Unmanned aerial vehicle image collapse intelligent identification method based on coordinate attention mechanism

Abstract

The invention relates to the technical field of collapse hill identification methods, in particular to an unmanned aerial vehicle image collapse hill intelligent identification method based on a coordinate attention mechanism. The system comprises an unmanned aerial vehicle image acquisition and preprocessing module, a CA-Unet model construction module, a coordinate attention calculation module and an intelligent collapse hill identification output module. According to the method, a high-resolution image and laser radar point cloud data are collected through an unmanned aerial vehicle, a digital orthoimage and a digital elevation model are generated through orthorectification, and a deep learning data set is constructed; a coordinate attention mechanism is embedded in a U-Net model, the recognition capability of long-distance gullies and steep edges of collapse hills is enhanced through feature coding and dynamic weighting in the horizontal and vertical directions, direction-sensitive feature enhancement is achieved through a coordinate attention calculation module, and finally collapse hill boundary vectors and morphological parameters are output. According to the method, the limitation of a traditional method in the aspects of spatial dependence modeling and boundary extraction is broken through, and accurate recognition and surveying and mapping of the collapse slope erosion landform are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of methods for identifying collapsed ridges, and in particular to an intelligent method for identifying collapsed ridges using drone images based on a coordinate attention mechanism. Background Art

[0002] Collapsed gullies, a prominent and widespread erosional landform in southern my country's granite regions, pose a serious threat to regional ecological security and sustainable development with their dense gullies and steep collapse walls. For a long time, identifying collapsed gullies relied primarily on time-consuming and labor-intensive field surveys and visual interpretation of remote sensing imagery, which were inadequate for large-scale, high-frequency dynamic monitoring. In recent years, the maturity of drone remote sensing technology has made it possible to obtain sub-meter and even centimeter-level high-resolution imagery. This, combined with semantic segmentation models such as U-Net, can be used to identify geological hazards such as landslides and debris flows.

[0003] Although existing technologies have achieved automated identification of landslides to a certain extent, they still have obvious shortcomings in capturing the long-range spatial dependencies and direction-sensitive features of the complex morphology of landslides. Specifically, although traditional semantic segmentation models such as U-Net have the ability to fuse multi-scale features, their core convolution operations are limited in modeling long-range pixel dependencies, making it difficult to effectively capture spatial features in the horizontal and vertical directions. This makes it difficult for the model to accurately identify the narrow gullies, steep collapse walls and irregular boundaries of landslides. Summary of the Invention

[0004] The present invention provides an intelligent method for identifying hillock collapse in drone images based on a coordinate attention mechanism. By embedding the coordinate attention mechanism in a U-Net decoder, feature encoding and dynamic weighting are performed along the horizontal and vertical directions respectively, thereby solving the problems raised in the above background technology, namely: Traditional semantic segmentation models (such as U-Net) have problems in capturing long-distance spatial dependencies and extracting fine boundary information when identifying erosion landforms such as landslides in drone images.

[0005] To achieve the above purpose, the method for intelligently identifying collapsed hillocks includes the following steps: S1. Collect aerial image data and LiDAR point cloud data, generate high-resolution digital orthophotos and digital elevation models through orthorectification, and construct a deep learning collapsing dataset; S2. Based on the deep learning Benggang dataset, train a U-Net semantic segmentation model embedded with a coordinate attention mechanism; S3, input the high-resolution digital orthophoto into the U-Net semantic segmentation model, and extract the multi-scale feature pyramid from the high-resolution digital orthophoto through the encoder; S4, fuse the multi-scale feature pyramids in the decoder feature splicing layer and output the spliced ​​fusion feature tensor; S5. Perform one-dimensional global average pooling on the spliced ​​fusion feature tensor in the horizontal direction and the vertical direction respectively to generate horizontal feature coding and vertical feature coding; S6. Process the horizontal feature code and the vertical feature code through nonlinear transformation, and output a horizontal attention weight map and a vertical attention weight map; S7. Perform element-wise multiplication on the horizontal attention weight map, the vertical attention weight map, and the spliced ​​fusion feature tensor, and output a binary segmentation mask map after iterative decoding to generate the gully erosion boundary vector data.

[0006] This technical solution, by embedding a coordinate attention mechanism into the feature concatenation layer of the U-Net decoder, specifically addresses the structural limitations of traditional convolutional neural networks in landslide detection. Using only the standard U-Net architecture, the local receptive field of its convolution kernels fails to model the long-range gully continuity characteristic of landslides (poor horizontal feature correlation), and the isotropic convolution operation cannot distinguish the vertical spatial features of the landslide edge (resulting in blurred boundaries). Simply introducing a conventional attention mechanism (such as the SE module) improves feature response strength but loses critical spatial directional information. This solution innovatively decomposes spatial features using a horizontal / vertical dual-path one-dimensional pooling approach. This allows the model to establish long-range dependencies along the gully extension direction (horizontally), preserving the continuity of the banded structure. Simultaneously, it strengthens the edge gradient response along the steepness direction (vertically), achieving sub-pixel boundary localization. This directionally decoupled feature enhancement approach, while retaining the multi-scale fusion advantages of the U-Net, overcomes the bottleneck of traditional methods in representing the anisotropic characteristics of erosional landforms.

[0007] On this basis, the UAV image acquisition and preprocessing module packages the high-resolution digital orthophotos, digital elevation models and deep learning collapsing gully datasets into a collapsing gully multi-source training data package and passes it to the CA-Unet model construction module.

[0008] In another technical solution, the CA-Unet model construction module adopts a binary cross entropy loss function, and uses high-resolution digital orthophotos in the deep learning collapsing ridge dataset as input and semantic segmentation labels as supervision signals for model training.

[0009] This technical solution, by constructing a multi-source training data package for collapsing gullies, achieves the systematic integration and efficient transfer of multi-source data. Relying solely on DOM imagery, the model lacks the ability to verify the three-dimensional morphology of the collapsing gullies using terrain auxiliary information (DEM), leading to annotation errors being passed on to the training phase. Furthermore, without standardized data packaging, the asynchronous transmission of heterogeneous multi-source data (imagery / point cloud / labels) compromises the integrity of the training process. The CA-Unet model construction module utilizes a binary cross-entropy loss function to specifically address the specific requirements of the collapsing gully binary segmentation task. Using a multi-class cross-entropy loss introduces redundant class noise (e.g., in the case of a simple classification of collapsing gullies / background), while using regional matching metrics like the Dice loss makes it difficult to handle the extreme class imbalance of collapsing gullies. This data-model collaborative optimization design ensures the effective integration of multi-source geographic information while adapting to the mathematical properties of erosional landform segmentation.

[0010] Compared with the prior art, the present invention has the following beneficial effects: 1. This study, using sub-meter drone remote sensing imagery as a data source, leverages the CA mechanism within the Unet model to effectively enhance the directional sensitivity of the complex spatial structure of landslides and gullies and its ability to capture long-range spatial dependencies. This results in higher accuracy landslide identification, a smaller range of misidentifications of landslides, and superior performance in segmenting landslide boundaries. The research findings are expected to provide a new technical approach for landslide monitoring, offering a solid theoretical basis and key technical support for landslide identification and extraction, ecological and environmental restoration, and disaster risk prevention and control.

[0011] 2. This invention creatively constructs a full-chain technology system of "image acquisition-intelligent recognition-engineering application", and realizes seamless connection from raw data to governance decision-making through multi-module collaborative optimization. The dynamic computing characteristics of the coordinate attention mechanism enable the model to have the ability to adaptively process complex terrain, and the standardized data interface design ensures lossless transmission of centimeter-level precision, which significantly improves the reliability of the results compared with the traditional staged processing method. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 It is a schematic diagram of the overall process structure of the present invention; Figure 3 A schematic diagram of the data set annotation process of the present invention; Figure 4 Schematic diagram of the CA-Unet structure of the present invention.

[0013] The meaning of each number in the figure is: 100. UAV image acquisition and preprocessing module; 200. CA-Unet model construction module; 300. Coordinate attention calculation module; 400. Collapse intelligent identification and output module. DETAILED DESCRIPTION

[0014] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0015] Currently, for erosion landforms with complex shapes and blurred boundaries, traditional semantic segmentation models (such as U-Net) have problems in capturing long-range spatial dependencies and extracting fine boundary information in drone image recognition. This paper provides an intelligent recognition method for erosion landforms in drone images based on the coordinate attention mechanism. Figure 1 As shown, the following steps are included: S1. Collect aerial image data and LiDAR point cloud data, generate high-resolution digital orthophotos and digital elevation models through orthorectification, and construct a deep learning collapsing dataset; S2. Based on the deep learning Benggang dataset, train a U-Net semantic segmentation model embedded with a coordinate attention mechanism; S3, input the high-resolution digital orthophoto into the U-Net semantic segmentation model, and extract the multi-scale feature pyramid from the high-resolution digital orthophoto through the encoder; S4, fuse the multi-scale feature pyramids in the decoder feature splicing layer and output the spliced ​​fusion feature tensor; S5. Perform one-dimensional global average pooling on the spliced ​​fusion feature tensor in the horizontal direction and the vertical direction respectively to generate horizontal feature coding and vertical feature coding; S6. Process the horizontal feature code and the vertical feature code through nonlinear transformation, and output a horizontal attention weight map and a vertical attention weight map; S7. Perform element-wise multiplication on the horizontal attention weight map, the vertical attention weight map, and the spliced ​​fusion feature tensor, and output a binary segmentation mask map after iterative decoding to generate the gully erosion boundary vector data.

[0016] See also Figure 2As shown, it specifically includes a UAV image acquisition and preprocessing module 100, a CA-Unet model construction module 200, a coordinate attention calculation module 300 and a collapse gully intelligent recognition output module 400. Through the collaborative work of multiple modules, the automated and high-precision semantic segmentation and recognition of collapse gully erosion landforms in high-resolution UAV images is achieved.

[0017] The drone image acquisition and pre-processing module 100 is the fundamental data input component of the entire system. This module uses a DJI M350RTK drone equipped with a Zenmuse L2 LiDAR system and a 20-megapixel visible-light camera for aerial photography, along with a Qianxun Star SE handheld RTK device for precise positioning. Aerial photography is completed within a specific timeframe, using the CGCS2000 coordinate system (with the central meridian at 117°) and the 1985 National Elevation Datum. This process directly captures raw, high-resolution drone aerial imagery and LiDAR point cloud data.

[0018] During implementation, the aerial photography mission was carried out according to the pre-defined route parameters. The flight altitude was set at 150 meters relative to the ground, the speed was 13 meters per second, and a five-pass scan pattern was used with a field of view of 70° × 15.5°. The heading overlap and lateral overlap were set to 70% and 60%, respectively. The aerial photography covered the entire survey area, completed over multiple flights, ultimately acquiring a large amount of high-resolution imagery. This imagery data had a specific ground sampling distance and resolution.

[0019] like Figure 3 As shown, the acquired raw image data was imported into DJI Map software for processing to generate an orthophoto (DOM) of the survey area. The processed DOM image was cropped to its boundaries, resulting in a true-color image containing red, green, and blue bands, achieving a specific high resolution (0.04 meters). Simultaneously, the LiDAR point cloud data was processed in DJI Map and LiDAR360 software through point cloud denoising, filtering, and classification (combined with manual interpretation and correction). This ultimately yielded ground point cloud data for the study area. This data was then used to construct a high-precision digital elevation model (DEM) with a specific resolution (0.5 meters). The coordinate system used was the CGCS2000 Gaussian projection, and the elevation datum was the 1985 National Elevation Datum. The DOM and DEM are the core preprocessing results of this module.

[0020] At the same time, the module also evenly distributes a specific number of ground control points throughout the survey area in accordance with relevant standards (GB / T13977-2012). These control points provide a guarantee for subsequent geometric accuracy control.

[0021] The data annotation phase primarily relied on preprocessed high-resolution digital orthophotos, supplemented by a 3D surface model and digital elevation model for comprehensive assessment. Professional interpreters then performed manual interpretation and annotation to demarcate the collapsing gullies. This ultimately resulted in a deep learning collapsing gully dataset, consisting of multiple sets of samples, each consisting of DOM image blocks of a specific size and their corresponding semantic segmentation label files. These sets were divided into training and test sets according to specific ratios.

[0022] The drone image acquisition and preprocessing module 100 ultimately generates a multi-source training data package for collapsing gullies and passes it to the CA-Unet model construction module 200. This data package, constructed strictly in accordance with aerial photography specifications, comprises three core components: high-resolution digital orthophotos, a digital elevation model, and a deep learning collapsing gullies dataset. The deep learning collapsing gullies dataset is an annotated dataset constructed through professional interpretation of DOM imagery. It consists of 512×512 pixel image sample blocks and their corresponding semantic segmentation label files (PNG format), divided into training and test sets at a standard 9:1 ratio. This rigorously quality-controlled data serves as the fundamental input for subsequent CA-Unet model training and optimization, ensuring that the model fully learns the multi-scale characteristics of collapsing gullies.

[0023] After receiving the multi-source training data package for collapsing ridges, the CA-Unet model construction module 200 performs model construction and training. Based on an improved U-Net architecture, this module uses the high-resolution digital orthophotos (DOM) in the data package as input and implements end-to-end collapsing ridge feature learning through a deep learning framework. The entire model strictly adheres to the encoder-decoder paradigm: the encoder, consisting of multiple convolutional and downsampling units, progressively extracts spatial semantic features from the image; the decoder gradually restores the feature map resolution through upsampling operations, while skip connections fuse feature information from different encoder scales.

[0024] The core innovation of the CA-Unet model building block 200 lies in the coordinate attention mechanism embedded in the decoder path. This mechanism is precisely located at the output of the feature concatenation layer (Concat layer) to spatially enhance the fused features. Its operation includes the following key steps: horizontally performing one-dimensional global pooling to generate horizontal feature encodings; vertically performing one-dimensional global pooling to generate vertical feature encodings; and finally, converting the bidirectional encodings into a spatial attention weight map, which is then element-wise multiplied with the original feature map to achieve feature reweighting.

[0025] A directionally sensitive feature enhancement mechanism significantly improves the model's landform recognition capabilities: horizontal weighting enhances the capture of the continuous structure of narrow gullies and ravines, while vertical weighting enhances sensitivity to the edges of steep collapse walls. This design inherits the multi-scale advantages of U-Net while specifically overcoming the shortcomings of traditional convolutional networks in long-range modeling and capturing direction-dependent features.

[0026] Model training strictly utilizes the deep learning collapsing hill dataset included in the data package. The training process uses 512×512 pixel DOM image patches as input, with corresponding binary semantic segmentation labels as supervision signals. Weight parameters are optimized via backpropagation using a binary cross-entropy loss function. Training is performed on a GPU computing platform, with convergence controlled by a dynamic learning rate adjustment strategy.

[0027] The CA-Unet model construction module 200 module finally generates a fully encapsulated CA-Unet prediction model. Figure 4 As shown in the figure, the model contains a complete network architecture and training parameters, which are passed to subsequent modules as the core reasoning engine for the actual collapse intelligent identification task.

[0028] While the CA-Unet model construction module 200 has successfully established an improved network architecture embedded with a coordinate attention mechanism and completed model training based on high-quality annotated data, resulting in a prediction model with optimized weight parameters, the static model output by this module does not address the critical requirement of dynamic feature enhancement during the actual inference phase of intelligent landslide identification: the coordinate attention weights within the model depend on the specific content of the input image and must be calculated and generated in real time during the inference process. This dynamic computational capability cannot be fixed during the model construction phase and requires dedicated computational units. To address this, we introduced the coordinate attention computation module 300, which is responsible for dynamically performing direction-aware feature weighting operations during model inference to address the fundamental limitation of static models: their inability to adaptively handle diverse landslide landforms.

[0029] The Coordinate Attention Calculation Module 300 receives the pretrained CA-Unet model (including fixed weight parameters) from the CA-Unet Model Construction Module 200 and the high-resolution digital orthophoto (DOM) image of the drone to be identified, and then initiates the inference process for collapsing hillock identification. This module first processes the input DOM image through the encoder path of the loaded model. The encoder performs multiple layers of convolution and downsampling operations, gradually extracting spatial features from the image and generating multi-scale feature pyramids, from shallow details to deep semantics. These feature pyramids lay the foundation for refined reconstruction in the subsequent decoding stage.

[0030] Based on the multi-scale feature pyramid generated by the encoder, the coordinate attention calculation module 300 then transitions to the decoder path to perform core enhancement operations. The multi-scale feature pyramid is fused in the decoder's feature concatenation layer, outputting a concatenated fused feature tensor. As the spatial resolution of the features is gradually restored through the upsampling layers, the dynamic coordinate attention calculation mechanism is triggered immediately after feature skip connections and concatenation of shallow and deep features (the Concat layer) are completed. This mechanism operates in a seamless manner: One-dimensional global feature compression is performed horizontally on the concatenated and fused feature tensor to generate a horizontal spatial correlation code. Simultaneously, the same compression operation is performed vertically to generate a vertical spatial correlation code. The two directional codes are then processed through a shared 1×1 convolutional layer with a sigmoid activation function, outputting the corresponding directional attention weight maps (horizontal and vertical).

[0031] The coordinate attention calculation module 300 then uses the generated directional attention weight map to perform real-time feature recalibration. The original spliced ​​feature map is element-wise multiplied with the horizontal and vertical weight maps, dynamically weighting the feature values ​​based on their spatial position and direction: horizontal weights enhance the continuity of narrow gullies, while vertical weights increase the edge sensitivity of steep collapse walls. This dynamically weighted output serves as the enhanced feature map and is passed to the next decoding layer, where the "upsampling-feature splicing-coordinate attention weighting" process is repeated until the original image resolution is restored.

[0032] After completing multi-level decoding enhancement, the output of the coordinate attention calculation module 300 generates a semantic prediction map. A sigmoid activation function is applied to the final high-resolution feature map to calculate the probability of each pixel belonging to a landslide area. A fixed threshold (0.5) is set for binary decision making, converting the probability map into a binary segmentation mask consisting of 0 (background) and 1 (collapse), which accurately annotates the spatial distribution of landslides in the image.

[0033] As a direct product of the dynamic calculation of the coordinate attention mechanism, this binary segmentation mask map specifically addresses two major defects in the traditional model's extraction of key features of landslide landforms: by strengthening the continuity of the narrow gully structure through horizontal weights, the strip-like extension morphology of the main area of ​​the landslide is effectively restored; by precisely sharpening the edges of steep collapse walls through vertical weights, the positioning accuracy of complex boundaries is significantly improved; and ultimately, the coordinated optimization of directional features and structural integrity is achieved, overcoming the inherent limitations of the original recognition method in long-distance dependency modeling and fine boundary expression.

[0034] The coordinate attention calculation module 300 transmits the binary segmentation mask image to the gully intelligent identification and output module 400, which performs subsequent boundary vectorization and statistical analysis. Through a fully closed-loop process, this module achieves adaptive enhancement and precise segmentation of the directional characteristics of the gully landform.

[0035] After receiving the binary segmentation mask from the coordinate attention calculation module 300, the intelligent gully identification and output module 400 initiates the results transformation process. This module first performs topological analysis on the connected domains within the mask to identify the contours of independent gully areas. It then uses a boundary smoothing algorithm to minimize jagged edges and employs vector processing techniques to generate a vector graphic of the gully erosion boundary that complies with geographic information standards. This process rigorously preserves the primary structural features of the gully head, walls, and bottom of the gully, ensuring the integrity of the spatial topology.

[0036] Next, the intelligent collapsing gully identification and output module 400 combines the georeferenced information from the original drone digital orthophoto imagery to perform spatial coordinate system conversion. This converts the vector boundary data from the image pixel coordinate system to the standard geodetic coordinate system, assigning it precise geographic location attributes. Simultaneously, based on spatial analysis algorithms, it automatically calculates morphological parameters such as the area, perimeter, and gully density of each collapsing gully patch, constructing a complete attribute data table and enabling the coordinated management of spatial entities and attribute characteristics.

[0037] Subsequently, multi-dimensional results verification was carried out, and key accuracy evaluation indicators were calculated through spatial overlay analysis with the true value labeled data in the preprocessing stage. Based on the preset quality control threshold, low-confidence recognition results were automatically screened out and a quality control report was generated to ensure that the output results met the technical standards requirements of soil and water conservation projects.

[0038] The Collapse and Gulch Intelligent Identification and Output Module 400 ultimately produces a three-in-one engineering output: a spatially distributed vector map of collapse and gully erosion boundaries, a table of morphological characteristic parameters, and an accuracy verification report. These results are integrated into a thematic map for collapse and gully intelligent monitoring through a geographic information system and simultaneously uploaded to the soil and water conservation intelligent management cloud platform, forming a dynamically updated erosion landform database.

[0039] By introducing a direction-aware dynamic feature weighting mechanism, namely, coordinate attention calculation that independently performs feature compression and recalibration in the horizontal and vertical directions, the present invention solves the problems of traditional semantic segmentation models' insufficient ability to capture long-distance spatial dependencies and difficulty in extracting fine boundary information in UAV image recognition. It realizes an integrated air-space-ground monitoring system from UAV raw image acquisition and intelligent feature recognition to the output of application-level results for soil and water conservation projects, providing accurate and reliable technical support for the prevention and control of gully erosion.

[0040] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent method for identifying hillock collapse in drone images based on the coordinate attention mechanism, characterized by: The following steps are involved: S1. Collect aerial image data and LiDAR point cloud data, generate high-resolution digital orthophotos and digital elevation models through orthorectification, and construct a deep learning collapsing dataset; S2. Based on the deep learning Benggang dataset, train a U-Net semantic segmentation model embedded with a coordinate attention mechanism; S3, input the high-resolution digital orthophoto into the U-Net semantic segmentation model, and extract the multi-scale feature pyramid from the high-resolution digital orthophoto through the encoder; S4, fuse the multi-scale feature pyramids in the decoder feature splicing layer and output the spliced ​​fusion feature tensor; S5. Perform one-dimensional global average pooling on the spliced ​​fusion feature tensor in the horizontal direction and the vertical direction respectively to generate horizontal feature coding and vertical feature coding; S6. Process the horizontal feature code and the vertical feature code through nonlinear transformation, and output a horizontal attention weight map and a vertical attention weight map; S7. Perform element-wise multiplication on the horizontal attention weight map, the vertical attention weight map, and the spliced ​​fusion feature tensor, and output a binary segmentation mask map after iterative decoding to generate the gully erosion boundary vector data.

2. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 1 is characterized by: The drone image acquisition and preprocessing module (100) uses a drone equipped with a laser radar system and a camera to acquire the aerial image data and the laser radar point cloud data.

3. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 1 is characterized by: The drone image acquisition and preprocessing module (100) manually interprets and annotates the high-resolution digital orthophotos to construct the deep learning collapse dataset.

4. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 1 is characterized by: The UAV image acquisition and preprocessing module (100) packages the high-resolution digital orthophoto, digital elevation model and deep learning collapsing gully dataset into a collapsing gully multi-source training data package and passes it to the CA-Unet model construction module (200).

5. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 4 is characterized in that: The CA-Unet model construction module (200) adopts a binary cross entropy loss function and uses high-resolution digital orthophotos in a deep learning collapse dataset as input and semantic segmentation labels as supervisory signals for model training.

6. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 5 is characterized by: The CA-Unet model construction module (200) controls the convergence process through a dynamic learning rate adjustment strategy, generates a CA-Unet prediction model, and passes it to the coordinate attention calculation module (300).

7. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 6 is characterized by: The coordinate attention calculation module (300) processes the horizontal feature code and the vertical feature code through a 1×1 convolution layer and a Sigmoid activation function respectively to generate the horizontal direction attention weight map and the vertical direction attention weight map.

8. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 7 is characterized by: The coordinate attention calculation module (300) transmits the binary segmentation mask image to the collapse intelligent identification output module (400) for performing subsequent boundary vectorization and statistical analysis tasks.

9. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 8, characterized in that: The collapse gully intelligent identification and output module (400) converts the binary segmentation mask image into collapse gully erosion boundary vector data and matches it to the CGCS2000 geographic coordinate system.

10. The method for intelligently identifying hillock collapse in drone images based on a coordinate attention mechanism according to claim 1, characterized in that: The collapse ridge intelligent identification and output module (400) calculates collapse ridge morphological parameters based on the boundary vector data, and outputs an engineering application package containing boundary vectors and characteristic parameters.

Images