Wide-area insar active landslide map reasoning method and system considering landslide prior knowledge

Abstract

The application relates to the technical field of geological disaster monitoring, and discloses a wide-area InSAR active landslide map reasoning method and system considering landslide prior knowledge, which comprises the following steps: acquiring wide-area InSAR deformation rate field and environmental factor data; adopting a high-precision InSAR deformation detection network to automatically detect a deformation area from the deformation rate field and taking the deformation area as a candidate active landslide node; based on spatial consistency and feature similarity double constraints, a topological relationship graph between the candidate nodes is constructed; the topological relationship graph is input into a graph neural network transformer reasoning model; the model combines the aggregation ability of the graph neural network to local neighborhood information and the capture ability of the transformer to global long-range dependence; and each node is classified and distinguished. The application effectively combines landslide geological environment prior knowledge, and simultaneously considers the local and global spatial correlation between landslides, thereby improving the accuracy and automation level of wide-area active landslide identification.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, specifically to a wide-area InSAR active landslide map reasoning method and system that takes into account prior knowledge of landslides. Background Technology

[0002] Landslides are a major geological hazard worldwide, posing a serious threat to people's lives, property, and infrastructure. Therefore, large-scale, high-precision early identification and monitoring of active landslides is of paramount importance. Synthetic Aperture Radar Interferometry (InSAR) technology, with its advantages of wide coverage, high monitoring accuracy, and independence from weather conditions, has become a core technology for monitoring minute surface deformations and is widely used in landslide hazard investigation and analysis.

[0003] The conventional process for identifying active landslides using InSAR technology typically involves acquiring a surface deformation rate field covering a large area and then identifying deformation areas associated with landslide activity using specific methods. However, existing techniques have several limitations in this process. Firstly, accurately extracting all deformation patches from a wide-area deformation field containing complex background noise and non-landslide deformation signals is inherently challenging. Traditional methods often rely on manual visual interpretation or simple thresholding, which is not only inefficient and labor-intensive, but also heavily depends on expert prior knowledge for accuracy and completeness, making it subjective and prone to missed or incorrect identifications.

[0004] On the other hand, after the initial detection of deformed areas, the more crucial step is to further infer and confirm which areas are truly active landslides. Existing techniques typically treat each detected deformed area as an independent analytical object, classifying it based on its own deformation characteristics and geographical environmental factors (such as slope and lithology). This approach overlooks an important physical fact: landslides within a specific area do not exist in isolation; they often exhibit inherent connections in spatial distribution and geological origin. For example, a group of spatially adjacent and similar deformed areas may collectively constitute a large, complex landslide mass or a landslide cluster. Existing technologies isolate these related areas for separate analysis, failing to effectively utilize the spatial contextual information and group characteristics between deformed areas, thus limiting the overall accuracy and reliability of landslide identification. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a wide-area InSAR active landslide map reasoning method and system that takes into account prior knowledge of landslides. Existing technologies, when using InSAR data to identify active landslides, suffer from low overall identification accuracy due to low deformation zone detection accuracy and the isolation analysis of each deformation zone while ignoring their inherent correlation within the group.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a wide-area InSAR active landslide map inference method that takes into account prior knowledge of landslides. This method aims to achieve accurate identification of active landslides by modeling and inferring the correlations between discrete deformation regions. The method includes the following steps:

[0008] First, obtain the wide-area InSAR surface deformation rate field covering the target study area.

[0009] Subsequently, a high-precision InSAR deformation detection network is used to process the wide-area InSAR deformation rate field to automatically detect and segment a set of deformation regions. Each segmented deformation region is then used as a candidate active landslide node.

[0010] Next, environmental factor data related to landslide occurrence within the study area are acquired. This environmental factor data, along with deformation statistical features extracted from the wide-area InSAR deformation rate field, are concatenated to form the initial feature attribute vector for each candidate active landslide node.

[0011] Then, based on the geographic coordinates of all candidate active landslide nodes and the initial feature attributes, a topological relationship graph between the nodes is constructed. This construction process employs dual constraints of spatial consistency and feature similarity. Specifically, for any two candidate active landslide nodes, a topological connection is established between them only if their spatial distance and feature similarity simultaneously meet preset conditions.

[0012] The topological graph is then input into a graph neural network transformer inference model. The model analyzes the initial characteristic attributes of the nodes in the graph and the topological relationships between them to classify each candidate active landslide node and output the probability that it belongs to an active landslide.

[0013] Finally, based on the classification results, an active landslide distribution map of the study area is generated.

[0014] In one specific embodiment, the high-precision InSAR deformation detection network includes a backbone network for extracting multi-level features from an input deformation field image, a feature pyramid neck for fusing the multi-level features to enhance expressive power, and a detection head for outputting the precise location and boundary range of the deformation region based on the fused features.

[0015] In one specific embodiment, the deformation statistics in the initial feature attributes include, but are not limited to, the average deformation rate, maximum deformation rate, and standard deviation of deformation rate within the deformation region corresponding to the node. The environmental factor data include at least one of the following: altitude, slope, aspect, distance from road, vegetation index, distance from river, lithology, soil type, distance from fault, and rainfall.

[0016] In one specific embodiment, the quantification method for spatial consistency and feature similarity in the step of constructing the topological relationship graph is as follows: the spatial consistency is calculated by quantifying any two candidate active landslide nodes. and Euclidean distance between the center points of the corresponding deformation regions To quantify it, the calculation formula is:

[0017] ;

[0018] in, and Represents any two candidate active landslide nodes; Represents a node and The Euclidean distance between them; and Representing nodes respectively and The geographic coordinates of the center point of the corresponding deformation area.

[0019] The feature similarity is calculated by comparing any two candidate active landslide nodes. and Initial feature attribute vector and Cosine similarity between To quantify it, the calculation formula is:

[0020] ;

[0021] in, Represents a node and The cosine similarity between the initial feature attribute vectors; and Representing nodes respectively and The initial feature attribute vector; Representing vectors and The dot product; and Representing vectors respectively and The Euclidean norm (L2 norm).

[0022] Preferably, the condition for establishing a topological connection between any two candidate active landslide nodes is that the Euclidean distance between the two nodes is... Less than the preset distance threshold And the cosine similarity between the two nodes Greater than the preset similarity threshold .

[0023] Preferably, the graph neural network transformer inference model includes a graph neural network module and a transformer module. The graph neural network module receives the topological relationship graph as input and aggregates the local neighborhood information of each candidate active landslide node through a message passing mechanism to update the node's feature representation. The transformer module receives the updated node feature representation from the graph neural network module and captures the global long-range dependencies between all candidate active landslide nodes through a self-attention mechanism to generate the final node representation used for classification.

[0024] Preferably, the wide-area InSAR deformation rate field is obtained through time-series InSAR processing technology, which is either permanent scatterer radar interferometry (PS-InSAR) or short baseline set radar interferometry (SBAS-InSAR).

[0025] In one specific embodiment, the step of generating an active landslide distribution map specifically involves: setting an active landslide classification probability threshold; identifying candidate active landslide nodes whose probability of being determined as active landslides in the classification process is greater than or equal to the probability threshold as final active landslides, and marking them on the map.

[0026] Preferably, after generating the active landslide distribution map, the method further includes: performing spatial clustering analysis on all nodes identified as final active landslides, for example using the density-based spatial clustering (DBSCAN) algorithm to delineate the boundaries of independent active landslide bodies.

[0027] A second aspect of the present invention provides a wide-area InSAR active landslide map inference system that takes into account prior landslide knowledge, the system being applied to perform the aforementioned method, comprising:

[0028] The data acquisition module is used to acquire wide-area InSAR deformation rate field and environmental factor data related to landslide occurrence covering the study area;

[0029] The deformation region detection module is connected to the data acquisition module. It is configured with a high-precision InSAR deformation detection network to automatically detect and segment a set of deformation regions from the wide-area InSAR deformation rate field, and to define each deformation region as a candidate active landslide node.

[0030] The topology graph construction module is used to combine the environmental factor data and the deformation statistical features extracted from the wide-area InSAR deformation rate field to form the initial feature attributes of each candidate active landslide node, and to construct a topology graph between the candidate active landslide nodes based on the geographic coordinates of the nodes and the initial feature attributes.

[0031] The active landslide inference module is connected to the topology graph construction module. It is configured with a graph neural network transformer inference model to receive the topology graph and classify each candidate active landslide node.

[0032] The landslide map generation module, connected to the active landslide inference module, is used to generate the final active landslide distribution map of the study area based on the classification results.

[0033] This invention provides a wide-area InSAR active landslide map inference method and system that takes into account prior landslide knowledge. It has the following beneficial effects:

[0034] 1. This invention employs a high-precision InSAR deformation detection network to process a wide-area deformation rate field, achieving automated and high-precision extraction of deformed regions. Compared to traditional methods relying on manual visual interpretation or simple threshold segmentation, this approach significantly improves detection efficiency, accuracy, and objectivity, providing a reliable data foundation for subsequent precise inference and resolving the problem of inaccurate initial data extraction affecting the final recognition effect.

[0035] 2. This invention constructs a topological relationship graph based on dual constraints of spatial consistency and feature similarity, transforming all discrete deformation regions into a structured network for overall analysis. This method solves the core defect of existing technologies that treat deformation regions as isolated objects and ignore their inherent correlations. It enables the model to comprehensively utilize the spatial context and group characteristics of landslides, thereby more realistically modeling the regional distribution patterns of landslides and improving the depth of analysis.

[0036] 3. This invention employs a graph neural network transformer composite inference model to perform landslide discrimination on the constructed topological graph. This model can collaboratively capture the local neighborhood information and global long-range dependencies of nodes in the graph. This dual inference mechanism, which combines local and global information, enables the model to more deeply understand the complex causes and spatial patterns of landslide activities, thereby achieving more reliable and robust identification of active landslides in complex geographical environments. Attached Figure Description

[0037] Figure 1 A general map of the Bailongjiang River basin study area;

[0038] Figure 2 A schematic diagram of environmental factor data;

[0039] Figure 3 This is a schematic diagram of the reasoning process for active landslides;

[0040] Figure 4 This is a schematic diagram of the HPDNet network structure;

[0041] Figure 5 This is a schematic diagram of the core module structure of HPDNet;

[0042] Figure 6 A schematic diagram illustrating the process of constructing a topological graph that incorporates prior knowledge of landslides;

[0043] Figure 7 This is a schematic diagram of the Graph-Transformer semi-supervised inference model;

[0044] Figure 8 This is a schematic diagram of the InSAR deformation detection results in the Bailongjiang River Basin.

[0045] Among them, (a) is a diagram showing the deformation detection results of the proposed model in the entire Bailongjiang River Basin; (b)–(e) are comparison diagrams of deformation detection results of different comparison methods in four typical areas;

[0046] Figure 9 This is a schematic diagram showing the distribution of active landslides in the Bailongjiang River basin.

[0047] Among them, (a) spatial distribution map of active landslides in the Bailongjiang River Basin; (b) map of the number of InSAR deformation areas detected in each administrative region; and (c) map of the number of active landslides inferred in each administrative region.

[0048] Figure 10 This is a schematic diagram illustrating the verification and overlay of InSAR results on Google Earth imagery.

[0049] Among them, (1-a)–(10-a) InSAR surface deformation detection results; (1-b)–(10-b) Google imagery visual interpretation results; (1-c)–(10-c) field verification results;

[0050] Figure 11 This is a diagram illustrating the precision-recall curve and the F1 score curve.

[0051] Among them, (a) is a comparison of precision-recall curves of different models when the intersection-union ratio threshold is 0.5; (b) is a comparison of F1 score curves of different models at different confidence thresholds.

[0052] Figure 12 Schematic diagrams of strategies for constructing different graphs;

[0053] Figure 13 This diagram illustrates the performance comparison of the GNN-Transformer model with other models across multiple evaluation metrics, including accuracy, precision, recall, F1 score, and area under the true / false / false curve.

[0054] Figure 14 A schematic diagram of the true report rate - false negative rate curves for each comparative model;

[0055] Figure 15 This is a flowchart of the steps of the present invention.

[0056] Figure 16 This is a multicollinearity analysis diagram of prior information for landslides. Detailed Implementation

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] See attached document Figure 3 and attached Figure 15 This embodiment uses the attached... Figure 1 The Bailongjiang River basin shown is the study area. A wide-area InSAR active landslide map inference method that takes into account prior landslide knowledge is provided. This method specifically includes the following steps:

[0059] Acquire a wide-area InSAR deformation rate field covering the study area, as well as environmental factor data related to landslide occurrence. Environmental factor data include one or more of the following: elevation, slope, aspect, distance from road, vegetation index, distance from river, lithology, soil type, distance from fault, and rainfall.

[0060] A high-precision InSAR deformation detection network is employed to automatically detect and segment a set of deformation regions from a wide-area InSAR deformation rate field. Each deformation region is then used as a candidate active landslide node.

[0061] Environmental factor data, along with deformation statistical features extracted from the wide-area InSAR deformation rate field, are used together as the initial feature attributes for each candidate active landslide node. The deformation statistical features include the average deformation rate, maximum deformation rate, and standard deviation of the deformation rate within the deformation region corresponding to the node.

[0062] Based on the geographic coordinates and initial feature attributes of all candidate active landslide nodes, a topological relationship graph between the nodes is constructed. This construction process employs dual constraints of spatial consistency and feature similarity; specifically, for any two candidate active landslide nodes… and The condition for establishing a topological connection between them is that the Euclidean distance between the two nodes is... Less than the preset distance threshold And the cosine similarity between the two nodes Greater than the preset similarity threshold European distance The calculation formula is:

[0063] ;

[0064] in, and Represents any two candidate active landslide nodes; Represents a node and The Euclidean distance between them; and Representing nodes respectively and The geographic coordinates of the center point of the corresponding deformation area.

[0065] Cosine similarity The calculation formula is:

[0066] ;

[0067] in, Represents a node and The cosine similarity between the initial feature attribute vectors; and Representing nodes respectively and The initial feature attribute vector; Representing vectors and The dot product; and Representing vectors respectively and The Euclidean norm (L2 norm).

[0068] The constructed topological graph is input into the graph neural network transformer inference model. The model analyzes the initial feature attributes of the nodes in the graph and the topological relationships between the nodes to classify and identify each candidate active landslide node, and outputs the probability that it belongs to an active landslide.

[0069] Based on the classification results and a probability threshold for active landslide classification, candidate active landslide nodes that are judged as active landslides with a probability greater than or equal to the probability threshold are identified as the final active landslides, and an active landslide distribution map of the study area is generated.

[0070] Accordingly, the present invention also provides a wide-area InSAR active landslide map inference system that takes into account prior knowledge of landslides, the system being used to execute the aforementioned method. The system includes: a data acquisition module for acquiring wide-area InSAR deformation rate field and environmental factor data covering the study area.

[0071] The deformation region detection module connects its input to the output of the data acquisition module. This module uses a high-precision InSAR deformation detection network to automatically detect and segment a set of deformation regions from a wide-area InSAR deformation rate field, defining each deformation region as a candidate active landslide node.

[0072] The topology graph construction module connects its input to the output of the data acquisition module and the deformation region detection module. This module uses environmental factor data and deformation statistical features extracted from the deformation field as initial feature attributes for each candidate active landslide node, and constructs a topological relationship graph between these nodes based on their geographic coordinates and initial feature attributes.

[0073] The active landslide inference module connects its input to the output of the topology graph construction module. This module takes the topology graph and inputs it into a graph neural network transformer inference model to classify and identify each candidate active landslide node.

[0074] The landslide map generation module connects its input to the output of the active landslide inference module. Based on the classification results, the landslide map generation module generates a distribution map of active landslides in the study area.

[0075] See attached document Figure 1This step first requires acquiring a wide-area InSAR surface deformation rate field covering the target study area (e.g., the Bailong River basin). This deformation rate field is obtained through time-series InSAR processing of multi-period synthetic aperture radar (SAR) satellite imagery. In one specific implementation, the SAR image data used is C-band data acquired by the Sentinel-1A satellite in interferometric wide swath (IW) mode. The time-series InSAR processing technique employed is Short Baseline Set Radar Interferometry (SBAS-InSAR). Through a series of processing steps, including registration, interferometric pair generation, differential interferometry, unwrapping, deformation rate inversion, and geocoding of multiple SAR images over time, a surface deformation rate field with precise geographic coordinates in raster data format is finally generated. In this data field, the value of each pixel represents the annual average deformation rate in the direction of linear surface subsidence or uplift at that geographic location.

[0076] See attached document Figure 2 and attached Figure 16 This step also requires acquiring data on various environmental factors related to landslide occurrence within the study area. This data constitutes prior knowledge for assessing landslide activity. In one specific implementation, the acquired environmental factor data includes: topographic factors derived from a digital elevation model (DEM), such as elevation, slope, and aspect; geological factors obtained from geological maps, such as soil type, lithology, and distance from faults; vegetation factors calculated from remote sensing imagery, such as vegetation indices; and human activity and hydrological factors obtained from infrastructure and hydrological data, such as distance from roads, distance from rivers, and rainfall.

[0077] After acquiring the InSAR deformation rate field and various environmental factor data, all data need to be preprocessed to ensure their spatial consistency. This preprocessing step involves resampling and projection transformation of all environmental factor data layers to unify them to the exact same spatial resolution, coordinate system, and raster grid as the InSAR deformation rate field. This operation ensures that any geographic location (i.e., raster pixel) within the study area can precisely correspond to a deformation rate value and a complete set of environmental factor values, providing a data foundation for subsequently constructing the initial feature attributes of each candidate active landslide node.

[0078] See attached document Figure 4This invention employs a high-precision InSAR deformation detection network (HPIDNet) to automatically detect and segment deformation regions from previously acquired wide-area InSAR deformation rate field images. The network uses an encoder-decoder architecture, and its overall structure includes a backbone network for extracting multi-level features from the input deformation field image, a feature pyramid neck for fusing multi-level features to enhance target representation, and a detection head for outputting the precise location and boundary range of the deformation region based on the fused features.

[0079] To improve detection performance, the network neck integrates a Spatial Pyramid Multi-Scale Feature Fusion (SPMFF) component, and the detection head integrates a Context-Guided Bounding Box Decoding (CGBD) component. (See attached diagram.) Figure 5 During processing, the network employs a dynamic upsampling (Dysample) operator, which addresses the feature space misalignment caused by downsampling through a learnable grid sampling process. This process is defined by the following formula:

[0080] ;

[0081] in, The feature map to be resampled represents the input; This represents a sampling grid that contains the coordinates of sampling points used to generate the output feature map. This sampling grid is not fixed but is determined by the network based on the input features. Dynamic prediction generation; This represents a grid sampling function that is based on the sampling grid. The coordinates defined in the input feature map Sampling is performed using bilinear interpolation. This represents the output feature map generated after dynamic resampling.

[0082] The processed deformation rate field image is input into a high-precision InSAR deformation detection network, which outputs a set of bounding boxes for the deformation regions. Each detected deformation region is treated as a candidate active landslide node. Subsequently, the center point of the bounding box of each node is mapped from the image pixel coordinate system to the geographic coordinate system. First, the center point of the bounding box in the local pixel coordinate system is calculated. The calculation formula is as follows:

[0083] ;

[0084] in, The pixel coordinates of the first corner point of the bounding box of the deformed region in the subgraph coordinate system; The pixel coordinates of the other corner point of the diagonal of the bounding box of the deformed region in the subgraph coordinate system; This represents the calculated pixel coordinates of the center point of the bounding box in the subgraph coordinate system.

[0085] Next, it is mapped to the global pixel coordinate system. The calculation formula is as follows:

[0086] ;

[0087] in, This represents the pixel coordinates of the center point in the subgraph coordinate system calculated above; This represents the offset of the top-left vertex of the subgraph in the global pixel coordinate system of the original wide-area rate field image; This represents the coordinates of the center point of the deformed region in the global pixel coordinate system, obtained from the calculation.

[0088] Finally, the geographic affine transformation parameters are used to convert it into geographic coordinates. The calculation formula is as follows:

[0089] ;

[0090] in, Represents global pixel coordinates; Represents the final output geographic coordinates. There are six affine transformation parameters. Specifically, and These are the top-left corner pixels of the original raster image. direction and Geographic coordinates of direction; Is it pixels? Width of direction (east-west resolution); Is it pixels? The height of the direction (north-south resolution, its value is usually negative); and These are the rotation or shearing coefficients of the image, assuming the original raster image has not undergone rotation or shearing deformation (i.e., the image's rotation or shearing coefficients). The axis is parallel to the geographical north-south direction. (Axis parallel to the geographic east-west direction), parameters and The values ​​are all 0.

[0091] See attached document Figure 8 This figure objectively demonstrates the detection results of HPDNet after its application in the Bailongjiang River Basin. Sub-figures (b), (c), (d), and (e) show that compared to other models, HPDNet can more accurately identify the true deformation regions and has a lower false positive and false negative rate.

[0092] In a specific embodiment of the present invention, the landslide topology map construction step in the method is described in detail. This step aims to transform a set of discrete, unrelated candidate active landslide nodes detected in the previous step into a structured topology map to explicitly represent the inherent relationships between these nodes.

[0093] First, for each candidate active landslide node Construct its initial feature attribute vector The vector is composed of two concatenated data parts. The first part consists of deformation statistics extracted from the wide-area InSAR deformation rate field, representing the deformation statistics within the deformation region corresponding to the node. Specifically, this includes the average, maximum, and standard deviation of the deformation rate of all pixels within the region. The second part consists of the corresponding values ​​extracted from the acquired and preprocessed environmental factor data layer at the geographic coordinates of the node's center point, such as the point's elevation, slope, lithology, and distance from the river.

[0094] See attached document Figure 6 After constructing the initial feature attribute vectors for all nodes, the system determines whether there is a topological connection (i.e., an edge in the graph) between any two candidate active landslide nodes based on the dual constraints of spatial consistency and feature similarity.

[0095] Spatial consistency is achieved by calculating any two candidate active landslide nodes. and Euclidean distance between the center points of the corresponding deformation regions To quantify.

[0096] Feature similarity is calculated by comparing any two candidate active landslide nodes. and Initial feature attribute vector and Cosine similarity between To quantify.

[0097] At any two candidate active landslide nodes and The condition for establishing a topological connection edge between two nodes is: the Euclidean distance between the two nodes. Less than a preset distance threshold And the cosine similarity between the two nodes Greater than a preset similarity threshold By traversing and evaluating all node pairs, a complete topological graph is ultimately generated. ,in It is the set of all candidate active landslide nodes. It is the set of all edges determined according to the above dual constraints.

[0098] In a specific embodiment of the present invention, the graph reasoning and classification steps in the method are described in detail. This step involves constructing a topological graph in the previous step, which contains all candidate active landslide nodes and their interrelationships. Input a graph neural network transformer inference model to classify and distinguish each node.

[0099] See attached document Figure 7 The model structurally integrates a graph neural network module and a transformer module to collaboratively capture local neighborhood information and global long-range dependencies of nodes in the graph.

[0100] The graph neural network module first processes the input topological graph. For each node in the graph... This module aggregates its data in the graph through a message passing mechanism. The feature information of directly connected neighboring nodes is used to update the node. Its own characteristic representation. In one specific implementation, the update process can be performed iteratively. Next, in the In the next iteration, the node eigenvectors The update process can be represented by the following formula:

[0101]

[0102] in, Indicates the first Layer nodes The output, This represents the activation function. Indicates the first The layer weight matrix is ​​responsible for mapping the input to the current layer, and CONCAT represents the concatenation operation. Indicates a node Adjacent nodes In the The features of the layer are summed, an operation used to synthesize information from neighboring nodes.

[0103] After the graph neural network module outputs a set of node feature vectors containing local context information, this set is input to the transformer module. The transformer module, through its internal multi-head self-attention mechanism, calculates the mutual influence weights between any two nodes in the graph, regardless of whether they are directly connected in the topological graph. Specifically, for all node features in the graph, a query matrix is ​​first generated through a linear transformation. Key matrix Value matrix Next, the attention score between nodes is calculated using the following formula:

[0104] ;

[0105] in, This represents the computation of the attention mechanism, accepting a query vector. Key vector Sum value vector As input, A query matrix is ​​typically a representation of the current word or element in an input sequence. The key matrix represents the features of all elements in the input sequence. The value matrix corresponds to the key features, and the attention mechanism will affect the weighting of these values ​​based on the degree of matching. Calculate the similarity between the query and the key. yes transpose, It is a normalization factor. The softmax function is used to convert attention scores into a probability distribution such that the sum of all scores is 1, which can then be used as weights to assign to values.

[0106] Through this self-attention mechanism, the transformer module is able to capture the global long-range dependencies of landslides in their geographic spatial distribution.

[0107] The final output layer of the graph neural network transformer inference model is a classifier. This classifier receives the final node feature representation output by the transformer module, which integrates local and global information, and outputs a probability value between 0 and 1 for each candidate active landslide node. This value represents the likelihood that the node belongs to a real active landslide.

[0108] In one specific embodiment of the present invention, the steps of generating, verifying, and optimizing the active landslide distribution map in the method are described in detail. This step aims to generate the final result map based on the reasoning results of the previous step and to verify its reliability.

[0109] After receiving the activity probability of each candidate active landslide node output by the graph neural network transformer inference model, an active landslide classification probability threshold is set. In one specific implementation, this probability threshold can be set to 0.5. All candidate active landslide nodes are traversed, and nodes whose probability of being classified as active landslides is greater than or equal to the probability threshold are identified as the final active landslides. The geographic coordinates of all nodes identified as final active landslides are projected onto a Geographic Information System (GIS) platform to generate an active landslide distribution map of the study area.

[0110] See attached document Figure 9To verify the identification results of the active landslide distribution map, several typical areas identified as active landslides were selected from the map, and their InSAR deformation results were overlaid and compared with high-resolution optical images. (See attached image.) Figure 10 Subplots (1-a) to (10-a) in the figure show the active landslide areas identified by the model, and the corresponding subplots (1-b) to (10-c) are high-resolution optical images of these areas. Through comparative analysis, the deformation areas identified by the model have a high degree of spatial consistency with the landslide backwall, sliding direction, and other geomorphic features shown on the optical images, thus verifying the reliability of the identification results by the method of this invention.

[0111] In an optional embodiment, after generating the active landslide distribution map, post-processing optimization can be performed on all nodes identified as ultimately active landslides. This optimization step aims to delineate the boundaries of independent active landslide bodies from the discrete set of landslide points. Specifically, a spatial clustering algorithm can be applied to the geographic coordinates of all ultimately active landslide nodes. In one specific implementation, this algorithm is the Density-Based Noisy Application Spatial Clustering (DBSCAN) algorithm. This algorithm achieves the identification and delineation of landslide bodies from point targets to area landslide bodies by grouping spatially adjacent landslide nodes into the same cluster, treating each cluster as an independent active landslide body.

[0112] In one specific embodiment of the present invention, to objectively evaluate the performance of the proposed method, the graph neural network transformer inference model used in this embodiment is quantitatively compared with various baseline models. Baseline models include traditional machine learning models such as multilayer perceptrons (MLP), support vector machines (SVM), and random forests (RF), as well as other graph neural network models such as graph convolutional networks (GCN) and graph sampling and aggregation networks (GraphSAGE).

[0113] See attached document Figure 13 The results in the figure show that the model (Ours) used in this invention outperforms all the comparison models in all evaluation metrics, including Area Under the Receiver Operating Characteristic (AUC), Accuracy, Precision, Recall, and F1 score.

[0114] See attached document Figure 14 The receiver operating characteristic (AUC-ROC) curve in the figure shows that the curve of the model of this invention completely envelops the curves of all other comparison models and achieves the highest AUC value, specifically 0.926. This indicates that the model used in this invention has the best overall classification performance at all classification thresholds.

[0115] See attached document Figure 11 The precision-recall (PR) curve in the figure further demonstrates that the model used in this invention can achieve a high recall level while maintaining a high precision, thus verifying the effectiveness of the model in the landslide identification task.

[0116] To demonstrate the effectiveness of the topological graph construction strategy employing both spatial consistency and feature similarity constraints used in this invention, an ablation experiment was also conducted. (See attached document.) Figure 12 This figure compares the model performance under three different graph construction strategies, including:

[0117] The paper presents three construction strategies: spatial-based, feature-based, and hybrid / Ours (a strategy that combines both). Numerical experimental results demonstrate that the hybrid strategy employed in this invention achieves optimal results across all performance metrics compared to strategies using only a single constraint, thus confirming the necessity and superiority of this dual-constraint graph construction method.

[0118] Accordingly, the present invention also provides a wide-area InSAR active landslide map inference system that takes into account prior knowledge of landslides. This system is used to execute all the aforementioned method steps. In one specific embodiment, the system can be deployed on one or more servers and its functionality is implemented through software programs. The system includes:

[0119] The data acquisition module is used to acquire wide-area InSAR deformation rate field and various environmental factor data covering the specified study area. This module includes interfaces for accessing and downloading public or commercial satellite data archives (such as the ESA Copernicus Data Center), as well as interfaces for extracting geological, hydrological, and transportation data from geospatial databases.

[0120] The deformation region detection module, whose input is connected to the output of the data acquisition module, receives wide-area InSAR deformation rate field data. Internally, this module contains a pre-trained high-precision InSAR deformation detection network (HPIDNet). Its function is to perform inference calculations on the deformation rate field image to automatically detect and segment a set of deformation regions, and then output the location and boundary information of each deformation region as a candidate active landslide node.

[0121] The topology graph construction module connects its inputs to the outputs of both the data acquisition and deformation area detection modules. This module receives information on candidate active landslide nodes and environmental factor data. Its function is to extract and combine the initial feature attributes of each candidate node, and based on the geographic coordinates of all nodes and their initial feature attributes, and under the dual constraints of spatial consistency and feature similarity, calculate and generate a topology graph representing the relationships between all candidate nodes.

[0122] The active landslide inference module connects to the output of the topology graph construction module to receive the topology graph. Internally, this module contains a pre-trained graph neural network transformer inference model. Its function is to perform graph inference calculations on the topology graph, outputting a probability value for each candidate active landslide node that it belongs to an active landslide.

[0123] The landslide map generation module connects its input to the output of the active landslide inference module. This module receives the activity probability values ​​of all candidate nodes. Its function is to filter all nodes based on a preset probability threshold, visualize the location information of the selected active landslide nodes, and generate and output an active landslide distribution map of the study area.

Claims

1. A wide-area InSAR active landslide map reasoning method that takes into account prior knowledge of landslides, characterized in that, Includes the following steps: Obtain the wide-area InSAR deformation rate field covering the study area; A high-precision InSAR deformation detection network is used to automatically detect and segment a set of deformation regions from the wide-area InSAR deformation rate field. Each of the aforementioned deformation regions is considered as a candidate active landslide node; Acquire environmental factor data related to landslide occurrence within the study area, and extract deformation statistical features of each deformation region from the wide-area InSAR deformation rate field; The environmental factor data and the deformation statistical features are concatenated to form the initial feature attribute vector of each candidate active landslide node; Based on the geographic coordinates of the candidate active landslide nodes and the initial feature attribute vector, a topological relationship graph between the candidate active landslide nodes is constructed using the dual constraints of spatial consistency and feature similarity. The topological relationship graph is input into a graph neural network transformer inference model. The graph neural network transformer inference model classifies and identifies each candidate active landslide node by analyzing the initial feature attributes of the nodes in the graph and the topological relationship between the nodes, and outputs the probability that it belongs to an active landslide. Based on the probability of the active landslides, an active landslide distribution map of the study area is generated.

2. The wide-area InSAR active landslide map inference method considering prior landslide knowledge as described in claim 1, characterized in that, The deformation statistical features include at least one of the following: average deformation rate, maximum deformation rate, and standard deviation of deformation rate within the deformation region corresponding to the node; The environmental factor data include at least one of the following: altitude, slope, aspect, distance from road, vegetation index, distance from river, lithology, soil type, distance from fault, and rainfall.

3. The wide-area InSAR active landslide map inference method considering prior landslide knowledge as described in claim 1, characterized in that, In the step of constructing the topological relationship graph: The spatial consistency is quantified by calculating the Euclidean distance between the center points of the deformation regions corresponding to any two candidate active landslide nodes. The feature similarity is quantified by calculating the cosine similarity between the initial feature attribute vectors of any two candidate active landslide nodes.

4. The wide-area InSAR active landslide map reasoning method considering prior landslide knowledge as described in claim 3, characterized in that, The condition for establishing a topological connection between any two candidate active landslide nodes is: The Euclidean distance between the two nodes is less than a preset distance threshold, and the cosine similarity between the two nodes is greater than a preset similarity threshold.

5. The wide-area InSAR active landslide map inference method considering prior landslide knowledge as described in claim 1, characterized in that, The graph neural network converter inference model includes a graph neural network module and a converter module. The graph neural network module is used to aggregate the local neighborhood information of each candidate active landslide node in order to update the feature representation of the node; The transformer module is used to capture the global long-range dependencies between all the candidate active landslide nodes, and to generate the final node representation for classification.

6. The wide-area InSAR active landslide map reasoning method considering prior landslide knowledge according to claim 1, characterized in that, The high-precision InSAR deformation detection network includes a backbone network for extracting multi-level features, a feature pyramid neck for fusing multi-level features, and a detection head for outputting the location and boundary of the deformation region.

7. The wide-area InSAR active landslide map reasoning method considering prior landslide knowledge according to claim 1, characterized in that, The wide-area InSAR deformation rate field is obtained through time-series InSAR processing technology, which is either permanent scatterer radar interferometry or short baseline set radar interferometry.

8. The wide-area InSAR active landslide map reasoning method considering prior landslide knowledge according to claim 1, characterized in that, The specific steps for generating the active landslide distribution map are as follows: Set a probability threshold for classifying active landslides; Candidate active landslide nodes whose probability of being classified as active landslides is greater than or equal to the probability threshold in the classification process will be identified as the final active landslides.

9. The wide-area InSAR active landslide map reasoning method considering prior landslide knowledge according to claim 1, characterized in that, After generating the active landslide distribution map, the process also includes: Spatial cluster analysis was performed on all nodes identified as final active landslides to delineate the boundaries of independent active landslide bodies.

10. A wide-area InSAR active landslide map inference system that takes into account prior knowledge of landslides, characterized in that, The method for inferring active landslide maps in wide-area InSAR, which takes into account prior knowledge of landslides, as described in any one of claims 1-9, includes: The data acquisition module is used to acquire wide-area InSAR deformation rate field and environmental factor data covering the study area; The deformation region detection module is used to automatically detect and segment a set of deformation regions from the wide-area InSAR deformation rate field using a high-precision InSAR deformation detection network, and to define each deformation region as a candidate active landslide node. The topology graph construction module is used to combine the environmental factor data and the deformation statistical features extracted from the wide-area InSAR deformation rate field to form the initial feature attributes of each candidate active landslide node, and to construct a topology graph between the candidate active landslide nodes based on the geographic coordinates of the nodes and the initial feature attributes. The active landslide inference module is used to input the topological relationship graph into a graph neural network transformer inference model to classify and identify each candidate active landslide node. The landslide map generation module is used to generate an active landslide distribution map of the study area based on the classification results.

Images