Landslide instance segmentation method fusing self-supervised visual features and topographic prior

Abstract

The application discloses a landslide instance segmentation method fusing self-supervised visual features and topographic prior, and belongs to the technical field of image processing, and comprises the following steps: S1, constructing a multi-source landslide training dataset; S2, performing self-supervised field adaptation training on a pre-trained visual basic model to obtain a landslide field feature extraction backbone network; S3, constructing a landslide semantic segmentation model, and extracting a landslide semantic probability graph and multi-scale landslide semantic features; S4, constructing a landslide geographic relation graph; S5, inputting landslide instance features into an instance decoder for training to obtain a landslide instance segmentation model; and S6, inputting the trained landslide instance segmentation model for prediction, and obtaining monomer landslide instance recognition results with independent instance identification through post-processing. The application effectively solves the problems of adjacent landslide adhesion and internal fragmentation of the same landslide, and significantly improves the boundary precision, spatial continuity and cataloging practicability of monomer landslide instance segmentation in a complex mountainous scene.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a landslide instance segmentation method that integrates self-supervised visual features and topographic priors. Background Technology

[0002] Landslides are common and highly destructive geological hazards in mountainous areas, potentially triggering secondary disasters such as landslide-dammed lakes, debris flows, and slope instability. Rapidly and accurately identifying the spatial location, boundary extent, and number of individual landslides is crucial for landslide surveys, geological hazard monitoring, engineering investigations, risk assessments, and emergency response. With the development of high-resolution satellite remote sensing, UAV photogrammetry, terrestrial photography, and mobile terminal image acquisition technologies, landslide identification data sources have expanded from single satellite images to multi-source data including satellite remote sensing images, UAV orthophotos, single UAV images, terrestrial camera photos, and smartphone photos. Images acquired by different sensors exhibit significant differences in spatial resolution, imaging angle, lighting conditions, color or spectral response, image scale, land cover characteristics, and spatial reference information. Traditional landslide identification methods, which rely on manually designed color, texture, shape, and topographic features, are typically sensitive to image quality and data source, making it difficult to simultaneously adapt to landslide identification tasks across different sensors, observation scales, and geomorphic environments.

[0003] In recent years, convolutional neural networks and basic vision models have been increasingly applied to landslide classification, object detection, and semantic segmentation. Landslide classification methods can determine whether an image or image patch contains a landslide, but they cannot provide the precise spatial boundaries of a landslide. Object detection methods typically use bounding boxes to mark landslide locations, which are difficult to describe the irregular contours of landslides. Semantic segmentation methods can classify each pixel as either a landslide or non-landslide, but they usually represent all landslide pixels of the same category as a single category, failing to directly distinguish between different individual landslides in an image. In areas with dense landslide distribution, areas with multiple landslides, and mountainous areas with strong topographical dissections, adjacent landslides may be spatially close to each other, have boundary contact, or even partially overlap. Ordinary semantic segmentation models tend to merge multiple adjacent landslides into a single continuous region, resulting in contiguous landslide instances. At the same time, the same landslide may be segmented into multiple discontinuous patches due to differences in vegetation cover, shadow, degree of exposure, material composition, and image features of the source area, transport area, and deposition area, resulting in fragmented instances. The aforementioned problems can lead to errors in the statistical analysis of landslide numbers, deviations in individual landslide boundaries, and inaccurate area calculations, making it difficult to meet the requirements of landslide cataloging and professional geomorphological mapping for the extraction of individual landslides.

[0004] Existing general instance segmentation methods typically distinguish instances based on color, texture, edges, and target shape in images, with their network structures and training mechanisms primarily geared towards natural scene targets. Landslides, as surface geological objects controlled by topographic conditions, have spatial morphology and distribution that are not only related to image semantic features but also closely linked to topographic features such as elevation, slope, aspect, curvature, topographic relief, valleys, watersheds, and slope units. Candidate regions located on the same slope and with similar topographic environments may belong to the same landslide instance; conversely, candidate regions separated by watersheds, valleys, or significant aspect changes may belong to different landslide instances, even if their image colors and textures are similar. Relying solely on image features for instance decoding makes it difficult to fully utilize the aforementioned topographic constraints.

[0005] General-purpose visual models are typically pre-trained on large-scale natural imagery, and the learned features can represent general targets and scenes. However, there are still domain differences between these models and the exposed surfaces, vegetation destruction, steep slopes, deposits, gullies, and complex mountainous backgrounds in landslide imagery. If general pre-trained weights are directly used for landslide segmentation, the model may lack stable feature representations for differences in imaging from different sensors, variations in landslide scale, and complex background objects. Existing methods typically use a limited number of labeled landslide samples for supervised fine-tuning of the model, failing to fully utilize the large number of unlabeled or weakly labeled landslide-related images, thus making them susceptible to insufficient labeled samples and data distribution bias. Self-supervised learning, on the other hand, can achieve multi-view... Figure 1Landslide imagery can be learned through methods such as consistency, teacher-student self-distillation, or masked image patch prediction to learn object-level and local structural features. Self-supervised domain adaptation of pre-trained visual models using multi-source landslide imagery allows the model to further learn the color or spectral response, texture structure, morphological boundaries, and scene composition features of landslide objects and non-landslide background objects under different sensors, scales, and observation conditions. However, existing intelligent landslide recognition processes typically implement self-supervised feature learning, semantic segmentation, and instance segmentation separately, lacking a unified approach to jointly model landslide semantic features, topographic priors, and spatial adjacency relationships. Furthermore, complex geographical relationships exist between different spatial units within landslide candidate regions. Traditional convolutional or attention operations primarily propagate information on regular pixel grids, making it difficult to explicitly describe spatial adjacency, slope connections, similarity in geographical environment configuration, and topographic barriers between candidate regions. Graph neural networks can represent candidate spatial units as graph nodes and describe the relationships between nodes through graph edges, providing a foundation for fusing image features and geospatial relationships. However, how to generate graph nodes suitable for instance separation based on landslide semantic probability graphs, how to construct node attributes that simultaneously include landslide semantic features, topographic features, and spatial morphological features, and how to determine the graph edges and edge weights between nodes based on geographical environment configuration similarity remain key issues that need to be addressed in landslide instance segmentation.

[0006] Chinese patent application document with publication number CN121259662A and publication date of January 2, 2026 discloses a landslide image instance segmentation method based on a dual adaptation mechanism, including: collecting multi-source landslide image data and processing the multi-source landslide image data; constructing a landslide image instance segmentation model based on a multi-scale state alignment mechanism and training it using the processed multi-source landslide image data; evaluating the trained landslide image instance segmentation model for subsequent landslide image instance segmentation.

[0007] The landslide image instance segmentation method disclosed in this patent application is based on a dual-adaptation mechanism. The designed model is based on a multi-scale state alignment mechanism. During the forward pass of the model, feature information within the same level is fused using exponential moving average, and feature information from different levels is transferred and fused, enhancing feature robustness and reducing error accumulation caused by local biases. However, relying on multi-source data alignment and meta-adaptive fine-tuning, it lacks explicit modeling of prior topography and spatial adjacency relationships, making it difficult to effectively solve the instance segmentation problem of adjacent landslide adhesion and fragmentation within the same landslide in complex mountainous scenarios. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a landslide instance segmentation method that integrates self-supervised visual features and topographic priors. Through self-supervised domain adaptation training, this invention makes full use of a large number of unlabeled multi-source landslide images, reducing the model's dependence on manually labeled samples. At the same time, it uses graph neural networks for geographic constraint feature propagation and instance decoding, effectively solving the problems of adjacent landslide adhesion and fragmentation within the same landslide. This significantly improves the boundary accuracy, spatial continuity, and cataloging practicality of single landslide instance segmentation in complex mountainous scenes.

[0009] This invention is achieved through the following technical solution: A landslide instance segmentation method integrating self-supervised visual features and topographic priors includes the following steps: S1. Acquire multi-source landslide images, topographic data corresponding to the spatial distribution of the multi-source landslide images, and landslide annotation data. After spatial registration, preprocessing, and sample segmentation, construct a multi-source landslide training dataset. S2. Using unlabeled image samples from the multi-source landslide training dataset, perform self-supervised domain adaptation training on the pre-trained visual basic model to obtain the backbone network for landslide domain feature extraction. S3. Connect the semantic segmentation decoder to the feature extraction backbone network in the landslide domain, use the semantic segmentation training samples for supervised transfer training, construct the landslide semantic segmentation model, and extract the landslide semantic probability map and multi-scale landslide semantic features. S4. Generate candidate landslide areas based on the landslide semantic probability map, divide the candidate landslide areas into spatial units and use them as graph nodes, integrate the landslide semantic features, topographic features and spatial morphological features of the graph nodes, and construct a landslide geographical relationship map based on spatial adjacency and similarity of geographical environment configuration. S5. Input the landslide geographic relationship map into the graph neural network for feature propagation to obtain landslide instance features constrained by geographic relationships. Then, input the landslide instance features into the instance decoder for training to obtain the landslide instance segmentation model. S6. Obtain image data and topographic data of the area to be identified, input them into the trained landslide instance segmentation model for prediction, and obtain the identification results of individual landslide instances with independent instance identifiers through post-processing.

[0010] The multi-source landslide training dataset includes landslide-related images from ground cameras, smartphones, drones, and satellite remote sensing sensors, as well as digital elevation models and landslide annotation data corresponding to the image space.

[0011] In S2, the self-supervised domain adaptation training includes multi-view enhancement of unlabeled image samples. Through feature consistency constraints between teacher and student networks, masked image patch learning, and cross-view local consistency constraints, the pre-trained visual base model is adapted to the feature distribution of multi-source landslide images.

[0012] In S3, the landslide semantic segmentation model uses a self-supervised domain-adapted landslide domain feature extraction backbone network as an encoder, and outputs a landslide semantic probability map through a semantic segmentation decoder. The multi-scale landslide semantic features are used as image semantic input for the subsequent construction of landslide geographic relationship maps.

[0013] In S4, the graph nodes of the landslide geographic relationship map include landslide semantic features obtained by semantic probability weighted aggregation, topographic features calculated by digital elevation model, and spatial morphological features of candidate spatial units.

[0014] Furthermore, graph edges are established based on the boundary contact relationship, centroid spatial distance, and nearest neighbor relationship between graph nodes. The weights of the graph edges are determined based on the similarity of geographical environment configuration, the semantic similarity of landslides, the spatial proximity similarity, and the terrain barrier relationship.

[0015] The graph neural network performs message passing and feature aggregation between adjacent graph nodes based on graph edge weights, which enhances the same instance representation of graph nodes with continuous terrain and similar semantics, and suppresses erroneous feature mixing across valleys, watersheds, or slope units.

[0016] In S5, the instance decoder generates an instance query vector based on the landslide instance features output by the graph neural network, and outputs the predicted instance category and the predicted instance mask. During training, at least one of the following is constructed by the matching relationship between the predicted instance and the real landslide instance: instance classification loss, mask loss, region overlap loss, boundary constraint loss and graph node instance relationship loss.

[0017] In step S6, the post-processing includes at least one of confidence filtering, overlapping instance resolution, connected component correction, hole filling, boundary smoothing, and vectorization.

[0018] In step S6, the identification results of individual landslide instances include instance number, spatial range, area, and confidence level statistics.

[0019] The beneficial effects of this invention are mainly reflected in the following aspects: 1. This invention, through self-supervised domain adaptation training, makes full use of a large number of unlabeled multi-source landslide images, reducing the model's dependence on manually labeled samples. At the same time, it uses graph neural networks for geographic constraint feature propagation and instance decoding, effectively solving the problems of adjacent landslide adhesion and fragmentation within the same landslide, and significantly improving the boundary accuracy, spatial continuity and cataloging practicality of single landslide instance segmentation in complex mountainous scenes.

[0020] 2. This invention combines multi-view enhancement, mask image patch learning, and cross-view local consistency constraints in self-supervised learning, which effectively improves the model's discriminative visual representation ability of landslide objects and background environment in complex mountainous backgrounds.

[0021] 3. This invention constructs a landslide semantic segmentation model through supervised transfer training, and extracts multi-scale semantic features while obtaining a high-precision landslide semantic probability map, providing a reliable image semantic input basis for subsequent candidate region selection and spatial unit division.

[0022] 4. This invention divides candidate landslide areas into spatial units and integrates topographic features to construct a landslide geographic relationship map, enabling explicit modeling of the spatial continuity, slope integrity, and topographic separation relationships of landslides, thereby enhancing the stability of semantic expression within landslides under complex topographic conditions.

[0023] 5. This invention comprehensively represents the similarity of geographical environment configuration, landslide semantic similarity, spatial proximity similarity, and terrain barrier relationship through graph edge weights, enabling the graph neural network to accurately distinguish the boundaries between internal units of the same landslide and adjacent landslide units during message transmission.

[0024] 6. This invention utilizes graph neural networks for feature propagation of geographic relationship constraints, effectively enhancing the same instance representation among spatial units with similar semantics and continuous terrain configuration on the same slope, while suppressing erroneous feature mixing across valleys, watersheds, or slope units.

[0025] 7. This invention inputs the landslide instance features constrained by geographical relationships and the multi-scale landslide semantic features into the instance decoder. Through joint training with multiple loss functions, it achieves accurate conversion from the landslide semantic probability map to a single landslide mask with independent instance identification.

[0026] 8. This invention significantly improves the separation capability of adjacent adherent landslides and the integrity of the segmentation of the same landslide region by jointly optimizing the matching loss, masking loss, boundary constraint loss and graph node instance relationship loss in the instance decoder.

[0027] 9. This invention combines post-processing operations such as confidence screening, overlapping instance resolution, connected region correction, hole filling, and boundary smoothing in the prediction stage, which effectively improves the boundary accuracy and spatial continuity of the identification results of individual landslide instances.

[0028] 10. The present invention outputs a single landslide instance identification result carrying complete attribute information such as instance number, spatial range, area and confidence level, which can directly serve the practical application needs of landslide cataloging, disaster investigation, engineering survey and geological disaster risk assessment.

[0029] 11. This invention fully utilizes multi-source unlabeled landslide images to adapt a pre-trained visual basic model to a specific domain. It obtains landslide semantic probabilities and multi-scale semantic features through supervised semantic segmentation, divides candidate landslide areas into spatial units, and integrates their semantic features, topographic features, and spatial features to construct a landslide geographic relationship map. It uses graph neural networks and instance decoders to separate instances of adjacent landslides, restore relationships in broken landslide areas, and optimize individual boundary, thereby effectively improving the completeness, accuracy, and cross-data source adaptability of landslide individual identification in complex mountainous environments.

[0030] 12. This invention can solve the problems of insufficient cross-data source adaptability, strong dependence on limited labeled samples, difficulty in identifying single landslide adhesion or fragmentation, and unstable landslide boundary extraction in complex mountainous scenarios of existing landslide intelligent identification methods, and has good applicability. Attached Figure Description

[0031] The present invention will now be further described in detail with reference to the accompanying drawings and specific embodiments, wherein: Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0032] Example 1 See Figure 1 A landslide instance segmentation method that integrates self-supervised visual features and topographic priors includes the following steps: S1. Acquire multi-source landslide images, topographic data corresponding to the spatial distribution of the multi-source landslide images, and landslide annotation data. After spatial registration, preprocessing, and sample segmentation, construct a multi-source landslide training dataset. S2. Using unlabeled image samples from the multi-source landslide training dataset, perform self-supervised domain adaptation training on the pre-trained visual basic model to obtain the backbone network for landslide domain feature extraction. S3. Connect the semantic segmentation decoder to the feature extraction backbone network in the landslide domain, use the semantic segmentation training samples for supervised transfer training, construct the landslide semantic segmentation model, and extract the landslide semantic probability map and multi-scale landslide semantic features. S4. Generate candidate landslide areas based on the landslide semantic probability map, divide the candidate landslide areas into spatial units and use them as graph nodes, integrate the landslide semantic features, topographic features and spatial morphological features of the graph nodes, and construct a landslide geographical relationship map based on spatial adjacency and similarity of geographical environment configuration. S5. Input the landslide geographic relationship map into the graph neural network for feature propagation to obtain landslide instance features constrained by geographic relationships. Then, input the landslide instance features into the instance decoder for training to obtain the landslide instance segmentation model. S6. Obtain image data and topographic data of the area to be identified, input them into the trained landslide instance segmentation model for prediction, and obtain the identification results of individual landslide instances with independent instance identifiers through post-processing.

[0033] This embodiment is the most basic implementation method. Through self-supervised domain adaptation training, it makes full use of a large number of unlabeled multi-source landslide images, reducing the model's dependence on manually labeled samples. At the same time, it uses graph neural networks for feature propagation and instance decoding of geographical constraints, effectively solving the problems of adjacent landslide adhesion and fragmentation within the same landslide. It significantly improves the boundary accuracy, spatial continuity and cataloging practicality of single landslide instance segmentation in complex mountainous scenes.

[0034] Example 2 See Figure 1 A landslide instance segmentation method that integrates self-supervised visual features and topographic priors includes the following steps: S1. Acquire multi-source landslide images, topographic data corresponding to the spatial distribution of the multi-source landslide images, and landslide annotation data. After spatial registration, preprocessing, and sample segmentation, construct a multi-source landslide training dataset. S2. Using unlabeled image samples from the multi-source landslide training dataset, perform self-supervised domain adaptation training on the pre-trained visual basic model to obtain the backbone network for landslide domain feature extraction. S3. Connect the semantic segmentation decoder to the feature extraction backbone network in the landslide domain, use the semantic segmentation training samples for supervised transfer training, construct the landslide semantic segmentation model, and extract the landslide semantic probability map and multi-scale landslide semantic features. S4. Generate candidate landslide areas based on the landslide semantic probability map, divide the candidate landslide areas into spatial units and use them as graph nodes, integrate the landslide semantic features, topographic features and spatial morphological features of the graph nodes, and construct a landslide geographical relationship map based on spatial adjacency and similarity of geographical environment configuration. S5. Input the landslide geographic relationship map into the graph neural network for feature propagation to obtain landslide instance features constrained by geographic relationships. Then, input the landslide instance features into the instance decoder for training to obtain the landslide instance segmentation model. S6. Obtain image data and topographic data of the area to be identified, input them into the trained landslide instance segmentation model for prediction, and obtain the identification results of individual landslide instances with independent instance identifiers through post-processing.

[0035] Preferably, the multi-source landslide training dataset includes landslide-related images from ground cameras, smartphones, drones, and satellite remote sensing sensors, as well as digital elevation models and landslide annotation data corresponding to the image space.

[0036] In S2, the self-supervised domain adaptation training includes multi-view enhancement of unlabeled image samples. Through feature consistency constraints between teacher and student networks, masked image patch learning, and cross-view local consistency constraints, the pre-trained visual base model is adapted to the feature distribution of multi-source landslide images.

[0037] In S3, the landslide semantic segmentation model uses a self-supervised domain-adapted landslide domain feature extraction backbone network as an encoder, and outputs a landslide semantic probability map through a semantic segmentation decoder. The multi-scale landslide semantic features are used as image semantic input for the subsequent construction of landslide geographic relationship maps.

[0038] This embodiment represents a preferred implementation, combining multi-view enhancement, mask image patch learning, and cross-view local consistency constraints in self-supervised learning to effectively improve the model's discriminative visual representation ability of landslide objects and their background environment in complex mountainous terrain. A landslide semantic segmentation model is constructed through supervised transfer training, extracting multi-scale semantic features while acquiring a high-precision landslide semantic probability map, providing a reliable image semantic input foundation for subsequent candidate region selection and spatial unit division. Candidate landslide regions are divided into spatial units, and topographic features are integrated to construct a landslide geographic relationship map, enabling explicit modeling of the spatial continuity, slope integrity, and terrain separation relationships of landslides, thus enhancing the stability of semantic representation within landslides under complex terrain conditions.

[0039] By comprehensively representing the similarity of geographical environment configuration, landslide semantic similarity, spatial proximity similarity, and terrain barrier relationship through graph edge weights, the graph neural network can accurately distinguish the boundaries between internal units of the same landslide and adjacent landslide units during message transmission.

[0040] By using graph neural networks for feature propagation with geographic relationship constraints, we can effectively enhance the same instance representation among spatial units with the same slope, similar semantics, and continuous terrain configuration, while suppressing erroneous feature mixing across valleys, watersheds, or slope units.

[0041] Example 3 See Figure 1 A landslide instance segmentation method that integrates self-supervised visual features and topographic priors includes the following steps: S1. Acquire multi-source landslide images, topographic data corresponding to the spatial distribution of the multi-source landslide images, and landslide annotation data. After spatial registration, preprocessing, and sample segmentation, construct a multi-source landslide training dataset. S2. Using unlabeled image samples from the multi-source landslide training dataset, perform self-supervised domain adaptation training on the pre-trained visual basic model to obtain the backbone network for landslide domain feature extraction. S3. Connect the semantic segmentation decoder to the feature extraction backbone network in the landslide domain, use the semantic segmentation training samples for supervised transfer training, construct the landslide semantic segmentation model, and extract the landslide semantic probability map and multi-scale landslide semantic features. S4. Generate candidate landslide areas based on the landslide semantic probability map, divide the candidate landslide areas into spatial units and use them as graph nodes, integrate the landslide semantic features, topographic features and spatial morphological features of the graph nodes, and construct a landslide geographical relationship map based on spatial adjacency and similarity of geographical environment configuration. S5. Input the landslide geographic relationship map into the graph neural network for feature propagation to obtain landslide instance features constrained by geographic relationships. Then, input the landslide instance features into the instance decoder for training to obtain the landslide instance segmentation model. S6. Obtain image data and topographic data of the area to be identified, input them into the trained landslide instance segmentation model for prediction, and obtain the identification results of individual landslide instances with independent instance identifiers through post-processing.

[0042] The multi-source landslide training dataset includes landslide-related images from ground cameras, smartphones, drones, and satellite remote sensing sensors, as well as digital elevation models and landslide annotation data corresponding to the image space.

[0043] In S2, the self-supervised domain adaptation training includes multi-view enhancement of unlabeled image samples. Through feature consistency constraints between teacher and student networks, masked image patch learning, and cross-view local consistency constraints, the pre-trained visual base model is adapted to the feature distribution of multi-source landslide images.

[0044] In S3, the landslide semantic segmentation model uses a self-supervised domain-adapted landslide domain feature extraction backbone network as an encoder, and outputs a landslide semantic probability map through a semantic segmentation decoder. The multi-scale landslide semantic features are used as image semantic input for the subsequent construction of landslide geographic relationship maps.

[0045] In S4, the graph nodes of the landslide geographic relationship map include landslide semantic features obtained by semantic probability weighted aggregation, topographic features calculated by digital elevation model, and spatial morphological features of candidate spatial units.

[0046] Furthermore, graph edges are established based on the boundary contact relationship, centroid spatial distance, and nearest neighbor relationship between graph nodes. The weights of the graph edges are determined based on the similarity of geographical environment configuration, the semantic similarity of landslides, the spatial proximity similarity, and the terrain barrier relationship.

[0047] The graph neural network performs message passing and feature aggregation between adjacent graph nodes based on graph edge weights, which enhances the same instance representation of graph nodes with continuous terrain and similar semantics, and suppresses erroneous feature mixing across valleys, watersheds, or slope units.

[0048] In S5, the instance decoder generates an instance query vector based on the landslide instance features output by the graph neural network, and outputs the predicted instance category and the predicted instance mask. During training, at least one of the following is constructed by the matching relationship between the predicted instance and the real landslide instance: instance classification loss, mask loss, region overlap loss, boundary constraint loss and graph node instance relationship loss.

[0049] In step S6, the post-processing includes at least one of confidence filtering, overlapping instance resolution, connected component correction, hole filling, boundary smoothing, and vectorization.

[0050] In step S6, the identification results of individual landslide instances include instance number, spatial range, area, and confidence level statistics.

[0051] This embodiment represents the optimal implementation. It inputs both geographically constrained landslide instance features and multi-scale landslide semantic features into the instance decoder. Through joint training with various loss functions, it achieves accurate conversion from a landslide semantic probability map to a mask of individual landslides with independent instance identifiers. Joint optimization of matching loss, mask loss, boundary constraint loss, and graph node instance relationship loss in the instance decoder significantly improves the separation ability of adjacent contiguous landslides and the completeness of regional segmentation within the same landslide. In the prediction stage, post-processing operations such as confidence screening, overlapping instance resolution, connected region correction, void filling, and boundary smoothing effectively enhance the boundary accuracy and spatial continuity of the individual landslide instance identification results. The output individual landslide instance identification results carry complete attribute information such as instance number, spatial range, area, and confidence level, directly serving practical applications such as landslide cataloging, disaster investigation, engineering survey, and geological hazard risk assessment.

[0052] This study leverages multi-source, unlabeled landslide imagery to adapt a pre-trained visual model to a specific domain. Supervised semantic segmentation yields landslide semantic probabilities and multi-scale semantic features. Candidate landslide regions are divided into spatial units, and a landslide geographic relationship map is constructed by fusing semantic, topographic, and spatial features. Graph neural networks and instance decoders are used to separate adjacent landslide instances, restore relationships within fractured landslide areas, and optimize individual landslide boundaries. This effectively improves the completeness, accuracy, and cross-data source adaptability of landslide identification in complex mountainous environments. It addresses the shortcomings of existing intelligent landslide identification methods in complex mountainous scenarios, including insufficient cross-data source adaptability, strong dependence on limited labeled samples, difficulty in identifying contiguous or fractured individual landslides, and unstable landslide boundary extraction. Therefore, it demonstrates good applicability.

[0053] The specific process of the segmentation method of the present invention is described below: S1. Construction of Multi-Source Landslide Training Dataset S11. Collect landslide-related images acquired by smartphones, ground cameras, drones, and satellite remote sensing sensors. The landslide-related images include close-up ground photos, single drone images, drone orthophotos, and satellite optical remote sensing images. Record the sensor type, spatial resolution, acquisition time, image size, and spatial reference information corresponding to each image. S12. Perform quality checks and preprocessing on multi-source landslide images, remove images that are severely blurred, overexposed, have insufficient effective area, or are severely affected by clouds, snow, shadows, or obstructions, and perform duplicate sample cleaning, geometric correction, radiometric adjustment, spatial registration, and resolution unification on the retained images. S13. Obtain the classification label, semantic segmentation label and instance segmentation label corresponding to the multi-source landslide image, wherein the classification label is used to characterize whether the image or image block contains a landslide, the semantic segmentation label is used to characterize the landslide category to which each pixel belongs, and the instance segmentation label is used to distinguish different individual landslide objects in the image. S14. Perform category coding and consistency checks on the classification labels, encode images or image blocks containing landslide targets as landslide categories, encode images or image blocks that do not contain landslide targets as non-landslide categories, and establish the correspondence between image samples and classification labels. S15. Perform rasterization and binarization processing on the semantic segmentation labels, assign the first preset pixel value to the pixels in the landslide area, assign the second preset pixel value to the pixels in the non-landslide area and the background area, and check and correct the holes, isolated pixels, invalid areas and boundary deviations in the labels to obtain the binary semantic labels corresponding to the image pixel positions. S16. Assign a distinct instance identifier value to each individual landslide object in the instance segmentation label, assign a background identifier value to the background area, and establish the association between each instance identifier value and the corresponding individual landslide category, spatial range, and geometric attributes to obtain instance labels that can distinguish different individual landslide objects. S17. Convert the multi-source landslide images into red, green and blue three-channel images, and perform numerical range scaling and standardization on the pixel values ​​of each channel; simultaneously crop the images and their corresponding labels according to the preset input size of 512 pixels × 512 pixels, scale and fill the images that are too small, and use a sliding window to divide the images that are larger than the preset input size into blocks to obtain image tiles, semantic label tiles and instance label tiles with consistent spatial location; S18. Perform at least one of the following data augmentation processes on the image tiles and their corresponding labels: rotation, flipping, scale transformation, brightness adjustment, contrast adjustment, and color perturbation. Divide the image tiles and their corresponding labels into training set, validation set, and test set according to the image source, landslide scene, or spatial region to construct a multi-source landslide training dataset including unlabeled image samples, classification training samples, semantic segmentation training samples, and instance segmentation training samples. S2, Self-supervised domain adaptation of pre-trained visual base models S21. Obtain the visual basic model and its pre-trained weights. Initialize the student network and teacher network with the same feature extraction structure based on the pre-trained weights. The feature extraction structure includes an image patch embedding module, multiple feature encoding modules and a feature normalization module. S22. Select ground close-up photos, UAV single images, UAV orthophotos and satellite optical remote sensing images from the multi-source landslide training dataset, and perform balanced sampling according to sensor type, spatial resolution, image scene and distribution of landslide objects and non-landslide objects to form a batch of unlabeled training samples for self-supervised domain adaptation. S23. Perform multi-view data enhancement on the same unlabeled image sample, and generate at least one global view and at least one local view respectively. The multi-view data enhancement includes at least one of random cropping, scale transformation, horizontal flipping, vertical flipping, brightness adjustment, contrast adjustment, saturation adjustment, color perturbation, blurring and random image block occlusion, and process each view to a preset input size. S24. Input the global view and local view into the student network, and input the global view into the teacher network to extract global semantic features that represent the overall scene and object category of the image, as well as image block-level local features that represent local color or spectral response, texture structure, edge morphology and spatial composition. S25. Construct a global feature consistency loss based on the global semantic features output by the student network and the teacher network for different views of the same image. Construct a mask feature learning loss based on the predicted features of the occluded image patch and the corresponding image patch features of the teacher network. Construct a cross-view local consistency loss based on the local features of the same object in views at different scales. S26. For images acquired by different sensors, constrain features with the same or similar surface object attributes to remain consistent in the feature space, and apply loss weighting to training samples corresponding to different sensor types and image resolutions to reduce the impact of differences in imaging scale, color or spectral response, lighting conditions and observation perspective on the expression of landslide features. S27. A self-supervised domain adaptation loss function is formed based on the global feature consistency loss, mask feature learning loss, and cross-view local consistency loss. The student network parameters are updated based on the self-supervised domain adaptation loss function, and the teacher network parameters are updated based on the student network parameters using an exponential moving average method until the preset training termination condition is met. S28. After training, retain the image patch embedding module, feature encoding module and feature normalization module in the student network or teacher network, and remove the projection module and prediction module used only for self-supervised training to obtain a landslide domain feature extraction backbone network that can extract landslide objects, non-landslide objects and their local boundary features in multi-source landslide images. S3. Landslide semantic segmentation model construction and semantic feature extraction S31. Load the model weights obtained through self-supervised domain adaptation training, remove the projection module and prediction module used for self-supervised training in the model, retain the image patch embedding module, position encoding module, feature encoding module and feature normalization module, construct a landslide semantic segmentation encoder, and connect a semantic segmentation head after the landslide semantic segmentation encoder to form an initial landslide semantic segmentation model. S32. Read red, green and blue three-channel image tiles with a size of 512 pixels × 512 pixels and their corresponding binary semantic labels from the multi-source landslide training dataset, perform channel standardization on the image tiles, and perform synchronous spatial transformation on the image tiles and binary semantic labels to form a batch of semantic segmentation training samples. S33. The image tile is divided into multiple non-overlapping or partially overlapping image blocks by the image block embedding module, and each image block is converted into an image block feature sequence. The image block feature sequence is input into multiple sequentially connected feature encoding modules, and global semantic information and local spatial information of landslide area and non-landslide background area are extracted by self-attention calculation and feedforward feature transformation. S34. Extract image patch features from multiple preset levels of the landslide semantic segmentation encoder, remove classification labels or additional labels that do not have a two-dimensional spatial correspondence, and reconstruct the image patch features of each level into a two-dimensional semantic feature map according to the spatial arrangement of the image patches in the original image; perform channel mapping, scale transformation and feature fusion on the two-dimensional semantic feature maps of different levels to obtain multi-scale landslide semantic features. S35. Input the multi-scale landslide semantic features into the semantic segmentation head, and restore the low-resolution semantic features to the spatial resolution corresponding to the input image through convolutional mapping, feature normalization, nonlinear activation and stepwise upsampling processing, and output the semantic prediction value that represents whether each pixel belongs to the landslide area or the non-landslide background area. S36. Perform activation calculation on the semantic prediction value to obtain the landslide semantic probability map, and construct a semantic segmentation loss function based on the difference between the landslide semantic probability map and the corresponding binary semantic label. The semantic segmentation loss function includes at least two of the following: binary cross-entropy loss, focus loss and boundary constraint loss. S37. Supervised transfer training is performed on the initial landslide semantic segmentation model according to the semantic segmentation loss function; at least some parameters of the landslide semantic segmentation encoder are frozen and the parameters of the semantic segmentation head are updated in the initial stage of training; at least some feature encoding modules of the landslide semantic segmentation encoder are unfrozen in the subsequent stage of training; and the parameters of the landslide semantic segmentation encoder and the semantic segmentation head are jointly updated using a differentiated learning rate until the preset training termination condition is met. S38. Save the model parameters of the trained landslide semantic segmentation encoder and semantic segmentation head to obtain the landslide semantic segmentation model; use the landslide semantic probability map as the basis for extracting candidate landslide areas, and use the multi-scale landslide semantic features output by the landslide semantic segmentation encoder and semantic segmentation head as the image semantic features for constructing the landslide geographic relationship map.

[0054] S4. Generation of candidate landslide areas and construction of landslide geographic relationship map S41. Obtain the landslide semantic probability map and multi-scale landslide semantic feature map output by the landslide semantic segmentation model, and obtain the digital elevation model corresponding to the space of the image to be processed; perform projection transformation, spatial resampling and pixel alignment processing on the digital elevation model, and calculate at least two topographic factors among elevation, slope, aspect, plane curvature, profile curvature, topographic relief, surface roughness, runoff accumulation and topographic humidity index, so that the landslide semantic probability map, multi-scale landslide semantic feature map and topographic factors have consistent spatial range, spatial resolution and pixel correspondence. S42. The landslide semantic probability map is segmented according to a preset semantic probability threshold, and a candidate landslide region mask is generated by at least one of hole filling, small region removal, morphological closing operation, and boundary smoothing. The calculation formula for the candidate landslide region mask is as follows: Formula 1; in: This represents the mask value of the candidate landslide region corresponding to pixel spatial location u; u represents the spatial location of a pixel in the image; P(u) represents the semantic probability that pixel u belongs to the landslide region; This represents the preset semantic probability threshold for landslides; when When =1, it means that the pixel spatial location u is divided into a candidate landslide region; when When =0, it means that the pixel spatial location u is divided into a non-candidate landslide region; S43. Within the spatial range defined by the candidate landslide area mask, the candidate landslide area is divided into multiple candidate spatial units based on at least one of image color or spectral response, texture variation, semantic feature difference, terrain boundary and slope differentiation features; the candidate spatial unit includes at least one of superpixel unit, regular feature block, slope sub-unit or subdivision unit within the candidate connected region; Each candidate spatial unit is treated as a graph node, and all graph nodes form a graph node set: Formula 2; in: V represents the set of nodes in the landslide geographic relationship diagram; This represents the i-th graph node; N This represents the total number of candidate spatial units, i.e., the total number of graph nodes; The granularity of the candidate spatial units is smaller than the expected spatial scale of a single landslide, so that the connected area of ​​the same candidate landslide can be represented by multiple graph nodes, so that adjacent or connected single landslides can be distinguished according to the geographical environment configuration relationship between the nodes. S44. For each graph node, perform semantic probability weighted aggregation on the multi-scale landslide semantic features within its corresponding candidate spatial unit range to obtain the landslide semantic feature vector of the graph node. The calculation formula for the landslide semantic feature vector is as follows: Formula 3; in: Represents the semantic feature vector of the landslide at the i-th graph node; This represents the set of pixels contained in the candidate spatial unit corresponding to the i-th graph node; This represents the multi-scale landslide semantic feature vector corresponding to the pixel spatial location u; This represents the semantic probability that pixel spatial location u belongs to the landslide region. This represents a preset positive number used to prevent the denominator from being zero; The semantic probability weighted aggregation enables pixels with higher landslide semantic probabilities to contribute more to the semantic features of nodes, thereby reducing the impact of edge noise in candidate landslide areas and non-landslide background pixels on the semantic expression of nodes. S45. Statistically analyze the topographic and geomorphic factors within the candidate spatial unit range corresponding to each map node, and calculate at least one of the following statistical measures for each topographic and geomorphic factor: mean, standard deviation, maximum value, minimum value, or median, to obtain the original topographic and geomorphic features of the map node; standardize the k-th topographic and geomorphic feature of the i-th map node using the following formula: Equation 4; in: This represents the original value of the k-th topographic feature of the i-th graph node; This represents the k-th standardized topographic feature. This represents the mean of the k-th topographic feature in the training dataset; This represents the standard deviation of the k-th topographic feature in the training dataset; The serial number indicating the topographic and geomorphic features; This represents a preset positive number used to prevent the denominator from being zero; The geographic environment configuration feature vector of the i-th graph node is composed of the standardized topographic features: Formula 5; in: This represents the geographic environment configuration feature vector of the i-th graph node; K represents the total number of topographic features used to describe the configuration of the geographical environment; For slope aspect features, the slope aspect angle is converted into sine and cosine components and then used to construct the geographic environment configuration feature vector to avoid numerical discontinuities in the slope aspect angle near 0 degrees and 360 degrees. S46. Extract the spatial location features and geometric morphology features of each graph node. The spatial location features and geometric morphology features include at least one of the following: node centroid x-coordinate, node centroid y-coordinate, node area, node perimeter, shape compactness, aspect ratio, main direction, and the identifier of the slope unit to which it belongs, to form the spatial morphology feature vector of the i-th graph node. ; The landslide semantic feature vector, geographic environment configuration feature vector, and spatial morphology feature vector are concatenated, and their dimensions are unified through a feature mapping layer to obtain the graph node attribute vector. The calculation formula is as follows: Formula 6; in: This represents the node attribute vector of the i-th graph node; Represents the semantic feature vector of the landslide at the i-th graph node; This represents the geographic environment configuration feature vector of the i-th graph node; This represents the spatial location and geometric feature vector of the i-th graph node; Represents the trainable feature map weight matrix; This represents a trainable bias vector; Represents a non-linear activation function; The node attribute vector simultaneously represents the image semantic information, topographic information, and spatial geometric information of the candidate spatial unit; S47. Establish candidate graph edges based on the boundary contact relationship, centroid spatial distance, and nearest neighbor relationship between graph nodes; when two graph nodes share a common boundary, or when the centroid spatial distance between two graph nodes is not greater than a preset distance threshold and they are each other's nearest neighbors by a preset number, establish candidate graph edges between the two graph nodes. The formula for calculating the candidate adjacency relationship is as follows: Formula 7; in: This represents the candidate adjacency relationship between the i-th graph node and the j-th graph node; This represents the length of the common boundary between the i-th graph node and the j-th graph node; This represents the spatial distance between the centroids of the i-th graph node and the j-th graph node; This indicates a preset spatial distance threshold; This represents the set of a predetermined number of nearest neighbor nodes for the i-th graph node; This indicates that a candidate graph edge is established between the i-th graph node and the j-th graph node; This indicates that no candidate graph edge is established between the i-th graph node and the j-th graph node; S48. Based on the principle of similarity in geographical environment configuration as characterized by the third law of geography, calculate the geographical environment configuration distance between adjacent map nodes; the geographical environment configuration distance is used to measure the differences between two map nodes in terms of slope, aspect, curvature, topographic relief, and other topographic conditions, and its calculation formula is as follows: Formula 8; in: This represents the geographical configuration distance between the i-th graph node and the j-th graph node; This represents the k-th standardized topographic feature of the i-th graph node; This represents the k-th standardized topographic feature of the j-th graph node; This represents the weight coefficient corresponding to the k-th topographic feature; K represents the total number of topographic and geomorphic features; The weighting coefficients satisfy: Equation 9; The geographical environment configuration distance is converted into geographical environment configuration similarity, and the calculation formula is as follows: Formula 10; in: This represents the similarity of geographical environment configuration between the i-th graph node and the j-th graph node; This represents the geographical configuration distance between the i-th graph node and the j-th graph node; The scale parameter represents the rate at which the similarity of geographic environment configurations decays. The smaller the geographical environment configuration distance, the more similar the terrain and landform conditions of the two map nodes are, and the greater the geographical environment configuration similarity. The larger the geographical environment configuration distance, the more likely the two map nodes are located on different slopes, different aspects, or different landform units, and the smaller the geographical environment configuration similarity. S49. Calculate the landslide semantic similarity and spatial proximity similarity between the two endpoints of the candidate graph; the normalization process limits the landslide semantic similarity to between 0 and 1. The greater the landslide semantic similarity, the more similar the two candidate spatial units are in terms of color or spectral response, texture, edge morphology and landslide object semantics. S410. Determine the terrain barrier coefficient based on whether the two graph nodes cross a watershed, valley line, or a significant terrain change location. S411. By integrating the geographical environment configuration similarity, landslide semantic similarity, spatial proximity similarity, and terrain barrier coefficient, the initial edge weights of the candidate graph edges are calculated. S412. Within the neighborhood of each graph node, the initial graph edge weights are normalized to obtain normalized graph edge weights for feature propagation in the graph neural network. S413. Based on the graph node set, graph edge set, graph node attribute matrix, and graph edge weight matrix, construct a landslide geographic relationship graph. The landslide geographic relationship graph is used to represent the image semantic relationship, geographic environment configuration relationship, spatial proximity relationship, and terrain connectivity relationship between different candidate spatial units within the candidate landslide area, and serves as the input for subsequent graph neural network to perform node feature propagation, instance relationship discrimination, and individual landslide separation. S5, Graph Neural Network Propagation, Instance Decoding, and Model Training S51. Determine the node instance attribution label; The node instance attribution label is used to construct the graph edge relationship supervision signal and the node instance separation supervision signal, so that graph nodes belonging to the same single landslide remain aggregated in the feature space, and graph nodes belonging to different single landslides or background areas remain separable. S52. Initialize graph node features; S53. Construction of graph-edge relationship features: The obtained geographical environment configuration similarity, landslide semantic similarity, spatial proximity similarity, terrain barrier coefficient and normalized graph edge weight are concatenated to construct a graph-edge relationship feature vector. S54. Calculate the attention coefficient of geographic relationship constraint. By adding the logarithmic term of the normalized graph edge weight as the graph edge prior to the attention score, graph nodes with similar geographic environment configuration, similar image semantics, close spatial distance and strong terrain connectivity can obtain higher information propagation weights, and suppress the erroneous feature mixing between graph nodes that cross watersheds, cross valleys or are on different slopes.

[0055] S55, Neighborhood Attention Normalization; S56. Geographically constrained message aggregation and node feature update: Candidate spatial units that are semantically and topographically continuous within the same landslide can enhance each other, while candidate spatial units between different slopes or different instances reduce feature mixing due to weak graph edge priors. S57. Node feature space backfilling: Based on the pixel attribution relationship between each graph node and the candidate spatial unit, the final node instance features are backfilled into the two-dimensional space to generate a geographic constraint graph feature map. S58. Semantic features and graph features are fused. The geographic constraint graph feature map is scaled and channel-mapped to make it have the same spatial size as the multi-scale landslide semantic feature map output in step S3. The two types of features are spliced ​​and nonlinearly mapped to obtain a fused feature map for instance decoding. S59. Instance query vector initialization: Calculate the initial response weight of the query vector to each graph node using the trainable initial query vector and the final instance features of the graph nodes. Through initialization, different instance queries will prioritize node sets with different geographical relationships and different landslide instance features, reducing the probability of multiple queries responding to the same landslide instance at the same time. S510, instance decoder cross-attention update; S511, Instance Category and Instance Mask Prediction: Calculate the probability of the m-th predicted instance belonging to the landslide category or the empty category based on the final instance query vector, map the final instance query vector to the instance mask embedding vector, and map the fused feature map to the pixel mask feature. S512. Matching predicted instances with real instances: To eliminate the inconsistency between the query order of predicted instances and the number of real individual landslide instances, a bipartite graph optimal matching method is adopted to match the set of predicted instances with the set of real instances one by one. S513, Construct the joint loss for instance segmentation; S514. Model Training and Parameter Saving: The graph neural network, feature fusion module, and instance decoder are trained by backpropagation using the joint training loss function. In the initial training phase, all or part of the parameters of the landslide semantic segmentation encoder obtained in step S3 are frozen, and the parameters of the graph neural network, feature fusion module, and instance decoder are updated preferentially. In the subsequent training phase, at least part of the feature encoding layers of the landslide semantic segmentation encoder are unfrozen, and joint fine-tuning is performed using a differentiated learning rate until the joint training loss converges or the preset training termination condition is met. The model parameters of the trained landslide semantic segmentation encoder, graph neural network, feature fusion module, and instance decoder are saved to obtain the landslide instance segmentation model. The landslide instance segmentation model incorporates topographical configuration continuity, spatial adjacency, and image semantic similarity into the feature learning of individual landslide instances through graph feature propagation constrained by geographical relationships. Feature mixing between adjacent landslides is reduced through graph edge relationship discrimination loss and node instance separation loss. Individual separation and boundary restoration of adjacent, contiguous, and internally inconsistent spectral or textural landslide objects are achieved through joint supervision of instance query decoding and instance masking. S6. Prediction of the area to be identified and output of results For the area to be identified, image data and digital elevation models consistent with those used in the training phase are acquired. Spatial registration, resolution unification, and topographic factor calculation are completed. The same standardized parameters as in the training phase are used to generate model input. The image to be identified is input into the trained landslide semantic segmentation model to obtain a semantic probability map and multi-scale semantic features. Graph nodes and edges are constructed based on candidate area masks. Topographic features, spatial morphology features, and edge weights are input into a graph neural network and instance decoder to generate landslide masks with independent instance identifiers. The prediction results are then subjected to confidence filtering, overlapping instance resolution, connected region correction, hole filling, boundary smoothing, and vectorization processing. The output includes individual landslide vector results containing instance number, area, average confidence, spatial extent, and post-processing parameters. This result can be further used for landslide cataloging, emergency investigation, and geological hazard risk assessment.

Claims

1. A landslide instance segmentation method integrating self-supervised visual features and topographic priors, characterized in that, Includes the following steps: S1. Acquire multi-source landslide images, topographic data corresponding to the spatial distribution of the multi-source landslide images, and landslide annotation data. After spatial registration, preprocessing, and sample segmentation, construct a multi-source landslide training dataset. S2. Using unlabeled image samples from the multi-source landslide training dataset, perform self-supervised domain adaptation training on the pre-trained visual basic model to obtain the backbone network for landslide domain feature extraction. S3. Connect the semantic segmentation decoder to the feature extraction backbone network in the landslide domain, use the semantic segmentation training samples for supervised transfer training, construct the landslide semantic segmentation model, and extract the landslide semantic probability map and multi-scale landslide semantic features. S4. Generate candidate landslide areas based on the landslide semantic probability map, divide the candidate landslide areas into spatial units and use them as graph nodes, integrate the landslide semantic features, topographic features and spatial morphological features of the graph nodes, and construct a landslide geographical relationship map based on spatial adjacency and similarity of geographical environment configuration. S5. Input the landslide geographic relationship map into the graph neural network for feature propagation to obtain landslide instance features constrained by geographic relationships. Then, input the landslide instance features into the instance decoder for training to obtain the landslide instance segmentation model. S6. Obtain image data and topographic data of the area to be identified, input them into the trained landslide instance segmentation model for prediction, and obtain the identification results of individual landslide instances with independent instance identifiers through post-processing.

2. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 1, characterized in that: The multi-source landslide training dataset includes landslide-related images from ground cameras, smartphones, drones, and satellite remote sensing sensors, as well as digital elevation models and landslide annotation data corresponding to the image space.

3. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 1, characterized in that: In S2, the self-supervised domain adaptation training includes multi-view enhancement of unlabeled image samples. Through feature consistency constraints between teacher and student networks, masked image patch learning, and cross-view local consistency constraints, the pre-trained visual base model is adapted to the feature distribution of multi-source landslide images.

4. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 1, characterized in that: In S3, the landslide semantic segmentation model uses a self-supervised domain-adapted landslide domain feature extraction backbone network as an encoder, and outputs a landslide semantic probability map through a semantic segmentation decoder. The multi-scale landslide semantic features are used as image semantic input for the subsequent construction of landslide geographic relationship maps.

5. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 1, characterized in that: In S4, the graph nodes of the landslide geographic relationship map include landslide semantic features obtained by semantic probability weighted aggregation, topographic features calculated by digital elevation model, and spatial morphological features of candidate spatial units.

6. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 5, characterized in that: Graph edges are established based on the boundary contact relationships, centroid spatial distance, and nearest neighbor relationships between graph nodes. The weights of the graph edges are determined based on the similarity of geographical environment configuration, the semantic similarity of landslides, the spatial proximity similarity, and the terrain barrier relationship.

7. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 6, characterized in that: The graph neural network performs message passing and feature aggregation between adjacent graph nodes based on graph edge weights, which enhances the same instance representation of graph nodes with continuous terrain and similar semantics, and suppresses erroneous feature mixing across valleys, watersheds, or slope units.

8. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 1, characterized in that: In S5, the instance decoder generates an instance query vector based on the landslide instance features output by the graph neural network, and outputs the predicted instance category and the predicted instance mask. During training, at least one of the following is constructed by the matching relationship between the predicted instance and the real landslide instance: instance classification loss, mask loss, region overlap loss, boundary constraint loss and graph node instance relationship loss.

9. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 1, characterized in that: In step S6, the post-processing includes at least one of confidence filtering, overlapping instance resolution, connected component correction, hole filling, boundary smoothing, and vectorization.

10. The landslide instance segmentation method based on the fusion of self-supervised visual features and topographic priors as described in claim 1, characterized in that: In step S6, the identification results of individual landslide instances include instance number, spatial range, area, and confidence level statistics.

Images