A method for intelligent identification and segmentation of linear structures in fault zones based on fusion of multi-source remote sensing and LiDAR point clouds

Abstract

The application discloses a kind of based on the fusion of multiple remote sensing and LiDAR point cloud linear structure intelligent identification and segmentation system and method of fracture zone.It is extracted to construct tensor by the deformation anomaly guided LiDAR acquisition, input three-dimensional attention deformable convolution network to output fracture probability body after the system obtains InSAR and LiDAR data, realizes kinematic segmentation after tracking fracture main line and time series clustering and mutation point detection.The method breaks through the limitation of single data source, improves the identification accuracy and segmentation quantification level under complex terrain, and adapts to the needs of seismic risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of geology and relates to an intelligent identification and segmentation system and method for linear structures of fault zones based on the fusion of multi-source remote sensing and LiDAR point clouds. Background Technology

[0002] Intelligent identification and kinematic feature analysis of linear fault zone structures are core foundational work for active tectonic research, seismic hazard assessment, and site selection for major engineering projects. However, this field has long been hampered by the following technical bottlenecks: Existing remote sensing technologies suffer from significant functional gaps in fault structure detection: while optical remote sensing imagery provides intuitive information on surface texture, it is easily obscured by vegetation cover and clouds, making it difficult to effectively reveal fault landforms in high-altitude canyon areas such as southern Tibet and western Sichuan; synthetic aperture radar interferometry can acquire surface deformation fields over a wide area, reflecting fault activity, but its spatial resolution is limited, making it difficult to characterize the fine geometric structure of fault zones; lidar point cloud data can penetrate vegetation to obtain high-precision micro-topography, but its coverage is limited by flight costs and data processing efficiency. This contradiction of "seeing a wide area but not being able to see it clearly, and seeing clearly but not being able to see a wide area" makes it difficult for a single data source to fully support the entire research chain of fault zones, from regional location to detailed local interpretation. Current mainstream paradigms in fault structure research still focus on "fault discovery" and "fracture location," with insufficient attention paid to the segmentation of kinematic properties along fault zones. However, active faults exhibit distinctly different seismic gestation patterns and engineering hazard characteristics across their various segments, including strike-slip, dip-slip, creep, and locked segments. For example, the Xianshuihe Fault Zone, with Daofu County as its boundary, shows significant differences in strike-slip rate, activity mode, and degree of locking between its northern and southern segments. This segmentation directly impacts the accuracy of regional seismic hazard assessment. Similarly, research on the Haiyuan Fault Zone reveals that the left-lateral slip rate decreases from 5 mm / a to 1 mm / a in its western, central, and eastern segments, while the compression rate increases from 1 mm / a to 4 mm / a, demonstrating highly significant segmentation characteristics. However, traditional methods rely heavily on geological surveys and manual interpretation to obtain segmentation information, resulting in low efficiency, high subjectivity, and difficulty in quantification. Although existing research has attempted to fuse multi-source remote sensing data for linear structure identification, current fusion models mostly remain at the data layer or are simply superimposed, failing to achieve functional complementarity between "deformation information" and "geomorphological information." For example, the fusion of multi-source information such as gravity, magnetism, and remote sensing is mostly used for mineralization prediction, while lacking in-depth exploration of the kinematic properties of fault structures. In recent years, deep learning methods have made progress in the segmentation of geological linear structures, but existing models lack the ability to adaptively extract fault bending morphology and have not designed dynamic feature weighting mechanisms for different geographical environments (such as vegetation cover areas), resulting in limited generalization ability of the models. During the data acquisition phase, there is a lack of linkage between the deformation anomaly areas identified by InSAR and the fine-grained LiDAR scans. Traditional multi-source data acquisition is often "post-processing fusion," meaning that registration and fusion are performed after data acquisition, rather than "deformation-guided acquisition." This approach makes it difficult to accurately match the two types of data at spatial scales, and the LiDAR scanning range lacks specificity, resulting in data redundancy and inefficiency. In summary, there is an urgent need for an intelligent technical solution that can overcome the limitations of a single data source, achieve complementary deformation and geomorphological functions, and possess the ability to quantitatively segment kinematic properties, in order to support high-precision identification of active faults and refined assessment of seismic hazard. Summary of the Invention

[0003] To address the problems existing in the background technology, this application proposes an intelligent identification and segmentation system and method for linear structures of fault zones based on the fusion of multi-source remote sensing and LiDAR point clouds. To achieve the above objectives, the technical solution adopted in this application is as follows: On one hand, this invention provides an intelligent identification and segmentation system for linear fault zone structures based on the fusion of multi-source remote sensing and LiDAR point clouds, comprising: The data acquisition module is configured to acquire spaceborne synthetic aperture radar interferometry (InSAR) data and UAV lidar (LiDAR) point cloud data covering the target area; The deformation anomaly guidance module is configured to identify regional-scale surface deformation anomaly areas based on the InSAR data, and to guide the acquisition range and encryption scanning density of the LiDAR point cloud data with the deformation anomaly areas. The multi-source feature construction module is configured to extract micro-topographic features from the LiDAR point cloud data and deformation features from the InSAR data. The feature fusion and recognition module is configured to construct a multi-source feature tensor from the micro-topographic features and the deformation features, input it into a three-dimensional attention mechanism deformable convolutional network, and output a probability volume of fracture existence. The fracture tracking module is configured to automatically track continuous fracture main lines from the fracture existence probability volume based on the minimum cost path algorithm. The kinematic feature segmentation module is configured to extract the radar line-of-sight deformation time series along the main fracture line, perform temporal clustering analysis and abrupt change point detection on the time series, and divide the fracture zone into segments with different kinematic properties based on the clustering and detection results. Furthermore, the micro-topographic features include at least one of the following: a high-precision digital elevation model generated from LiDAR point cloud data after filtering and classifying to remove vegetation points; terrain roughness; slope; curvature; terrain openness; terrain relief; and point cloud echo intensity. The deformation characteristics include at least one of annual average deformation rate, deformation gradient, and time series coherence. Furthermore, the three-dimensional attention mechanism deformable convolutional network includes: Deformable convolutional layers, configured to adaptively extract curved or discontinuous linear constructs; The attention-gated fusion module is configured to dynamically adjust the fusion weights of the micro-topographic features and the deformation features according to the terrain complexity. In areas where the vegetation coverage exceeds a preset threshold, the weights of optical image features are reduced and the weights of LiDAR penetration features are increased. Furthermore, the temporal clustering algorithm configured in the kinematic feature segmentation module includes the K-Shape clustering algorithm, and the mutation point detection algorithm includes the Pettitt test; The kinematic properties include at least one of the following: a run-slip dominant segment, a tilt-slip dominant segment, a creep segment, and a locked segment. On the other hand, the present invention also provides a method for intelligent identification and segmentation of linear fault zone structures based on the fusion of multi-source remote sensing and LiDAR point clouds, comprising the following steps: S1. Acquire spaceborne InSAR data and UAV LiDAR point cloud data covering the target area; S2. Identify regional-scale surface deformation anomaly areas based on the InSAR data, and use the deformation anomaly areas to guide the acquisition range and encryption scanning density of the LiDAR point cloud data. S3. Extract micro-topographic features from the LiDAR point cloud data and extract deformation features from the InSAR data; S4. Construct the micro-topographic features and the deformation features into a multi-source feature tensor, input it into a three-dimensional attention mechanism deformable convolutional network, and output the probability volume of fracture existence. S5. Automatically track the main line of continuous fracture from the fracture existence probability volume based on the minimum cost path algorithm; S6. Extract the radar line-of-sight deformation time series along the main fracture line, perform temporal clustering analysis and abrupt change point detection on the time series, and divide the fracture zone into segments with different kinematic properties based on the clustering and detection results. Furthermore, the micro-topographic features mentioned in S3 include at least one of the following: a high-precision digital elevation model generated from LiDAR point cloud data after filtering and classifying to remove vegetation points; terrain roughness; slope; curvature; terrain openness; terrain relief; and point cloud echo intensity. The deformation characteristics include at least one of annual average deformation rate, deformation gradient, and time series coherence. Furthermore, the processing procedure of the deformable convolutional network with the 3D attention mechanism described in S4 includes: Adaptive extraction of curved or discontinuous linear constructs using deformable convolutional layers; The attention-gated fusion module dynamically adjusts the fusion weights of the micro-topographic features and the deformation features based on the terrain complexity. In areas where the vegetation coverage exceeds a preset threshold, the weights of optical image features are reduced and the weights of LiDAR penetration features are increased. Furthermore, the time-series clustering analysis described in S6 employs the K-Shape clustering algorithm, and the mutation point detection uses the Pettitt test; The kinematic properties include at least one of the following: a run-slip dominant segment, a tilt-slip dominant segment, a creep segment, and a locked segment. Compared with the prior art, this application has the following beneficial effects: This invention functionally integrates the regional deformation monitoring capabilities of InSAR with the fine micro-topographic characterization capabilities of LiDAR, forming a collaborative detection mode that combines wide-area perspective and localized detailed investigation. InSAR data provides information on the surface deformation field and activity of fault zones, while LiDAR data reveals the geometric structure and micro-topographic features of the faults. The two complement each other, effectively solving the core contradictions of optical remote sensing being susceptible to vegetation obstruction, InSAR losing geometric details, and LiDAR having limited coverage. Experiments show that in densely vegetated areas, this method, by fusing topographic data obtained through LiDAR penetration of vegetation with the InSAR deformation field, can identify concealed fault structures that are difficult to detect using traditional optical remote sensing. This invention introduces a deformation information guidance mechanism during the data acquisition phase. It utilizes wide-area spaceborne InSAR to identify deformation gradient bands, automatically generates region of interest coordinates, and imports them into the UAV flight path planning system to guide LiDAR for targeted terrain-following flight and intensive scanning. Compared to traditional random or gridded scanning, this acquisition mode can reduce invalid data acquisition by more than 30%, while ensuring accurate spatial matching between InSAR and LiDAR data from the source, laying a solid foundation for subsequent feature-level fusion. This invention designs a three-dimensional deformable convolutional layer, enabling the convolutional kernel to adaptively fit the bending morphology of fractures for feature sampling, effectively solving the problem of discontinuity in fracture recognition caused by the rigid structure of traditional convolutional kernels. Simultaneously, it innovatively introduces an attention-gated fusion module, which dynamically adjusts the fusion weights of InSAR deformation features and LiDAR micro-topographic features based on terrain complexity (such as vegetation cover and slope): in densely vegetated areas, the network automatically reduces the weight of optical features and increases the weight of LiDAR penetration features; in exposed bedrock areas, it strengthens the contribution of geomorphic texture features. This design allows the model to maintain stable recognition performance in different geographical environments, and it is expected to improve fracture recognition accuracy by more than 20% compared to traditional methods in densely vegetated and complex terrain areas. Detailed Implementation

[0004] The technical solutions in the embodiments of this application will be clearly and completely described below. 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 this application. This embodiment uses the Daofu-Kangding section of the Xianshuihe Fault Zone as the test area. This section has a complex geological structure, high vegetation coverage, and strong fault activity, making it an ideal site for testing this method. Example 1: Intelligent Recognition and Kinematic Segmentation of Daofu-Kangding Segment of Xianshuihe Fault Zone Single-look complex imagery in IW mode from Sentinel-1A / B satellites covering the study area was acquired, spanning from January 2018 to December 2022, totaling 120 scenes (74 scenes in ascending orbit and 46 scenes in descending orbit). Topographic phase was removed using precise orbital data and SRTM 30m DEM. SBAS-InSAR technology was then applied to generate an annual average surface deformation rate field (20m resolution) and deformation time series. Three anomaly zones with deformation gradients exceeding 15 mm / km were identified, indicating potential active fault segments. The boundary coordinates of three deformation anomaly areas identified by InSAR were imported into the flight path planning software of the lidar system on the DJI M300 drone. The flight altitude was set to 150m, with a 70% forward overlap, a 40% lateral overlap, and a point cloud density of ≥30 points / m². A total of 5 sorties were conducted, covering an area of ​​120 km², acquiring approximately 1.5 billion raw point cloud data points. Simultaneously acquire Sentinel-2 multispectral images (10m resolution) for vegetation index calculation and subsequent dynamic weighting of the attention module. Point clouds were classified using TerraSolid software: Iteratively refine the triangular mesh filtering to separate ground points from non-ground points (vegetation, buildings). Ground points are interpolated to a 1m resolution DEM. Based on the DEM, slope, topographic relief (3×3 window), topographic opening (positive opening, negative opening), and curvature are calculated. Radiometric correction is performed on the echo intensity to generate an intensity image (1m). A spatial baseline threshold of 200m and a temporal baseline threshold of 60 days were set, generating a total of 856 interferometric pairs. Least squares unwrapping and singular value decomposition were used to invert the deformation time series, and atmospheric delay errors were removed. Annual average deformation rate maps, deformation gradient maps (using the Sobel operator), and temporal coherence maps were generated, and uniformly resampled to 1m resolution and registered with LiDAR features. The aforementioned multi-source features were unified to the same geographic coordinate system (UTM 47N) and resampled to 1m resolution to construct a five-dimensional feature tensor. This tensor contains 12 channels: LiDAR-derived channels (7): DEM, slope, topographic relief, positive aperture, negative aperture, curvature, and point cloud intensity; InSAR-derived channels (3): annual deformation rate, deformation gradient, and coherence; and optical-derived channels (2): Normalized Difference Vegetation Index (NDVI) and Modified Normalized Difference Water Index (MNDWI). The values ​​for each channel were normalized to the [0,1] interval. A 3D attention-gated deformable convolutional network was designed. This network takes a 256×256×12 feature tensor block as input, first passing through a 3D deformable convolutional layer (3×3×3 kernel size, 64 kernels) to learn deformable offsets to adaptively fit fracture and bending morphologies; then batch normalization and ReLU activation are performed. Next, a second 3D deformable convolutional layer (also 3×3×3, 64 kernels) is passed, followed by an attention-gated fusion module. This module dynamically adjusts the fusion weights of InSAR deformation features and LiDAR micro-topographic features based on the normalized vegetation index: in areas with high vegetation cover, the weights of optical image features are reduced and the weights of LiDAR penetration features are increased; in bare areas, topographic texture features are enhanced. The fused feature map size remains 256×256×64. Then, after a 2×2×2 max pooling, the feature map size is reduced to 128×128×64. Next, a third 3D deformable convolutional layer (3×3×3, 128 kernels) is applied, again with attention-gated fusion and a second max pooling, resulting in an output size of 64×64×128. Finally, a fourth 3D deformable convolutional layer (3×3×3, 256 kernels) extracts higher-level features. Subsequently, upsampling is used to restore the feature map to a size of 128×128, and it is then concatenated with the output of the second convolutional layer via skip connections to obtain a 128×128×256 feature map. Finally, a 1×1×1 3D convolutional layer is used to compress the number of channels to 1, followed by a sigmoid activation function, outputting a fracture probability map of size 128×128×1, where each pixel value represents the probability that the location belongs to a fracture zone. The locations of fault zones within the experimental area were manually interpreted as ground truth labels. The study area was divided into 512×512 sliding windows with a step size of 256 to generate training sample pairs. A total of 3264 sample pairs were obtained, which were randomly divided into training and validation sets at an 8:2 ratio. The model is trained using Dice loss and binary cross-entropy loss as the loss functions, Adam as the optimizer, with an initial learning rate of 0.001 that decays by 0.1 every 20 epochs, a batch size of 8, and 100 training epochs. After training, the model weights with the highest Dice coefficients on the validation set are saved. The feature tensor of the entire study area is cut into overlapping blocks of 256×256 (step size 128), which are then input into a trained 3D-AMDCNet for inference. The blocks are then stitched together to obtain a probability map of the existence of fractures in the entire image. The pixel value represents the probability that the point belongs to the fracture zone. Using the probabilistic graph P as the cost surface, and setting starting and ending points, an improved Dijkstra algorithm is employed to automatically trace the path with the minimum cumulative cost. During the algorithm iteration process, fracture direction constraints are added to prevent path deviation. Finally, continuous fracture principal vector lines are obtained. Within the Xianshuihe experimental area, the main fault, approximately 65 km long, was automatically traced, passing through Yala River, Kangding, and other areas, and completely connecting the fault segments that were interrupted by vegetation obstruction using traditional methods. A sampling point was set every 20m along the main fault line. The LOS deformation time curve of each sampling point was extracted from the original SBAS-InSAR time series results, for a total of 120 time nodes. Low-quality points with coherence <0.6 were removed, resulting in 3125 valid sequence points. The K-Shape clustering algorithm was used to cluster 3125 time series data. This algorithm, based on shape similarity, can identify curves with different motion patterns. The cluster number K was set to 4 (determined by the silhouette coefficient). The clustering results are as follows: Category 1 strike-slip dominant segment: The time series maintains a linear trend over a long period, with stable annual variability and no obvious seasonal fluctuations, corresponding to the left-lateral strike-slip segment of the fault. Category 2: The dominant dip-slip segment: The time series shows periodic fluctuations (possibly related to precipitation or freeze-thaw), with obvious vertical displacement components, corresponding to the dip-slip segment of the fault. Category 3 creep segment: The time series shows a continuous increase without significant acceleration or deceleration, with an average annual rate of 3-5 mm / a, corresponding to the fault creep segment. Category 4 Locked Segment: The time series is close to zero with small fluctuations, indicating that the fault is in a locked state. The Petitt abrupt change test was performed on the deformation rate sequence of each sampling point to identify the spatial locations where significant changes in rate occurred. The test results showed an abrupt change point (p<0.05) approximately 12 km north of Daofu County, where the deformation rate dropped sharply from 3.2 mm / a to 0.8 mm / a. This location corresponds to the transition boundary of the fault from a creeping segment to a locked segment. Clustering labels and abrupt change point detection results were overlaid to divide the fault zone into segments: The northern segment, Daofu-Longdengba, approximately 22 km long, is clustered as a Category 3 creep segment with an average annual velocity (LOS) of 2.5-4.0 mm / a and no abrupt changes. The middle segment, Longdengba-Yala River Estuary, approximately 18 km long, is clustered as a Category 4 locked segment with an LOS velocity <1.0 mm / a and an abrupt change point located at the northern boundary. The southern segment, Yala River Estuary-Kangding, approximately 25 km long, is clustered as alternating Categories 1 and 2, exhibiting a combination of strike-slip and dip-slip, with a velocity of 2.0-5.0 mm / a and variable orientation. The segmentation results are consistent with geological survey data and reveal new insights into creep in the northern segment and closure in the central segment. This method still maintains an F1-score of 91.5% in densely vegetated areas (NDVI>0.6), which is superior to other methods. This embodiment successfully implemented a complete process in the Xianshuihe fault zone, from multi-source data acquisition, intelligent fusion identification, fault tracking to kinematic feature segmentation, verifying the feasibility and superiority of the method of this invention. The output fault zone kinematic property segmentation map can be directly used for regional seismic hazard assessment and site selection of major projects. Example 2: Adaptability verification under different geographical environments To verify the generalization ability of this method, experiments were conducted in the western segment of the Haiyuan Fault Zone (loess-covered area) on the northeastern edge of the Tibetan Plateau and the southern segment of the Longmenshan Fault Zone in western Sichuan (high mountain and canyon area), with the same procedure as in Example 1. The results show that in the loess-covered area (sparse vegetation but fragmented terrain), the attention module automatically reduces the weight of LiDAR features and strengthens InSAR deformation features, achieving an accuracy of 91.2%. In the high mountain and canyon area (steep terrain and dense vegetation), the attention module increases the weight of LiDAR penetration features, achieving an accuracy of 89.7%. This demonstrates the adaptability of this method to different geographical environments. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A system for intelligent identification and segmentation of linear fault zone structures based on the fusion of multi-source remote sensing and LiDAR point clouds, characterized in that, include: The data acquisition module is configured to acquire spaceborne synthetic aperture radar interferometry (InSAR) data and UAV lidar (LiDAR) point cloud data covering the target area; The deformation anomaly guidance module is configured to identify regional-scale surface deformation anomaly areas based on the InSAR data, and to guide the acquisition range and encryption scanning density of the LiDAR point cloud data with the deformation anomaly areas. The multi-source feature construction module is configured to extract micro-topographic features from the LiDAR point cloud data and deformation features from the InSAR data. The feature fusion and recognition module is configured to construct a multi-source feature tensor from the micro-topographic features and the deformation features, input it into a three-dimensional attention mechanism deformable convolutional network, and output a probability volume of fracture existence. The fracture tracking module is configured to automatically track continuous fracture main lines from the fracture existence probability volume based on the minimum cost path algorithm. The kinematic feature segmentation module is configured to extract the radar line-of-sight deformation time series along the main fracture line, perform temporal clustering analysis and abrupt change point detection on the time series, and divide the fracture zone into segments with different kinematic properties based on the clustering and detection results.

2. The system according to claim 1, characterized in that, The micro-topographic features include at least one of the following: a high-precision digital elevation model generated from LiDAR point cloud data after filtering, classification and removal of vegetation points; terrain roughness; slope; curvature; terrain openness; terrain relief; and point cloud echo intensity. The deformation characteristics include at least one of annual average deformation rate, deformation gradient, and time series coherence.

3. The system according to claim 1, characterized in that, The three-dimensional attention mechanism deformable convolutional network includes: Deformable convolutional layers, configured to adaptively extract curved or discontinuous linear constructs; The attention-gated fusion module is configured to dynamically adjust the fusion weights of the micro-topographic features and the deformation features according to the terrain complexity. In areas where the vegetation coverage exceeds a preset threshold, the weights of optical image features are reduced and the weights of LiDAR penetration features are increased.

4. The system according to claim 1, characterized in that, The temporal clustering algorithm configured in the kinematic feature segmentation module includes the K-Shape clustering algorithm, and the mutation point detection algorithm includes the Pettitt test. The kinematic properties include at least one of the following: a run-slip dominant segment, a tilt-slip dominant segment, a creep segment, and a locked segment.

5. A method for intelligent identification and segmentation of linear fault zone structures based on the fusion of multi-source remote sensing and LiDAR point clouds, characterized in that, Includes the following steps: S1. Acquire spaceborne InSAR data and UAV LiDAR point cloud data covering the target area; S2. Identify regional-scale surface deformation anomaly areas based on the InSAR data, and use the deformation anomaly areas to guide the acquisition range and encryption scanning density of the LiDAR point cloud data. S3. Extract micro-topographic features from the LiDAR point cloud data and extract deformation features from the InSAR data; S4. Construct the micro-topographic features and the deformation features into a multi-source feature tensor, input it into a three-dimensional attention mechanism deformable convolutional network, and output the probability volume of fracture existence. S5. Automatically track the main line of continuous fracture from the fracture existence probability volume based on the minimum cost path algorithm; S6. Extract the radar line-of-sight deformation time series along the main fracture line, perform temporal clustering analysis and abrupt change point detection on the time series, and divide the fracture zone into segments with different kinematic properties based on the clustering and detection results.

6. The method according to claim 5, characterized in that, The micro-topographic features mentioned in step 3 include at least one of the following: a high-precision digital elevation model generated from LiDAR point cloud data after filtering, classification and removal of vegetation points; terrain roughness; slope; curvature; terrain openness; terrain relief; and point cloud echo intensity. The deformation characteristics include at least one of annual average deformation rate, deformation gradient, and time series coherence.

7. The method according to claim 5, characterized in that, The processing procedure of the 3D attention mechanism deformable convolutional network in step 4 includes: Adaptive extraction of curved or discontinuous linear constructs is achieved through deformable convolutional layers.

8. The method according to claim 5, characterized in that, The temporal clustering analysis in step 6 uses the K-Shape clustering algorithm, and the mutation point detection uses the Pettitt test. The kinematic properties include at least one of the following: a run-dominant segment, a tilt-dominant segment, a creep segment, and a locked segment.