-attention mechanism on the basis of the MVM module.

[0021] The DEHB module comprises: horizontal depth convolution (DWConv3x3 horizontal), vertical depth convolution (DWConv3x3 horizontal), feature concatenation (Concat), channel attention mechanism (SE), 1x1 convolution (Conv1x1), and a VSS (Virtual Slope Simulation) module; the feature concatenation is used for concatenating the outputs of the horizontal depth convolution and the vertical depth convolution; the feature concatenation, the channel attention mechanism, the 1x1 convolution, and the VSS module are sequentially connected; the VSS module comprises: a main branch and a residual branch, the main branch comprises: normalization (Norm), linear mapping (Linear), depth separable convolution (DWConv), SiLU activation function, 2D-Scan operator, and normalization (Norm), and the residual branch comprises: linear mapping (Linear); after the outputs of the main branch and the residual branch are added and fused, the final output is generated through linear mapping (Linear).

[0022] Further, step 1 comprises the following steps:

[0023] Step 1.1: Collect high-resolution remote sensing image data by deploying a UAV platform in the warehouse area to capture multi-angle aerial photography of typical landslide scenes and obtain rich topographic and geomorphic information covering the landslide area.

[0024] Step 1.2: Data cleaning, to address image noise and data missing problems in remote sensing image data, use multi-frame fusion and interpolation methods for image quality enhancement processing.

[0025] Step 1.3: Normalization, use Min-Max normalization method to linearly scale the image to the [0, 1] interval, the specific calculation formula is as follows:

[0026] ;

[0027] In the formula, is the original image pixel value, , respectively represent the minimum and maximum pixel values in the image;

[0028] Step 1.4: Construct the dataset, combine historical landslide remote sensing images, geological disaster survey reports, and field-collected geographic annotation information to manually fine-label the landslide area in the remote sensing image to construct the dataset.

[0029] Further, the CSP module in the Backbone part includes: 1x1 convolution (Conv1x1), channel split, residual convolution unit, convolution subnet, cross-layer concatenation (Concat), and 1x1 convolution (Conv1x1); the first layer 1x1 convolution is input into the residual convolution unit and the convolution subnet after channel split; the cross-layer concatenation is used to concatenate the outputs of the residual convolution unit and the convolution subnet, and input into the second layer 1x1 convolution.

[0030] Further, the MVM module in the Backbone part and the Neck part includes: 1x1 convolution (Conv1x1), channel split, two-stage DEHB module, cross-layer concatenation (Concat), and 1x1 convolution (Conv1x1) connected in turn; the cross-layer concatenation is used to concatenate the outputs of the first layer 1x1 convolution and the two-stage DEHB module.

[0031] Further, the MSAVM module in the Backbone part and the Neck part includes: 1x1 convolution (Conv1x1), channel split, DEHB module, multi-head self-attention mechanism (MHSA), cross-layer concatenation (Concat), and 1x1 convolution (Conv1x1) connected in turn; the cross-layer concatenation is used to concatenate the outputs of the first layer 1x1 convolution and the DEHB module.

[0032] Further, the BHFM module in the Neck part comprises: multi-scale orthogonal convolution, channel attention, double-flow feature fusion, and normalization.

[0033] Further, the CCGM module in the Neck part comprises: a low-resolution feature map processing branch, a high-resolution feature map processing branch, and a 1×1 convolution; the low-resolution feature map processing branch is processed in sequence through a depth split convolution (DWConv3×3), a 1×1 convolution, and a grid sample operation; the high-resolution feature map processing branch is processed in sequence through a sparse convolution (Spance Conv3×3), a 1×1 convolution, a Gaussian Error Linear Unit (GELU), and a 1×1 convolution; the outputs of the low-resolution feature map processing branch and the high-resolution feature map processing branch are added and then processed through a 1×1 convolution to obtain a final output.

[0034] Further, the SPPF module in the Neck part comprises, in sequence: a 1×1 convolution (Conv1×1), a three-level 5×4 max-pooling, splicing, and a 1×1 convolution (Conv1×1).

[0035] Further, the Head part adopts a three-scale detection head to comprehensively judge the fused features, and outputs the position coordinates, category information, and confidence score of the landslide target.

[0036] Further, in the model training and testing process of step 2, the loss function is composed of three parts: a bounding box regression loss, a classification loss, and a confidence loss; the bounding box regression loss adopts a generalized intersection over union loss function, and the formula is:

[0037] ;

[0038] In the formula, is a predicted box, is a real labeled box, and the spatial overlap and distance between the two are calculated to improve the positioning accuracy.

[0039] The classification loss uses a cross-entropy loss function, and the formula is:

[0040] ;

[0041] In the formula, is a real category label, is a predicted category probability, and the loss function measures the difference between the predicted probability and the real label to improve the discrimination ability of the landslide type.

[0042] The confidence loss adopts a binary cross-entropy loss function, and the formula is:

[0043] ;

[0044] In the formula, t represents landslide / background label, p is prediction confidence, is used to suppress background false detection and improve detection reliability;

[0045] The comprehensive loss function is defined as:

[0046] ;

[0047] In the formula, =1.0, =1.0, =0.5, through back propagation, the loss value of a single batch of images is fed back to each layer of the network to dynamically adjust the parameters and improve the prediction accuracy;

[0048] In terms of optimization strategy, first, the adaptive momentum optimizer is selected, the initial learning rate is set to , and the weight decay coefficient is 0.0005; second, the cosine annealing learning rate scheduling mechanism is adopted, and the learning rate is smoothly decayed from the initial value to close to zero within 150 training rounds; during the training process, random enhancement operations are applied to the input batch of images, including cropping, scaling, flipping and brightness adjustment, to improve the robustness of the model in complex scenes; finally, when the average precision mean of the validation set does not appear significant improvement in continuous 5 rounds, trigger the early stopping mechanism and save the current optimal model parameters, to avoid overfitting and shorten the training time.

[0049] Compared with the prior art, the present application has the following advantages:

[0050] 1. The existing convolutional neural network generally relies on limited local receptive field in the feature extraction process, when the landslide area is similar to the surrounding forest, water, bare soil and other background textures, feature confusion and detection omission are easy to occur. To solve this problem, the present application designs a depth separable convolution in the DEHB module in the horizontal and vertical directions, explicitly extracts the gradient change and edge details in different directions, and then strengthens the high-weight features related to the landslide boundary through the channel attention mechanism, so as to highlight the key boundary and suppress the background noise. With the help of the module, the model's ability to distinguish the outline of the landslide area in complex background is significantly improved, realizing higher precision identification of small-scale landslide targets.

[0051] 2、The formation process of landslide disaster is closely related to topographic slope and structural deformation, and the existing method often lacks deep modeling ability for spatial geometric properties, leading to misidentification in the transition area of landform. To solve this problem, the VSS module is introduced in the DEHB module. The module first normalizes and linearly maps the features to stabilize the distribution, then uses deep separable convolution and nonlinear activation to extract local slope texture, and performs cumulative weighted scanning in the horizontal and vertical directions through the 2D-Scan operator, and encodes the elevation change pattern into the feature representation. Thus, the network can explicitly perceive the tilting trend and structural change of the landform, and has stronger geometric structure semantic modeling and discrimination ability under complex slope conditions, effectively reducing the false detection rate in the transition area of landform.

[0052] 3、The mainstream deep model has large parameter quantity and high computational complexity, which is difficult to directly deploy on resource-limited unmanned aerial vehicle platform, affecting its practicability and response speed in real-time landslide monitoring in the field. To solve this problem, the model design idea based on lightweight network skeleton is proposed, and the direction information enhancement mechanism and lightweight attention structure are innovatively integrated. In the overall architecture, the CSP module in the Backbone part effectively reduces redundant calculation through channel shunting and residual reuse, and the DEHB module in the MVM module and the MVM module uses deep separable convolution and virtual slope modeling to reduce FLOPs while enhancing boundary and slope feature expression; The multi-head self-attention mechanism in the MSAVM module adopts parallel grouping mode, which significantly compresses the computational overhead; The BHFM module in the Neck part uses sparse convolution through the CCGM module and replaces the large convolution kernel with sequential pooling through the SPPF module, which embodies the lightweight design idea. Finally, the detection head uses 1×1 convolution compression and multi-scale prediction branch to ensure detection accuracy while reducing inference delay. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 It is a flowchart of a landslide recognition method based on multi-directional edge-slope information enhancement;

[0054] Figure 2 It is an overall architecture diagram of a landslide recognition model;

[0055] Figure 3 It is a structure diagram of DEHB module;

[0056] Figure 4 It is a structure diagram of CSP module;

[0057] Figure 5 It is a structure diagram of MVM module;

[0058] Figure 6 It is a structure diagram of MSAVM module;

[0059] Figure 7 Structure diagram of BHFM module;

[0060] Figure 8 Structure diagram of CCGM module;

[0061] Figure 9 Structure diagram of SPPF module. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application.

[0063] Embodiment one:

[0064] In complex mountainous environments, the monitoring and identification of landslide areas require high real-time and accuracy of equipment. In addition, the remote sensing images of the reservoir area are often accompanied by low image resolution, strong background interference, fuzzy boundaries, and small target size, which greatly increases the technical difficulty of landslide detection tasks. Traditional target detection methods generally have limitations such as insufficient detection accuracy, fuzzy boundary discrimination, and difficulty in deploying lightweight platforms in such scenarios. To address the above challenges, the present embodiment proposes a landslide identification method based on multi-directional edge-slope information enhancement, as shown in Figure 1 The method includes the following steps:

[0065] Step 1: Obtain the reservoir image and perform preprocessing;

[0066] Step 2: Based on the deep learning framework, construct a landslide identification model;

[0067] Step 3: Train and test the landslide identification model;

[0068] Step 4: Use the trained landslide identification model to identify landslides.

[0069] As shown in Figure 2 The landslide identification model is a three-stage architecture of Backbone-Neck-Head.

[0070] The Backbone part includes, in sequence, two CBS modules (Convolution + Batch Normalization + SiLU Activation, CBS) for basic texture and edge feature extraction of the image and preliminary down-sampling, two CSP modules (Cross Stage Partial Network, CSP) for multi-stage feature fusion, an MVM module (Multi-scale Variable Module, MVM) for deep feature extraction and enhancement of boundary and slope information expression, and an MSAVM module (Multi-Scale Attention Variable Module, MSAVM) for global attention mechanism and long-range dependency modeling. 32Resolution global semantic modeling and feature alignment are performed to extract bottom features.

[0071] In this embodiment, the Backbone part is used for hierarchical spatial feature extraction. Specifically, the CBS module is used for convolution, batch normalization and SiLU activation function processing of the input image to complete preliminary feature extraction; the CSP module is used for multi-stage feature fusion to further improve the extraction ability and feature expression efficiency of the model and provide stable and hierarchical representation for subsequent modeling; in the deep feature extraction stage, the MVM module is introduced to enhance the expression ability of the landslide boundary and slope related features; in the bottom stage, the MSAVM module is further used to strengthen the long-range dependency modeling through the global attention mechanism and provide high semantic support for the multi-scale feature fusion of the Neck part.

[0072] In this embodiment, the multi-layer features generated in the hierarchical spatial feature extraction stage undertake differentiated functions in the detection head. First, the shallow high-resolution features from the CBS module and the CSP module retain a large amount of edge contour and texture detail information, and these features are directly input into the small target and medium target prediction branches of the detection head to improve the boundary delineation accuracy and positioning stability of the micro landslide area. Then, the deep features processed by the MVM module are significantly enhanced in direction sensitivity and slope modeling, and can highlight the fine-grained geometric structure of the landslide area. On this basis, the MSAVM module as a supplementary enhancement unit performs joint modeling of direction-slope and global dependency on deep and medium features, and through cross-layer information interaction and multi-scale fusion mechanism, the enhanced semantic features are fed back to the medium-scale and large-scale branches, thereby making up for the feature defects caused by the lack of long-range dependency and complex background interference.

[0073] The Neck part includes: three BHFM modules (Bi-branch Hybrid Fusion Module, BHFM) connected with two CSP modules and the MVM module of the Backbone part respectively, for decomposing features into high-frequency and low-frequency parts and performing weighted fusion; three MVM modules connected with the three BHFM modules respectively, the MVM module of the first layer connected with the MVM module of the second layer, and the MVM module of the second layer connected with the MVM module of the third layer, for realizing multi-scale semantic enhancement; two CCGM modules (Contextual Cue Guidance Module, CCGM) connected with the MVM module of the first layer and the GGM module of the lower layer, and connected with the MVM modules of the second layer and the third layer, for supplementing context information and fine-grained feature expression; an SPPF module (Spatial Pyramid Pooling - Fast, SPPF) connected with the MSAVM module of the Backbone part, for completing multi-receptive field fusion; and the MSAVM module connected with the SPPF module, for realizing feature alignment and enhancement, and connected with the MVM module of the third layer, for keeping the backbone information flowing through the residual path.

[0074] In the embodiment, the Neck part is used for multi-scale context perception enhancement, specifically: the three BHFM modules decompose features into high-frequency and low-frequency parts and perform weighted fusion by referring to the frequency division mechanism of the human visual system, automatically balance the importance of structural information and detail information, and make the network have stronger discriminability when processing landslide edge and background transition area; the three MVM modules are double-input, two paths of features are spliced in the channel dimension after resolution alignment, and then input into the MVM module as a single input, so as to realize the fusion and enhancement of cross-layer information while keeping the single-input design of the module; the CCGM module is used for supplementing context information and fine-grained feature expression, enhancing the discriminability of small-scale targets, and realizing spatial cue modeling of cross-layer features; for the 1 / 32 feature map of the bottom layer of the Backbone part, first, the SPPF module is used to complete multi-receptive field fusion, further aggregate multi-receptive field features, improve detection efficiency and structural robustness, and then input the MSAVM module to realize feature alignment and enhancement by using the global self-attention mechanism, and keep the backbone information flowing through the residual path; finally, the bottom layer features and the middle and high layer features are processed by multi-scale and context enhancement before entering the Head part, so as to significantly improve the robustness and precision of the model in small target detection and complex background scene.

[0075] In this embodiment, in the multi-scale context-aware enhancement stage, different modules further enrich the global dependence and semantic expression ability of the features. The BHFM module introduces cross-scale information interaction in each scale feature stream, improves the context consistency between different resolution features, and provides semantic-structural complementary input for the subsequent MVM module and MVM module. The MSAVM module fuses local directional slope information and global multi-head attention results in the lowest resolution feature, so that the large-scale prediction branch has stronger global modeling and background suppression ability when dealing with landslide detection tasks with large shape span. The CCGM module feeds back the global semantic information of the deep layer to the shallow and middle branches through cross-layer semantic guidance, which significantly improves the positioning and recognition accuracy of small target branches in complex background. The SPPF module at the end of the Backbone expands the receptive field in the lowest resolution feature, and supplements the large-scale prediction branch with more rich global spatial pattern perception ability.

[0076] The MVM module and the MSAVM module are embedded with a DEHB module (Direction-Enhanced Hybrid Block, DEHB), and the MSAVM module fuses a multi-head self-attention mechanism on the basis of the MVM module.

[0077] In this embodiment, the MVM module and the MSAVM module are embedded with a DEHB module (Direction-Enhanced Hybrid Block, DEHB), which combines multi-directional depth separable convolution, channel attention mechanism and slope simulation unit, effectively capturing the boundary details and terrain slope changes of the landslide area. Among them, the MSAVM module additionally fuses a multi-head self-attention mechanism on the basis of the MVM module to obtain more comprehensive global dependence information, providing a feature representation with both fine-grained and global perception for landslide recognition.

[0078] As Figure 3As shown, the DEHB module includes: horizontal depthwise convolution (DWConv3x3 horizontal), vertical depthwise convolution (DWConv3x3 horizontal), feature concatenation (Concat), channel attention mechanism (SE), 1x1 convolution (Conv1x1), and pseudo-slope simulation module (VSS); the feature concatenation is used for concatenating the outputs of the horizontal depthwise convolution and the vertical depthwise convolution; the feature concatenation, the channel attention mechanism, the 1x1 convolution, and the pseudo-slope simulation module are sequentially connected; the pseudo-slope simulation module includes: a main branch and a residual branch, the main branch includes: normalization (Norm), linear mapping (Linear), depthwise separable convolution (DWConv), SiLU activation function, 2D-Scan operator, and normalization (Norm), the residual branch includes: linear mapping (Linear), and after the main branch output and the residual branch output are added and fused, the final output is generated through linear mapping (Linear).

[0079] In the embodiment, the design concept of the DEHB module is the principle of "direction-slope detail modeling + structural feature enhancement + residual / cross-layer information reuse". Specifically, the depthwise separable convolution in the horizontal and vertical directions realizes multi-directional boundary detail modeling, which alleviates the insufficient description of complex edges by the local receptive field; the feature concatenation and the channel attention mechanism further highlight the direction and slope features related to landslides and suppress redundant background information, thereby realizing the enhancement of structural features; the 1x1 convolution and the pseudo-slope simulation module effectively reuse the shallow and deep information through linear mapping and 2D-Scan operators under the guidance of the residual path, and cooperatively encode the local geometry and high-level semantic features of the terrain slope change. Under the premise of maintaining lightweight calculation, the module organically combines the terrain boundary and the slope change pattern, and provides structural perception fine feature expression for landslide identification

[0080] In the DEHB module, first, the input features are respectively subjected to depthwise separable convolution in the horizontal and vertical directions, and the horizontal collapse boundary and the vertical slope texture are explicitly extracted. Compared with standard convolution, this operation has more stability and lightweight advantage in modeling of fine line structures such as cracks and fracture zones.

[0081] ;

[0082] In the formula, is the horizontal edge feature, For vertical edge features. Subsequently, the horizontal and vertical edge features are spliced in the channel dimension to realize cross-direction feature fusion, thereby alleviating the noise diffusion problem of single-path convolution in complex backgrounds. The fused features are input into the channel attention mechanism module, and the channel weights are generated through global average pooling and two fully connected layers to enhance the high-weight channels related to landslide discrimination while preserving low-frequency geometric and positional information. And through 1x1 convolution in the channel dimension, the direction features related to landslide discrimination are highlighted. Finally, the features are sent to the pseudo-slope simulation module, and the input features are normalized (Norm) and linearly mapped to stabilize the feature distribution and complete the preliminary feature projection. Subsequently, local texture and geometric morphology are extracted through depthwise separable convolution (DWConv), and then SiLU activation is performed to enhance the non-linear expression ability. Next, the 2D-Scan operator is used to perform cumulative weighted scanning in the horizontal and vertical directions to encode the elevation change patterns in different directions into the feature representation. Then, the features are normalized again to ensure numerical stability and cross-layer alignment. Finally, the main branch output and the residual branch features are added and then input into the linear mapping layer to generate the final output. In this way, the VSS module can highlight the slope changes and subtle topographic changes while maintaining lightweight, providing enhanced structural semantic features for accurate identification of landslide boundaries and slope structures. In summary, the DEHB module not only preserves the original geometric and positional information, but also deeply fuses direction-slope details and topographic structure enhancement features. Compared with the structure using only conventional convolution, it exhibits better performance and stronger robustness in dealing with typical landslide recognition scenarios such as complex terrain, ambiguous boundaries, and subtle slope changes.

[0083] In summary, the landslide recognition method based on multi-directional edge-slope information enhancement of the embodiment uses a lightweight detection network as the skeleton structure, adopts a three-stage architecture of Backbone-Neck-Head, and is composed of multiple functionally complementary modules. It can realize high-precision detection and robust identification of landslide areas in complex remote sensing scenarios. The DEHB module is integrated to enhance the perception of landform slope trend and landslide boundary based on the extraction of spatial structure features, thereby significantly improving the recognition accuracy and system robustness in complex terrain backgrounds. While considering the detection performance, the model complexity is effectively controlled, and good real-time performance and edge deployment ability are achieved, realizing the organic unification of landslide target detection accuracy and inference efficiency.

[0084] The landslide recognition method based on multi-directional edge-slope information enhancement of the embodiment proposes the DEHB module to address the problem of ambiguous landslide area boundaries and unclear outlines. This module integrates DWConv, VSS, and SE to accurately extract multi-directional landslide edge information, while guiding the network to focus on the changes in landform slope features, effectively enhancing the model's ability to discriminate small-scale landslide target structures.

[0085] The landslide recognition method based on multi-directional edge-slope information enhancement of the embodiment considers the characteristics of strong context correlation and significant target scale difference in remote sensing images, constructs an end-to-end detection framework including multiple feature enhancement units, and further deepens the spatial structure modeling capability. Specifically, the MSAVM module is used to enhance the spatial scale consistency; the BHFM module realizes the collaborative perception of frequency domain edge and semantic features; the CCGM module improves the semantic modeling capability of the model for small targets; and the SPPF module integrates multi-scale receptive field information. The above modules jointly act on the depth modeling and fusion of the DEHB output features, and improve the comprehensive perception capability of the model from multiple angles such as multi-scale modeling, frequency domain guidance and context enhancement. Finally, all the fused features are input into the lightweight detection head to complete the high-precision positioning and classification prediction of the landslide region.

[0086] In a specific embodiment, step 1 includes the following steps:

[0087] Step 1.1: Collect high-resolution remote sensing image data by deploying a UAV platform in the area of the reservoir area to take multi-angle aerial photography of typical landslide scenes, and obtain rich topographic and geomorphic information covering the landslide area;

[0088] In this embodiment, the image data is in RGB format, and the image resolution ranges from 512x512 to 2048x2048, which can cover the geomorphic features of different types of landslide areas. All collected images and related metadata information are stored in the edge computing node or local server carried by the UAV to avoid data loss problems caused by unstable network signals in mountainous areas, and to ensure the integrity and availability of remote sensing data.

[0089] Step 1.2: Data cleaning, for the image noise and data missing problems existing in the remote sensing image data, use multi-frame fusion and interpolation method for image quality enhancement processing;

[0090] In this embodiment, for the noise part in a single frame image, Gaussian filtering and median filtering and other smoothing strategies are applied to realize image denoising. For the missing area in the image, according to the pixel information of the corresponding area in the adjacent multiple frames of images, the mean filling method or the bidirectional interpolation method is used to complete the missing area, so as to improve the integrity and spatio-temporal continuity of the image data.

[0091] Step 1.3: Normalization, the Min-Max normalization method is used to linearly scale the image to the [0, 1] interval, and the specific calculation formula is as follows:

[0092] ;

[0093] In the formula, For the original image pixel value, , respectively represent the minimum and maximum pixel values in the image.

[0094] Due to the different light conditions and sensor device environments in which the original remote sensing image is collected, the pixel distribution is inconsistent. In order to unify the image data input format and improve the stability and generalization ability of model training, in this embodiment, the Min-Max normalization method is used to perform linear scale adjustment on the image, and all pixel values are scaled to the [0, 1] interval. The normalization operation helps to eliminate the pixel dynamic range offset caused by device differences or changes in collection conditions, unify the model input scale, and enhance the training stability.

[0095] Step 1.4: Constructing a dataset, combining historical landslide remote sensing images, geological disaster survey reports, and field collected geographic annotation information, manually fine labeling the landslide area in the remote sensing image to construct a dataset;

[0096] In this embodiment, after completing the image preprocessing, the landslide area in the remote sensing image is manually fine labeled by combining historical landslide remote sensing images, geological disaster survey reports, and field collected geographic annotation information. The labeling method includes two types of spatial expression forms: bounding box and polygon mask, and is uniformly converted into a YOLO format label file. The constructed dataset has high-quality supervision information, providing a solid data foundation for subsequent model training and evaluation.

[0097] In one specific embodiment, as shown in FIG. 1, the CSP module in the Backbone part includes: 1×1 convolution (Conv1×1), channel split, residual convolution unit, convolution subnet, cross-layer concatenation (Concat), and 1×1 convolution (Conv1×1); the first layer 1×1 convolution is input into the residual convolution unit and the convolution subnet after channel split; the cross-layer concatenation is used to concatenate the outputs of the residual convolution unit and the convolution subnet, and input into the second layer 1×1 convolution. Figure 4

[0098] In this embodiment, the input feature map is divided into two parts along the channel direction: one part is directly used as the residual branch , and the other part is sent to the convolution subnet composed of several CBS layers and Bottleneck structures for deep feature extraction to obtain the processing result .

[0099] Subsequently, the and are concatenated in the channel dimension and fused by 1×1 convolution to obtain the final output feature:

[0100] ​ ;

[0101] In the formula, represents the features of the residual branch, which retains the original low-level texture information. represents the deep features extracted by the convolutional subnetwork. is the output feature after fusion, which has both detail and semantic expression ability.

[0102] Compared with the traditional full-channel convolution structure, the CSP module of the embodiment effectively reduces redundant calculation and gradient repeated transmission through channel shunting, and cross-stage fusion ensures the synchronous retention of shallow details and deep semantics. While ensuring the feature expression ability, it significantly reduces the computational complexity and memory overhead, especially suitable for efficient processing of high-resolution images in unmanned aerial vehicle remote sensing landslide detection tasks, and can achieve higher operation efficiency and recognition robustness under limited hardware conditions.

[0103] In a specific embodiment, as shown in Figure 5 , the MVM module in the Backbone part and the Neck part includes, connected in turn: 1x1 convolution (Conv1x1), channel shunting, two-stage DEHB module, cross-layer splicing (Concat) and 1x1 convolution (Conv1x1); the cross-layer splicing is used to splice the outputs of the first layer 1x1 convolution and the two-stage DEHB module.

[0104] In the embodiment, the design concept of the MVM module is the principle of "lightweight direction edge extraction + slope modeling + channel selection + residual / cross-layer information reuse". The input features are first mapped and shunted by the first layer 1x1 convolution, one way as a residual branch to retain low-level semantics and position information, and the other way into the main branch for direction and slope feature enhancement, thereby embodying the idea of residual / cross-layer information reuse.

[0105] The two-stage DEHB modules connected in the main branch in turn gradually extract the landslide boundary and slope detail information through the depth separable convolution and the slope simulation structure in the horizontal and vertical directions, while keeping the lightweight calculation, thereby strengthening the edge-slope modeling capability; in addition, the progressive design of the two-stage DEHB modules enables the network to gradually transition from coarse-grained boundary positioning to fine-grained slope and texture modeling, thereby enhancing the representation ability of the directionality and slope features of the landslide target.

[0106] Finally, the main branch and the residual branch are spliced in the channel dimension, and then fused and compressed by the second layer 1x1 convolution, so as to realize channel selection and feature integration, thereby obtaining output features with direction sensitivity and structural integrity. Thus, the module effectively enhances the ability of the model to depict the edge clues and fine-grained structure of the landslide under limited computational overhead.

[0107] In the embodiment, the structure of the MVM module not only retains the original space and geometric information of the input features, but also significantly enhances the modeling capability of the landslide target directionality, slope, and structural integrity. Compared with the traditional module that only relies on the local receptive field of the convolution kernel, the MVM module exhibits stronger feature separation capability and target discrimination in broken slope bodies, small landslide targets, and complex mountainous backgrounds, thereby providing more robust and multi-scale consistent feature inputs for the subsequent detection head.

[0108] In one specific embodiment, as shown in Figure 6 The MVM module in the Backbone part and the Neck part includes, in sequence, a 1×1 convolution (Conv1×1), a channel split, a DEHB module, a multi-head self-attention mechanism (MHSA), a cross-layer concatenation (Concat), and a 1×1 convolution (Conv1×1). The cross-layer concatenation is used to concatenate the outputs of the first layer 1×1 convolution and the DEHB module.

[0109] In the embodiment, the design concept of the MVM module is the idea of “directional detail modeling + global dependency capturing + residual / cross-layer information reuse”. The input features are first mapped and split in channels by the first layer 1×1 convolution, one of which is reserved as a residual branch to retain low-frequency geometric and position information, and the other of which enters the main branch for direction and slope enhancement, embodying the mechanism of residual and cross-layer information reuse.

[0110] In the main branch, the DEHB module effectively extracts multi-directional boundaries and slope details through the combination of horizontal and vertical depth convolutions to simulate slopes, realizes directional detail modeling, and is consistent with the local geometric enhancement process in the MVM module. The enhanced features are further sent to the MHSA module to capture long-distance dependencies through multi-head parallel computing, thereby realizing global semantic modeling and context information integration.

[0111] Finally, the features of the main branch and the residual branch are concatenated in the channel dimension and compressed and fused through the 1×1 convolution to obtain multi-scale output features that have both local geometric sensitivity and global semantic consistency.

[0112] Through this structure, the MVM module realizes the collaborative modeling of local direction-slope information and global context dependency without significantly increasing the computational complexity and retaining the original geometric and position information, thereby providing fine and global feature expression for landslide recognition.

[0113] Among them, the MHSA module generates query (Q), key (K) and value (V) matrices through three parallel 1x1 convolutions, and divides them into multiple heads to capture semantic features in different subspaces. Q and K perform matrix multiplication in the feature dimension and are normalized to obtain the attention weight matrix A, which represents the correlation between any two spatial positions. Then, A is multiplied by V to realize the weighted aggregation of global context information to each spatial position feature. Finally, the outputs of all heads are concatenated in the channel dimension and fused through 1x1 convolution to realize the unified integration of multi-scale global information. The introduction of MHSA enables the module to explicitly capture the dependencies of the landslide area in a long distance range, thereby showing stronger robustness and discriminability under the conditions of cross-regional morphological changes and complex background interference.

[0114] In one specific embodiment, as shown in FIG. 2, the BHFM module in the Neck part includes: multi-scale orthogonal convolution, channel attention, double-flow feature fusion and normalization. Figure 7

[0115] In this embodiment, first, the input features are first subjected to directional decomposition by orthogonal convolution to efficiently model the structural information in the horizontal and vertical directions, and large-scale orthogonal convolution is used to further expand the receptive field, thereby balancing local detail description and global context perception while maintaining computational efficiency, embodying the design idea of multi-stage receptive field integration; specifically, the input features are subjected to multi-stage receptive field expansion. The first stage uses orthogonal convolution combination (1x5 and 5x1) to simulate the direction selectivity of neural cells, effectively capturing edge structural features in the horizontal and vertical directions. The second stage further expands the receptive field to 19x19 (formed by stacking 1x7 and 7x1 convolutions), while maintaining parameter efficiency and structural lightness while introducing large-scale context information; specifically, the channel attention mechanism is introduced to adaptively adjust the importance of multi-scale features. This mechanism compresses the features into channel description vectors through global average pooling, generates channel weights through two fully connected layers, and multiplies them with the original features channel by channel, achieving significant feature enhancement. This process can highlight the frequency band information highly related to landslide discrimination, while effectively suppressing background noise.

[0116] ​Subsequently, a channel attention mechanism is introduced to weighted select features at different scales and frequency bands, highlighting features relevant to the landslide target height and suppressing redundant interference, thereby improving the discriminative power of feature representation while preserving details. Specifically, the channel attention mechanism adaptively adjusts the importance of multi-scale features. This mechanism compresses features into channel description vectors through global average pooling, generates channel weights through two fully connected layers, and multiplies them channel by channel with the original features to achieve significant feature enhancement. This process can highlight frequency band information relevant to the landslide discrimination height while effectively suppressing background noise.

[0117] Finally, the dual-branch fusion path retains both the fine-grained features of the original resolution and the global features enhanced at multiple scales, achieving contextual complementarity after fusion and compression to form output features that combine detail sensitivity and global robustness. Specifically, a dual-stream feature fusion strategy is adopted: one stream retains the fine-grained features at the original resolution, while the other stream fuses the global features enhanced at multiple scales. The two streams are fused in the spatial dimension after convolution and normalization processing to form a feature representation that takes into account both edge details and context awareness.

[0118] In this embodiment, the BHFM module effectively embodies the design concept of "multi-stage receptive field integration mechanism", enabling the model to exhibit higher robustness and sensitivity in the detection of small-target landslides in complex backgrounds.

[0119] In a specific embodiment, such as Figure 8 As shown, the CCGM module in the Neck section includes: a low-resolution feature map processing branch, a high-resolution feature map processing branch, and a 1×1 convolution. The low-resolution feature map processing branch is processed sequentially by depthwise splittable convolution (DWConv3×3), 1×1 convolution, and grid sampling. The high-resolution feature map processing branch is processed sequentially by sparse convolution (Spance Conv3×3), 1×1 convolution, Gaussian Error Linear Unit (GELU), and 1×1 convolution. The outputs of the low-resolution feature map processing branch and the high-resolution feature map processing branch are added together and then processed by a 1×1 convolution to obtain the final output.

[0120] In this embodiment, the CCGM module aims to improve the detection and classification capabilities of small targets in high-resolution feature maps. The CCGM module enhances the modeling of fine-grained spatial cues through the complementary fusion of low-resolution and high-resolution features, thereby significantly reducing computational overhead while maintaining detection accuracy. It is particularly suitable for depicting detailed areas such as cracks and slope boundaries.

[0121] Specifically, for low-resolution feature maps, first, a depth separable convolution is adopted for feature extraction, which enhances the context awareness of the global scene and reduces the redundancy between channels; then, a dynamic upsampling mechanism is introduced in the low-resolution branch. The sampling point coordinates are generated by a 1x1 convolution, and the low-resolution features are aligned to the high-resolution feature space using the grid sample operation to achieve fine-grained supplementation of key areas.

[0122] For high-resolution feature maps, sparse convolution is used to reduce computational complexity while preserving local details and key spatial information; then, a 1x1 convolution and a Gaussian Error Linear Unit (GELU) are used for feature transformation, followed by a Multi-Layer Perceptron (MLP) to generate enhanced representations between channels, highlighting important context clues related to small targets.

[0123] Finally, the aligned low-resolution features and the enhanced high-resolution features are added channel by channel, and then compressed and fused by a 1x1 convolution to obtain the output features of the CCGM module.

[0124] In one specific embodiment, as shown in FIG. 1, the SPPF module in the Neck part includes, connected in sequence, a 1x1 convolution (Conv1x1), a three-level 5x4 max-pooling, a concatenation, and a 1x1 convolution (Conv1x1). Figure 9

[0125] In this embodiment, the SPPF module aims to expand the feature receptive field and capture multi-scale context information while maintaining low computational complexity, thereby enhancing the robustness of the model to different scale landslide targets, especially for crack, block, and vegetation disturbance areas with significant scale differences. Compared with the traditional Spatial Pyramid Pooling (SPP) module, the SPPF module achieves equivalent large kernel receptive field expansion through sequentially stacked max-pooling operations, significantly reducing redundant computational complexity.

[0126] Specifically, the input feature map is compressed in the channel by the first 1x1 convolution, reducing the computational complexity of subsequent pooling operations and suppressing redundant information. Then, three 5x5 max-pooling operations are performed in sequence, and the result of each pooling is used as the input of the next pooling, forming a pyramid structure with gradually expanding receptive fields. The formula can be expressed as:

[0127] ;

[0128] Next, the original input feature is concatenated with the three pooling results in the channel dimension:

[0129] ​ ;

[0130] Finally, the multi-scale pooling features are fused by a second 1x1 convolution to obtain the output of the SPPF module:

[0131] ;

[0132] This structure effectively improves the model's perception of multi-scale landslide areas without significantly increasing computational overhead, ensuring both fine-grained structure detection accuracy and retaining large-scale background information to assist in context discrimination.

[0133] In a specific embodiment, as shown in Figure 2 , the Head part uses a three-scale detection head to comprehensively discriminate the fused features, outputting the position coordinates, category information, and confidence score of the landslide target.

[0134] In this embodiment, the Head part uses a three-scale detection head to comprehensively discriminate the fused features, outputting the position coordinates, category information, and confidence score of the landslide target. This structure can achieve accurate identification of landslide areas under multi-scale target and complex background conditions, balancing detection accuracy and computational efficiency, and is suitable for real-time deployment and application on unmanned aerial vehicle platforms.

[0135] Specifically, different scale branches receive inputs from the hierarchical spatial feature extraction and multi-scale context perception enhancement stages, and first complete channel compression and feature reorganization through 1x1 convolution to reduce computational load and improve feature discriminability. Each branch then regresses the target's center point coordinates, width and height parameters, target confidence, and class probability to generate a multi-scale candidate box set. Finally, Sigmoid function and Non-Maximum Suppression (NMS) are used to filter the candidate boxes, remove redundant boxes, and output the final detection results, thereby achieving the synergistic effect of shallow structure details and high-level global semantics, balancing detection accuracy and robustness in multi-scale landslide detection tasks.

[0136] In a specific embodiment, during the step 2 model training process, a multi-task joint loss function and optimization configuration are designed to improve detection accuracy and convergence efficiency. The loss function consists of three parts: bounding box regression loss, classification loss, and confidence loss. The bounding box regression loss uses a generalized intersection over union loss function, with the formula:

[0137] ;

[0138] In the formula, is the predicted box, is the true labeled box, and the spatial overlap and distance between the two are calculated to improve positioning accuracy.

[0139] The classification loss uses a cross-entropy loss function, and the formula is as follows:

[0140]

[0141] In the formula, y represents the true class label, and p represents the predicted class probability. The loss function measures the difference between the predicted probability and the true label to improve the discrimination of landslide types. The confidence loss uses a binary cross-entropy loss function, and the formula is as follows:

[0142]

[0143] In the formula, t represents the landslide / background label, and p represents the predicted confidence. The confidence loss is used to suppress false detections of the background and improve the reliability of detection.

[0144] The comprehensive loss function is defined as follows:

[0145]

[0146]

[0147] In the formula, λ represents the weight of the classification loss, and λ represents the weight of the confidence loss. The comprehensive loss function is used to dynamically adjust the parameters and improve the prediction accuracy through backpropagation. = 1.0, = 1.0, = 0.5. Through backpropagation, the loss value of a single batch of images is fed back to each layer of the network to dynamically adjust the parameters and improve the prediction accuracy.

[0148] In terms of optimization strategy, first, the adaptive momentum optimizer (AdamW) is selected, with an initial learning rate of and a weight decay coefficient of 0.0005. This combination balances convergence speed and generalization ability while suppressing overfitting. Second, the cosine annealing learning rate scheduling mechanism (CosineAnnealing) is used to smoothly decay the learning rate from the initial value to near zero within 150 training rounds, achieving stable convergence in the later training period. During training, random augmentation operations are applied to the input batch of images, including cropping, scaling, flipping, and brightness adjustment, to improve the model's robustness in complex scenarios. Finally, when the mean average precision (mAP) of the validation set does not show significant improvement for 5 consecutive rounds, the early stopping mechanism is triggered, and the current optimal model parameters are saved to avoid overfitting and shorten the training time.

[0149] ​​​​​After the model training is completed, the model containing the optimal parameter weight is deployed on the edge computing node of the unmanned aerial vehicle platform or the ground station server. The obtained remote sensing image data is preprocessed and directly sent into the model as input to realize end-to-end automatic detection.

[0150] In summary, for the problems in the prior art:

[0151] (1) Insufficient small-scale landslide target recognition ability in complex topographic background

[0152] The existing convolutional neural network generally relies on limited local receptive fields in the feature extraction process. When the landslide area is similar to the surrounding forest land, water body, bare soil and other background textures, feature confusion and detection omission are prone to occur. To solve this problem, the embodiment designs a depth separable convolution in the DEHB module in the horizontal and vertical directions, explicitly extracts the gradient change and edge details in different directions, and then strengthens the high-weight features related to the landslide boundary through the channel attention mechanism, so as to highlight the key boundary and suppress the background noise. With the help of the module, the model has significantly improved discrimination ability for the outline of the landslide area in the complex background, and has realized higher precision recognition of small-scale landslide targets.

[0153] (2) Insufficient landslide target geometry modeling ability

[0154] The formation process of landslide disaster is closely related to the topographic slope and structural deformation, and the existing method often lacks deep modeling ability for spatial geometric properties, leading to misidentification in the topographic transition zone. To solve this problem, the embodiment introduces the VSS module in the DEHB module. The module first normalizes and linearly maps the features to stabilize the distribution, then uses depth separable convolution and nonlinear activation to extract local slope texture, and performs cumulative weighted scanning in the horizontal and vertical directions through the 2D-Scan operator to encode the elevation change pattern into the feature representation. In this way, the network can explicitly perceive the tilting trend and structural change of the topography, has stronger geometric structure semantic modeling and discrimination ability in complex slope conditions, and effectively reduces the false detection rate in the topographic transition zone.

[0155] (3) High-performance model deployment problem under limited computing resources

[0156] The mainstream deep model has large parameter quantity and high computational complexity, which is difficult to be directly deployed on the resource-limited UAV platform, affecting its practicability and response speed in real-time landslide monitoring in the field. To solve this problem, the embodiment proposes a model design idea based on a lightweight network skeleton, and innovatively integrates a direction information enhancement mechanism and a lightweight attention structure. In the overall architecture, the Backbone part effectively reduces redundant calculation through channel shunting and residual reuse of the CSP module, and the DEHB module inside the MVM module and the MSAVM module uses depth separable convolution and virtual slope modeling to reduce FLOPs while enhancing boundary and slope feature expression; the multi-head self-attention mechanism in the MSAVM module uses a parallel grouping method, which significantly reduces the computational overhead; the BHFM module of the Neck part uses sparse convolution through the CCGM module and replaces the large convolution kernel with sequential pooling through the SPPF module, both of which embody the lightweight design idea. Finally, the detection head uses 1×1 convolution compression and multi-scale prediction branches to ensure detection accuracy while reducing inference latency.

[0157] In summary, the embodiment implements the lightweight principle in each stage of feature extraction, context modeling and detection prediction, so that the model significantly reduces the parameter quantity and computational complexity while maintaining high accuracy and robustness, taking into account the real-time and low-power deployment requirements of the UAV platform, and providing a practical technical solution for efficient identification and intelligent monitoring of landslide in the reservoir area.

[0158] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1.A landslide recognition method based on multi-directional edge-gradient information enhancement, characterized by, The method comprises the following steps: Step 1: Obtain the reservoir area image and perform preprocessing; Step 2: Based on a deep learning framework, construct a landslide identification model; Step 3: Train and test the landslide identification model; Step 4: Use the trained landslide identification model to identify landslides; The landslide identification model in step 2 is a three-stage architecture of Backbone-Neck-Head; The Backbone part includes two CBS modules connected in turn, which are used to extract basic texture and edge features of the image and realize preliminary down-sampling; and two CSP modules, which are used to realize multi-stage feature fusion; MVM module for extracting deep features and enhancing the expression of boundary and slope information; MSAVM module, in 32Resolution global semantic modeling and feature alignment are performed for extracting underlying features; The Neck part includes three BHFM modules, which are respectively connected to the two CSP modules and the MVM module of the Backbone part, and are used to decompose the features into high-frequency and low-frequency parts and perform weighted fusion; three MVM modules, which are respectively connected to the three BHFM modules, the first layer MVM module is connected to the second layer MVM module, and the second layer MVM module is connected to the third layer MVM module, which are used to realize multi-scale semantic enhancement; two CCGM modules, the upper layer CCGM module is connected to the first layer MVM module and the lower layer GGM module, and the lower layer CCGM is connected to the second layer and third layer MVM modules, which are used to supplement the context information and fine-grained feature expression; an SPPF module connected to the MSAVM module of the Backbone part, which is used to complete multi-receptive field fusion; an MSAVM module connected to the SPPF module, which realizes feature alignment and enhancement, and is connected to the third layer MVM module through a residual path to keep the main information flowing; The MVM module and the MSAVM module are embedded with DEHB modules, and the MSAVM module fuses a multi-head self-attention mechanism based on the MVM module; The DEHB module includes horizontal depth convolution, vertical depth convolution, feature splicing, channel attention mechanism, 1x1 convolution, and slope simulation module; the feature splicing is used to splice the outputs of the horizontal depth convolution and the vertical depth convolution; the feature splicing, the channel attention mechanism, the 1x1 convolution, and the slope simulation module are connected in turn; the slope simulation module includes a main branch and a residual branch, the main branch includes normalization, linear mapping, depth separable convolution, SiLU activation function, 2D-Scan operator, and normalization, and the residual branch includes linear mapping; after the output of the main branch and the output of the residual branch are added and fused, the linear mapping is performed to generate the final output. 2.The landslide identification method based on multi-directional edge-gradient information enhancement according to claim 1, wherein, Step 1 includes the following steps: Step 1.1: Through the unmanned aerial vehicle platform deployed in the reservoir area, multi-angle aerial photography is performed on typical landslide scenes, high-resolution remote sensing image data is collected, and rich topographic and geomorphic information covering the landslide area is obtained; Step 1.2: Data cleaning, for the image noise and data missing problems existing in the remote sensing image data, multi-frame fusion and interpolation method is used for image quality enhancement processing; Step 1.3: Normalization, the Min-Max normalization method is used to perform linear scale adjustment on the image, and all pixel values are scaled to the [0, 1] interval, and the specific calculation formula is as follows: ； In the formula, is the original image pixel value, , respectively denote the minimum and maximum pixel values in the image; Step 1.4: Constructing the dataset, combining historical landslide remote sensing images, geological disaster survey reports, and field-collected geographic annotation information, manually fine-labeled the landslide areas in the remote sensing images to construct the dataset. 3.The landslide identification method based on multi-directional edge-gradient information enhancement according to claim 1, characterized in that, The CSP module in the Backbone part includes: 1x1 convolution, channel split, residual convolution unit, convolution subnet, cross-layer splicing, and 1x1 convolution; the first layer 1x1 convolution is input into the residual convolution unit and the convolution subnet after channel split; the cross-layer splicing is used to splice the outputs of the residual convolution unit and the convolution subnet, and input into the second layer 1x1 convolution. 4.The landslide identification method based on multi-directional edge-gradient information enhancement according to claim 1, wherein, The MVM module in the Backbone part and the Neck part includes: 1x1 convolution, channel split, two-stage DEHB module, cross-layer splicing, and 1x1 convolution connected in turn; the cross-layer splicing is used to splice the outputs of the first layer 1x1 convolution and the two-stage DEHB module. 5.The landslide identification method based on multi-directional edge-gradient information enhancement according to claim 1, wherein, The MSAVM module in the Backbone part and the Neck part includes: 1x1 convolution, channel split, DEHB module, multi-head self-attention mechanism, cross-layer splicing, and 1x1 convolution connected in turn; the cross-layer splicing is used to splice the outputs of the first layer 1x1 convolution and the DEHB module. 6.The landslide identification method based on multi-directional edge-gradient information enhancement according to claim 1, wherein, The BHFM module in the Neck part includes: multi-scale orthogonal convolution, channel attention, double-flow feature fusion, and normalization. 7.The landslide identification method based on multi-directional edge-gradient information enhancement according to claim 1, wherein, The CCGM module in the Neck part includes: low-resolution feature map processing branch, high-resolution feature map processing branch, and 1x1 convolution; the low-resolution feature map processing branch is processed by depth split convolution, 1x1 convolution, and grid sample operation in turn; the high-resolution feature map processing branch is processed by sparse convolution, 1x1 convolution, Gaussian error linear unit, and 1x1 convolution in turn; the outputs of the low-resolution feature map processing branch and the high-resolution feature map processing branch are added and processed by 1x1 convolution to obtain the final output. 8.The landslide identification method based on multi-directional edge-gradient information enhancement according to claim 1, wherein, The SPPF module in the Neck part includes: 1x1 convolution, three-stage 5x4 max pooling, splicing, and 1x1 convolution connected in turn. 9.The landslide identification method based on multi-directional edge-gradient information enhancement of claim 1, wherein, The Head part uses a three-scale detection head to comprehensively judge the fused features, and outputs the position coordinates, category information, and confidence score of the landslide target. 10.The landslide identification method based on multi-directional edge-gradient information enhancement according to claim 1, wherein, In the model training and testing process of step 2, the loss function is composed of three parts: bounding box regression loss, classification loss, and confidence loss; the bounding box regression loss uses the generalized intersection over union loss function, the formula is: ； In the formula, is a prediction box, is a real label box, and the spatial overlap and distance between the two are calculated to improve positioning accuracy; The classification loss uses the cross-entropy loss function, the formula is: ； In the formula, is the true class label, is the predicted class probability, and the loss function improves the ability to distinguish landslide types by measuring the difference between the predicted probability and the true label. The confidence loss uses the binary cross-entropy loss function, the formula is: ； In the formula, t represents the landslide / background label, and p is the predicted confidence, which is used to suppress background false positives and improve detection reliability; The comprehensive loss function is defined as: ； In the formula, = 1.0, = 1.0, = 0.5, the loss value of a single batch of images is fed back to each layer of the network through back propagation to dynamically adjust the parameters and improve the prediction accuracy. In the optimization strategy, firstly, the adaptive momentum optimizer is selected, the initial learning rate is set to , and the weight decay coefficient is 0.0005; secondly, the cosine annealing learning rate scheduling mechanism is adopted, and the learning rate is smoothly decayed from the initial value to close to zero within 150 training rounds. During training, random augmentation operations are applied to the input batch of images, including cropping, scaling, flipping, and brightness adjustment, to improve the robustness of the model in complex scenarios; Finally, when the average precision mean of the validation set does not show significant improvement for 5 consecutive rounds, the early stopping mechanism is triggered and the current optimal model parameters are saved to avoid overfitting and shorten the training time.

Images