InSAR deformation anomaly extraction method and system based on multi-modal fusion and vector optimization

Abstract

The application provides an InSAR deformation anomaly extraction method and system based on multi-modal fusion and vector optimization, relates to the technical field of artificial intelligence and geological disasters, and comprises the following steps: acquiring regional multi-source remote sensing data, calculating surface deformation phase data, terrain relief, slope, vegetation index, water index and soil index, and resampling to obtain multi-source remote sensing full-band data; manually interpreting the InSAR surface deformation anomaly range of the region based on expert knowledge to obtain vector data and binary mask data of the InSAR surface deformation anomaly range; constructing a double-branch multi-modal feature fusion network, extracting InSAR surface deformation anomaly information based on the binary mask data and the multi-source remote sensing full-band data; constructing a dynamic vector optimization network, and optimizing the boundary of the InSAR surface deformation anomaly information based on the multi-source remote sensing full-band data to dynamically convert into a vector result. The application realizes seamless connection from remote sensing images to GIS vectors.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and geological disaster technology, and relates to an InSAR deformation anomaly extraction method and system based on multimodal fusion and vector optimization. Background Technology

[0002] In recent years, with the rapid development of remote sensing technology, the integrated "air-space-ground" monitoring system has gradually become the mainstream method for early identification of landslide hazards. Integrated "air-space-ground" remote sensing technology, represented by InSAR (Synthetic Aperture Radar Interferometry), has demonstrated an irreplaceable role in landslide hazard identification due to its unique advantages—achieving large-scale (single-scene imagery covering hundreds to thousands of square kilometers), high-precision (millimeter-level deformation monitoring), and all-weather (unaffected by clouds, rain, or fog) continuous observation without ground-based sensor deployment. Compared to traditional ground surveys (low efficiency and long cycle), GNSS monitoring (point-based monitoring with limited coverage), and UAV aerial surveys (high resolution but poor timeliness), InSAR technology can quickly capture minute surface deformation signals, making it particularly suitable for the "general survey" stage of potential landslide bodies over a wide area, providing crucial clues for subsequent focused investigations.

[0003] However, despite the progress made by InSAR technology in wide-area landslide identification, its technical bottlenecks in operational applications remain significant, mainly in the following three aspects: First, low identification efficiency. Deformation anomaly information extraction relies primarily on manual visual interpretation, with a single data source and severe noise interference, making it difficult to meet the operational needs of large-scale, efficient identification. Second, weak discrimination capability. The causes of surface deformation are complex and diverse. How to effectively integrate multi-source remote sensing information such as topography and optics to accurately identify deformation anomalies related to landslides still requires in-depth research. Third, poor practicality of results. Existing deep learning-based identification methods produce output results with rough edges and are mostly in raster format, which easily leads to secondary information loss during subsequent vectorization, failing to meet the needs of accurate landslide hazard location and mapping.

[0004] In summary, overcoming the bottlenecks in efficiency, differentiation, and practicality, and constructing a full-chain technical system of "efficient extraction - accurate identification - practical output" is key to achieving early detection and early warning of landslide hazards, and is also an important breakthrough in promoting the transformation of geological disaster prevention and control from "passive response" to "proactive prevention and control". Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an InSAR deformation anomaly extraction method and system based on multimodal fusion and vector optimization. By constructing a dual-branch deep network, it achieves intelligent fusion of InSAR deformation phase with multimodal data such as optical and topographic data, effectively suppressing noise interference at the feature level and improving recognition robustness.

[0006] To achieve the above objectives, this invention provides an InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization, comprising: Acquire regional multi-source remote sensing data, calculate surface deformation phase data, topographic relief, slope, vegetation index, water index and soil index, and resample them to obtain multi-source remote sensing full-band data; Based on expert knowledge, the InSAR surface deformation anomaly range in this area was manually interpreted to obtain vector data and binarized mask data of the InSAR surface deformation anomaly range. A dual-branch multimodal feature fusion network is constructed, and the binarized mask data is used as the semantic segmentation ground truth. The dual-branch multimodal feature fusion network initially extracts InSAR surface deformation anomaly information based on the semantic segmentation ground truth and the multi-source remote sensing full-band data. A dynamic vector optimization network is constructed to obtain shallow spatial features based on the surface deformation phase data in the multi-source remote sensing full-band data. Based on the shallow spatial features, the InSAR surface deformation anomaly information is optimized at the boundary and dynamically converted into vector results. Based on the vector data, vector boundary anchor points are generated. These vector boundary anchor points are used as the truth values ​​for dynamic vectorization operations, and their differences from the vector results are measured to optimize the dynamic vector optimization network.

[0007] As a further improvement of the present invention, regional multi-source remote sensing data is acquired, including Sentinel-1 radar data, terrain data and Sentinel-2 data; The surface deformation phase data of the demonstration area is calculated based on the Sentinel-1 radar data, the topographic relief and slope are calculated based on the topographic data, and the vegetation index, water index and soil index are calculated based on the Sentinel-2 data. The resampled data are arranged in sequence according to the surface deformation phase data, topographic relief, slope, vegetation index, water index and soil index to form multi-source remote sensing full-band data.

[0008] As a further improvement of the present invention, the surface deformation phase data of the demonstration area is calculated based on the Sentinel-1 radar data, including: Data from the Sentinel-1 radar was processed using Stacking-InSAR multi-temporal analysis technology, including splicing and registering sub-strip data, selecting appropriate image pairs for differential interferometry calculation, phase unwrapping, phase superposition processing, and removal of trend stripes, to obtain surface deformation phase data of the demonstration area.

[0009] As a further improvement of the present invention, the formula for calculating the terrain relief is as follows: In the formula, Midpoint of the digital elevation model The elevation value, N is the number of grid points in the window, z mean This is the average elevation of all points within the window.

[0010] As a further improvement of the present invention, the formula for calculating the vegetation index is as follows: In the formula, NIR represents the near-infrared band in Sentinel-2 data, and Red represents the red band in Sentinel-2 data; The formula for calculating the water quality index is as follows: In the formula, Green represents the green light band in Sentinel-2 data, and SWIR represents the shortwave infrared band in Sentinel-2 data; The formula for calculating the soil index is as follows: In the formula, Blue represents the blue light band in Sentinel-2 data.

[0011] As a further improvement of the present invention, the dual-branch multimodal feature fusion network includes a single-modal feature extraction branch network, a multimodal feature extraction branch network, and a dual-branch fusion network; The single-modal feature extraction branch network extracts anomalous boundary features based on the surface deformation phase data in the multi-source remote sensing full-band data. The multi-modal feature extraction branch network distinguishes InSAR surface deformation anomaly features from noise in the data based on the multi-source remote sensing full-band data. The dual-branch fusion network fuses the anomalous boundary features with the InSAR surface deformation anomaly features to obtain the InSAR surface deformation anomaly information.

[0012] As a further improvement of the present invention, the loss function of the dual-branch multimodal feature fusion network adopts the cross-entropy function. The cross-entropy function establishes a mapping from the surface deformation phase data in the multi-source remote sensing full-band data to the semantic segmentation ground truth based on the InSAR surface deformation anomaly information and the semantic segmentation ground truth. The specific formula is as follows: In the formula, The cross-entropy loss value is a scalar. The smaller the value, the closer the InSAR surface deformation anomaly information Pred_0 is to the semantic segmentation ground truth value Seg_GT. This represents the total number of pixels in a feature map; Indicates the total number of categories; Indicates the pixel index of the current summation; Indicates the category index of the current summation; The dual-branch multimodal feature fusion network is optimized based on this loss function.

[0013] As a further improvement of the present invention, the dual-branch fusion network includes a dynamic fusion attention module; The fusion attention module uses a lightweight neural network in the channel attention part. The lightweight neural network dynamically predicts the compression ratio of the channel dimension features in the anomaly boundary features and InSAR surface deformation anomaly features, compressing the number of channels to a preset range. In the spatial attention section, the fusion attention module fuses the average pooling, max pooling, and mean structure of the original input features with the anomaly boundary features and InSAR surface deformation anomaly features to construct a comprehensive feature map, which is then convolved to generate a spatial attention map. The fusion attention module fuses the compressed channel attention and spatial attention maps to generate InSAR surface deformation anomaly information.

[0014] As a further improvement of the present invention, a dynamic vector optimization network is constructed to obtain shallow spatial features based on the surface deformation phase data in the multi-source remote sensing full-band data. Based on these shallow spatial features, the InSAR surface deformation anomaly information is boundary-optimized and dynamically converted into vector results, including: The shallow spatial features are added to the InSAR surface deformation anomaly information to obtain the optimized segmentation result; The boundary features of the optimized segmentation result of the convolution multiplication are characterized by linear features to obtain a boundary feature map; The optimized segmentation result is subjected to feature encoding, initial vertex prediction, and sequence decoding to generate an ordered vector vertex sequence, which is the vector result.

[0015] As a further improvement of the present invention, the boundary features of the optimized segmentation result are characterized by using linear features to characterize the convolution, thereby obtaining a boundary feature map; including: The horizontal and vertical boundary features of the optimized segmentation result are extracted by using two linear features to characterize the convolution kernel. The horizontal and vertical boundary features are fused to obtain a complete boundary feature map; Specifically, a cross-entropy loss function is established between the optimized segmentation result and the semantic segmentation ground value to constrain the segmentation accuracy; Establish a boundary loss function to constrain the consistency between the boundary feature map of the optimized segmentation result and the boundary feature map of the semantic segmentation ground truth.

[0016] This invention also provides an InSAR deformation anomaly extraction system based on multimodal fusion and vector optimization, including: a multi-source data acquisition module, a model truth acquisition module, a surface deformation information extraction module, and a surface deformation vector conversion module; The multi-source data acquisition module includes: Acquire regional multi-source remote sensing data, calculate surface deformation phase data, topographic relief, slope, vegetation index, water index and soil index, and resample them to obtain multi-source remote sensing full-band data; The model truth value acquisition module includes: Based on expert knowledge, the InSAR surface deformation anomaly range in this area was manually interpreted to obtain vector data and binarized mask data of the InSAR surface deformation anomaly range. The surface deformation information extraction module includes: A dual-branch multimodal feature fusion network is constructed, and the binarized mask data is used as the semantic segmentation ground truth. The dual-branch multimodal feature fusion network initially extracts InSAR surface deformation anomaly information based on the semantic segmentation ground truth and the multi-source remote sensing full-band data. The surface deformation vector conversion module includes: A dynamic vector optimization network is constructed to obtain shallow spatial features based on the surface deformation phase data in the multi-source remote sensing full-band data. Based on the shallow spatial features, the InSAR surface deformation anomaly information is optimized at the boundary and dynamically converted into vector results. Based on the vector data, vector boundary anchor points are generated. These vector boundary anchor points are used as the truth values ​​for dynamic vectorization operations, and their differences from the vector results are measured to optimize the dynamic vector optimization network.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention first constructs a dual-branch deep network to achieve intelligent fusion of InSAR deformation phase data with multimodal data such as optical and topographic data, effectively suppressing noise interference at the feature level and improving recognition robustness. Simultaneously, addressing the challenge of ambiguous boundaries in the recognition results, this invention innovatively designs a linear feature-characterized convolution combined with a boundary loss function, optimizing the segmentation results from "coarse" to "fine." Based on this, an end-to-end sequence vectorization network directly decodes the optimized boundaries into structurally regular vector polygons, solving the two major pain points of inaccurate boundary positioning and information loss during secondary transformation. This invention effectively improves the efficiency and effectiveness of wide-area InSAR deformation anomaly recognition, and is of great significance for promoting the transformation of geological disaster remote sensing monitoring from "human judgment" to "intelligence."

[0018] This invention addresses the challenge of ambiguous boundaries in identification results by innovatively designing a linear feature-characterized convolution combined with a boundary loss function, achieving optimization of segmentation results from "coarse" to "fine." Building upon this, an end-to-end sequence vectorization network directly decodes the optimized boundaries into structurally regular vector polygons, resolving two major pain points: inaccurate boundary localization and information loss during secondary transformation. This invention effectively improves the efficiency and effectiveness of wide-area InSAR deformation anomaly identification, and is of great significance for promoting the transformation of geological disaster remote sensing monitoring from "human judgment" to "intelligence."

[0019] This invention presents an InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization, featuring three core innovations: First, it constructs a dual-branch deep fusion network, achieving intelligent collaboration of multimodal remote sensing data through a dynamic attention mechanism, significantly improving feature discrimination and model robustness. Second, it establishes a "segmentation-optimization" cascade architecture, utilizing linear features to characterize convolution and boundary loss functions to achieve refined boundary processing, effectively improving edge localization accuracy. Third, it proposes an end-to-end dynamic vector conversion technology, completely abandoning traditional post-processing procedures and achieving seamless integration from remote sensing imagery to GIS vectors. This method addresses the pain points of existing technologies, such as blurred boundaries, high false alarm rates, and cumbersome processes, providing a solution for efficient and rapid monitoring of wide-area surface deformation. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the basic structure of a dual-branch multimodal feature fusion network disclosed in one embodiment of the present invention; Figure 2 This is a schematic diagram of the basic structure of a convolutional unit in a dual-branch fusion network disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the dynamic fusion attention module structure in a dual-branch fusion network disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the semantic segmentation process based on a dual-branch multimodal feature fusion network disclosed in one embodiment of the present invention; Figure 5 This is a schematic diagram of a process for dynamic vectorization extraction based on a dynamic vector optimization network, as disclosed in one embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, 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.

[0022] The present invention will now be described in further detail with reference to the accompanying drawings: like Figures 1-5 As shown, the InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization provided by this invention includes: S1. Acquire regional multi-source remote sensing data, calculate surface deformation phase data (GD), topographic relief (TRI), slope (SP), vegetation index (NDVI), water index (NDWI) and soil index (BSI), and resample them to obtain multi-source remote sensing full-band data. in, Regional multi-source remote sensing data, including Sentinel-1 radar data, terrain data, and Sentinel-2 data; The surface deformation phase data of the demonstration area were calculated based on the Sentinel-1 radar data, the topographic relief and slope were calculated based on the topographic data, and the vegetation index, water index and soil index were calculated based on the Sentinel-2 data. The resampled data are arranged in sequence according to the surface deformation phase data, topographic relief, slope, vegetation index, water index and soil index to form multi-source remote sensing full-band data. For example, the data are resampled at 30 meters and arranged in sequence according to GD, TRI, SP, NDVI, NDWI and BSI to form multi-source remote sensing full-band data (a total of 8 channels, of which GD has 3 channels).

[0023] Specifically, Data from the Sentinel-1 radar was processed using Stacking-InSAR multi-temporal analysis technology, including splicing and registering sub-strip data, selecting appropriate image pairs for differential interferometry calculation, phase unwrapping, phase superposition processing, and removal of trend stripes, to obtain surface deformation phase data of the demonstration area.

[0024] The formula for calculating terrain relief is: In the formula, Midpoint of the digital elevation model The elevation value, N is the number of grid points in the window, z mean This is the average elevation of all points within the window.

[0025] The formula for calculating the vegetation index (NDVI) is: In the formula, NIR represents the near-infrared band (B8: 842 nm) in Sentinel-2 data, and Red represents the red band (B4: 665 nm) in Sentinel-2 data. The formula for calculating the National Water Index (NDVI) is: In the formula, Green represents the green light band (B3: 560nm) in Sentinel-2 data, and SWIR represents the shortwave infrared band (B11: 1610nm / B12: 2190nm) in Sentinel-2 data. The formula for calculating the Soil Index (BSI) is as follows: In the formula, Blue represents the blue light band (B2: 490nm) in the Sentinel-2 data.

[0026] S2. Based on expert knowledge, the InSAR surface deformation anomaly range of the region is manually interpreted to obtain vector data and binarized mask data of the InSAR surface deformation anomaly range. in, Vector data is used to generate vector boundary anchor points, which serve as the ground truth (Opt_GT) for subsequent dynamic vectorization operations, while binarized mask data serves as the ground truth (Seg_GT) for subsequent semantic segmentation.

[0027] Specifically, The generation of the binarized mask is derived from manually drawn vector data and consists of 0s and 1s, where negative samples are 0s and positive samples are 1s.

[0028] Vector boundary anchors are a simplified representation of the extent of InSAR surface deformation anomalies.

[0029] S3. Construct a dual-branch multimodal feature fusion network (BMFFNet), using binarized mask data as the semantic segmentation ground truth. The dual-branch multimodal feature fusion network initially extracts InSAR surface deformation anomaly information based on the semantic segmentation ground truth and multi-source remote sensing full-band data. in, The dual-branch multimodal feature fusion network includes a single-modal feature extraction branch network, a multimodal feature extraction branch network, and a dual-branch fusion network; The single-modal feature extraction branch network extracts anomalous boundary features from the surface deformation phase data in the multi-source remote sensing full-band data. The multi-modal feature extraction branch network distinguishes InSAR surface deformation anomaly features from noise in the data based on the multi-source remote sensing full-band data. The dual-branch fusion network fuses the anomalous boundary features with the InSAR surface deformation anomaly features to obtain InSAR surface deformation anomaly information.

[0030] Furthermore, The network structure of BMFFNet is as follows: Figure 1As shown, the single-modal feature extraction branch network and the multi-modal feature extraction branch network have the same structure, both adopting a multi-scale feature cross-fusion structure similar to Unet++, but their input data differs. The input of the single-modal feature extraction branch network is 3-channel GD data, and its main function is to extract the boundary features of anomalies from InSAR surface deformation phase data; the input of the multi-modal feature extraction branch is multi-source remote sensing full-band data with 8 channels, and its main function is to use the rich information of multi-modal data to help distinguish InSAR surface deformation anomalies from noise in the data.

[0031] The two-branch fusion network in the BMFFNet network is responsible for fusing the features from the two branches mentioned earlier. Its structure is similar to the two branches, but the convolutional units are different. The convolutional units of the first two branches are inherited from the original Unet++ network, while the convolutional units of the two-branch fusion network are constructed as follows: Figure 2 As shown, its key feature is the introduction of a Dynamic Fusion Attention (DFAM) module to learn the weight relationships between different features, thereby improving the fusion effect. The specific structure of this module is as follows: Figure 3 As shown, it can simultaneously integrate spatial attention and channel attention.

[0032] Specifically, (1) In the channel attention part, DFAM uses a lightweight neural network to dynamically predict the optimal compression ratio of channel dimension features, thereby compressing the number of channels to a preset range and achieving adaptive adjustment of the compression degree. In the spatial attention part, DFAM constructs a comprehensive feature map by fusing the results of average pooling, max pooling, and the mean of the original input features; subsequently, a spatial attention map is generated through a two-layer convolutional structure. This design significantly enhances the model's ability to capture global deformation features without significantly increasing the number of model parameters.

[0033] (2) The BMFFNet network has two inputs: 3-channel GD data and 8-channel multi-source remote sensing full-band data. These are processed by two separate network branches. The features are then fused in the dual-branch fusion network to output a preliminary semantic segmentation result, Pred_0 (e.g., ...). Figure 4 As shown in the image, this data consists of floating-point numbers between 0 and 1, representing the probability value of a correct recognition result. The ground truth value during network training is Seg_GT. The loss function is a standard cross-entropy loss, with Pred_0 and Seg_GT as inputs. Its function is to establish a mapping from GD to Seg_GT, allowing the model to extract effective features from the GD data as much as possible. The specific formula is as follows: In the formula, The cross-entropy loss value is a scalar. The smaller the value, the closer the InSAR surface deformation anomaly information Pred_0 is to the semantic segmentation ground truth value Seg_GT. This represents the total number of pixels in a feature map; Indicates the total number of categories; Indicates the pixel index of the current summation; Indicates the category index of the current summation; This loss function is used to optimize the bi-branch multimodal feature fusion network.

[0034] S4. Construct a Dynamic Vector Optimization Network (DVONet) to obtain shallow spatial features based on surface deformation phase data from multi-source remote sensing full-band data. Based on the shallow spatial features, optimize the boundaries of InSAR surface deformation anomaly information and dynamically convert it into vector results. in, The optimized segmentation result is obtained by adding shallow spatial features with InSAR surface deformation anomaly information; The boundary features of the convolutional multi-optimization segmentation result Seg_GD are characterized using linear features to obtain a boundary feature map; The optimized segmentation results are subjected to feature encoding, initial vertex prediction, and sequence decoding to generate an ordered vector vertex sequence Opt_GD. The vector vertex sequence is the vector result, realizing the end-to-end conversion from pixel-level segmentation map to structured vector data.

[0035] Furthermore, linear features are used to characterize the boundary features of the convolutional segmentation to obtain a boundary feature map; including: Two linear features are used to characterize the convolution kernel, which extracts the horizontal and vertical boundary features of the optimized segmentation result respectively; The horizontal and vertical boundary features are fused to obtain a complete boundary feature map; Among them, a cross-entropy loss function is established between the optimized segmentation result and the semantic segmentation ground value to constrain the segmentation accuracy; Establish a boundary loss function to constrain the consistency between the boundary feature map of the optimized segmentation result and the boundary feature map of the semantic segmentation ground truth.

[0036] Specifically, (1) The role of the DVONet network is to use linear features to characterize convolution to extract boundary information, and combine it with the edge loss function to improve the fineness of the boundary recognition result of the model. The overall technical process is as follows: Figure 5 As shown.

[0037] The network takes a 3-channel GD as input and the Pred_0 obtained in the previous step, with Opt_GT and Seg_GT as reference ground truth values. The 3-channel GD data is processed by two sets of convolutional units and then added to Pred_0 to obtain the optimized segmentation result Seg_GD. Linear feature descriptors K_h and K_v are designed to extract boundary features in the horizontal and vertical directions, respectively. K_h is a horizontal gradient convolution kernel, enhancing vertical edge features; K_v is a vertical gradient convolution kernel, enhancing horizontal edge features. The edge responses in the two directions are fused to obtain a boundary feature map (K_c) with directional consistency and structural integrity. The specific calculation formula is as follows: K_h = [ -1, -2, -1 ] [ 0, 0, 0 ], [ 1, 2, 1 ] K_v = [ -1, 0, 1 ] [ -2, 0, 2 ], [ -1, 0, 2 ] K_c = Where G_h and G_v are gradient feature maps calculated from K_h and K_v, respectively; During network training, Seg_GD and Seg_GT are respectively input into the linear feature characterization convolution, and the corresponding linear fusion features (K_cp and K_cs) are output.

[0038] This process is jointly supervised by Loss_1 and Loss_2. Loss_1 is the cross-entropy loss between Seg_GD and Seg_GT (consistent with the loss function used in step 3), used to constrain segmentation accuracy; Loss_2 is the boundary loss function, used to optimize the consistency between K_cp and K_c. Loss_2 consists of two parts: feature alignment loss and boundary structure loss, and the specific calculation formula is shown below: in and To balance the weighting coefficients, recommended values ​​are 0.7 and 0.3, respectively. Φ(·) represents the boundary feature extraction function. and Let K_cp and K_cs represent the boundary feature maps K_cp and K_cs obtained from Seg_GD and Seg_GT, respectively, in pixels. The mean of features within a local window centered on the locator. and These represent the feature standard deviations of feature maps K_cp and K_cs within their respective local windows. This represents the feature covariance of feature maps K_cp and K_cs within the corresponding local window. c1 and c2 represent constants to prevent division by zero errors.

[0039] (2) The feature encoding network consists of three sets of convolutional units, whose main function is to convert Seg_GD into a deep feature representation rich in semantic information. The initial vertex prediction module is mainly used to locate the starting vertex of the polygon from the deep feature map. This module first converts the feature map into a vertex heatmap through 1×1 convolution, and then determines the initial vertex coordinates through spatial softmax operation. The specific calculation formula is as follows: H = Sigmoid(Conv1×1(F, 1)) (x0, y0) = Softmax(H) Where Conv1×1 represents 1×1 convolution, Sigmoid and Softmax are activation functions, and the output (x0, y0) is the initial vertex coordinates normalized to the range [0,1].

[0040] The sequence decoding module works by recursively predicting all subsequent vertices needed to construct a complete polygon, starting from the initial vertex, in an autoregressive manner. This module is built upon convolutionally gated recurrent units (GRUs) and iteratively fuses image depth features, historical vertex information, and the initial vertex positions to gradually deduce the complete boundary contour. At each time step t, the module receives the current information and predicts the coordinates v_t of the next vertex. By executing this loop T times, an ordered sequence from the initial vertex to the last vertex is generated. Finally, by connecting the start and end points of the sequence, a closed vector polygon is formed. The output of this dynamic vectorization process, the ordered set of anchor points ultimately used to construct the vector polygon, is expressed by the following formula: Opt_GD = {v0,v1,v2,...,v t , v0} in, v0 is the initial vertex, {v0,v1,v2,...,v t The sequence of intermediate vertices generated by the sequence decoding module is looped, and v0 is added again at the end to ensure that the polygon is closed.

[0041] S5. Generate vector boundary anchor points based on vector data, use the vector boundary anchor points as the truth values ​​of dynamic vectorization operations, measure the difference between them and the vector results, and optimize the dynamic vector optimization network.

[0042] in, During the vectorization process, a composite loss function, Loss_3, is constructed to measure the difference between the predicted vector sequence Opt_GD and the ground value Opt_GT. The calculation formula is as follows: Loss3= α·D_Chamfer + β·L_sequence + γ·L_closure in, D_Chamfer represents the chamfer distance, used to measure the overall matching degree between the predicted point set and the ground truth point set. Its calculation formula is as follows: m represents the prediction point set The total number of vertices in the array Represents the set of truth points The total number of vertices in the array. Representative prediction point set The first in vertex coordinates Represents the set of truth points The first in The coordinates of each vertex.

[0043] L_sequence represents the sequence order loss, which is used to ensure a reasonable spacing between adjacent vertices. Its calculation formula is as follows: 2 Where T is the total length of the vertex sequence. Indicates the first [number]th ... The coordinates of the vertices, Indicates the first [number]th ... The coordinates of each vertex.

[0044] L_closure represents the polygon closure loss, which promotes the formation of a closed polygon by aligning the first and last vertices. The calculation formula is as follows: To predict the coordinates of the last vertex in the vertex sequence, To predict the initial vertices in the vertex sequence.

[0045] Both the BMFFNet and DVONet networks in this invention can be trained end-to-end. Training uses the Adam optimizer with an initial learning rate of 0.0001. Training is recommended to be performed on a GPU server equipped with an NVIDIA Tesla A100 GPU, using the PyTorch deep learning framework. Model training consists of two phases: first, the BMFFNet network is trained independently until convergence; then, the output semantic segmentation map of BMFFNet is used as the input to DVONet, and the weights of BMFFNet and DVONet are fixed for joint training.

[0046] This invention also provides an InSAR deformation anomaly extraction system based on multimodal fusion and vector optimization, which implements the above-mentioned InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization, including: a multi-source data acquisition module, a model truth value acquisition module, a surface deformation information extraction module, and a surface deformation vector conversion module; The multi-source data acquisition module includes: Acquire regional multi-source remote sensing data, calculate surface deformation phase data, topographic relief, slope, vegetation index, water index and soil index, and resample them to obtain multi-source remote sensing full-band data; The model truth value acquisition module includes: Based on expert knowledge, the InSAR surface deformation anomaly range in this area was manually interpreted to obtain vector data and binarized mask data of the InSAR surface deformation anomaly range. The surface deformation information extraction module includes: A dual-branch multimodal feature fusion network was constructed, using binarized mask data as the semantic segmentation ground truth. Based on the semantic segmentation ground truth and multi-source remote sensing full-band data, the dual-branch multimodal feature fusion network initially extracted InSAR surface deformation anomaly information. The surface deformation vector conversion module includes: A dynamic vector optimization network is constructed to obtain shallow spatial features based on surface deformation phase data from multi-source remote sensing full-band data. Based on the shallow spatial features, the boundary of InSAR surface deformation anomaly information is optimized and dynamically converted into vector results. Vector boundary anchors are generated based on vector data. These anchors are then used as the ground truth for dynamic vectorization operations. The difference between these anchors and the vector results is measured, and the dynamic vector optimization network is optimized.

[0047] Advantages of this invention: This invention first constructs a dual-branch deep network to achieve intelligent fusion of InSAR deformation phase data with multimodal data such as optical and topographic data, effectively suppressing noise interference at the feature level and improving recognition robustness. Simultaneously, addressing the challenge of ambiguous boundaries in the recognition results, this invention innovatively designs a linear feature-characterized convolution combined with a boundary loss function, optimizing the segmentation results from "coarse" to "fine." Based on this, an end-to-end sequence vectorization network directly decodes the optimized boundaries into structurally regular vector polygons, solving the two major pain points of inaccurate boundary positioning and information loss during secondary transformation. This invention effectively improves the efficiency and effectiveness of wide-area InSAR deformation anomaly recognition, and is of great significance for promoting the transformation of geological disaster remote sensing monitoring from "human judgment" to "intelligence."

[0048] This invention addresses the challenge of ambiguous boundaries in identification results by innovatively designing a linear feature-characterized convolution combined with a boundary loss function, achieving optimization of segmentation results from "coarse" to "fine." Building upon this, an end-to-end sequence vectorization network directly decodes the optimized boundaries into structurally regular vector polygons, resolving two major pain points: inaccurate boundary localization and information loss during secondary transformation. This invention effectively improves the efficiency and effectiveness of wide-area InSAR deformation anomaly identification, and is of great significance for promoting the transformation of geological disaster remote sensing monitoring from "human judgment" to "intelligence."

[0049] This invention presents an InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization, featuring three core innovations: First, it constructs a dual-branch deep fusion network, achieving intelligent collaboration of multimodal remote sensing data through a dynamic attention mechanism, significantly improving feature discrimination and model robustness. Second, it establishes a "segmentation-optimization" cascade architecture, utilizing linear features to characterize convolution and boundary loss functions to achieve refined boundary processing, effectively improving edge localization accuracy. Third, it proposes an end-to-end dynamic vector conversion technology, completely abandoning traditional post-processing procedures and achieving seamless integration from remote sensing imagery to GIS vectors. This method addresses the pain points of existing technologies, such as blurred boundaries, high false alarm rates, and cumbersome processes, providing a solution for efficient and rapid monitoring of wide-area surface deformation.

[0050] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization, characterized in that, include: Acquire regional multi-source remote sensing data, calculate surface deformation phase data, topographic relief, slope, vegetation index, water index and soil index, and resample them to obtain multi-source remote sensing full-band data; Based on expert knowledge, the InSAR surface deformation anomaly range in this area was manually interpreted to obtain vector data and binarized mask data of the InSAR surface deformation anomaly range. A dual-branch multimodal feature fusion network is constructed, using the binarized mask data as the semantic segmentation ground truth. The dual-branch multimodal feature fusion network initially extracts InSAR surface deformation anomaly information based on the semantic segmentation ground truth and the multi-source remote sensing full-band data. The dual-branch multimodal feature fusion network includes a single-modal feature extraction branch network, a multimodal feature extraction branch network, and a dual-branch fusion network. The single-modal feature extraction branch network extracts anomaly boundary features based on the surface deformation phase data in the multi-source remote sensing full-band data. The multimodal feature extraction branch network distinguishes InSAR surface deformation anomaly features from noise in the data based on the multi-source remote sensing full-band data. The dual-branch fusion network fuses the anomaly boundary features with the InSAR surface deformation anomaly features to obtain the InSAR surface deformation anomaly information. A dynamic vector optimization network is constructed to obtain shallow spatial features based on surface deformation phase data from the multi-source remote sensing full-band data. Based on these shallow spatial features, the InSAR surface deformation anomaly information is optimized at the boundary and dynamically converted into vector results. This includes: adding the shallow spatial features to the InSAR surface deformation anomaly information to obtain an optimized segmentation result; characterizing the boundary features of the optimized segmentation result using linear features to characterize convolution, obtaining a boundary feature map; and performing feature encoding, initial vertex prediction, and sequence decoding on the optimized segmentation result to generate an ordered sequence of vector vertices, where the vector vertex sequence is the vector result. Based on the vector data, vector boundary anchor points are generated. These vector boundary anchor points are used as the truth values ​​for dynamic vectorization operations, and their differences from the vector results are measured to optimize the dynamic vector optimization network.

2. The InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization according to claim 1, characterized in that: Acquire regional multi-source remote sensing data, including Sentinel-1 radar data, terrain data, and Sentinel-2 data; The surface deformation phase data of the demonstration area is calculated based on the Sentinel-1 radar data, the topographic relief and slope are calculated based on the topographic data, and the vegetation index, water index and soil index are calculated based on the Sentinel-2 data. The resampled data are arranged in sequence according to the surface deformation phase data, topographic relief, slope, vegetation index, water index and soil index to form multi-source remote sensing full-band data.

3. The InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization according to claim 2, characterized in that: Data from the Sentinel-1 radar was processed using Stacking-InSAR multi-temporal analysis technology, including splicing and registering sub-strip data, selecting appropriate image pairs for differential interferometry calculation, phase unwrapping, phase superposition processing, and removal of trend stripes, to obtain surface deformation phase data of the demonstration area. The formula for calculating the terrain relief is: In the formula, Midpoint of the digital elevation model The elevation value, N is the number of grid points in the window, z mean This is the average elevation of all points within the window.

4. The InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization according to claim 2, characterized in that: The formula for calculating the vegetation index is as follows: In the formula, NIR represents the near-infrared band in Sentinel-2 data, and Red represents the red band in Sentinel-2 data; The formula for calculating the water quality index is as follows: In the formula, Green represents the green light band in Sentinel-2 data, and SWIR represents the shortwave infrared band in Sentinel-2 data; The formula for calculating the soil index is as follows: In the formula, Blue represents the blue light band in Sentinel-2 data.

5. The InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization according to claim 1, characterized in that: The loss function of the dual-branch multimodal feature fusion network adopts the cross-entropy function. This cross-entropy function establishes a mapping from the surface deformation phase data in the multi-source remote sensing full-band data to the semantic segmentation ground truth value, based on the InSAR surface deformation anomaly information and the semantic segmentation ground truth value. The specific formula is as follows: In the formula, The cross-entropy loss value is a scalar. The smaller the value, the closer the InSAR surface deformation anomaly information Pred_0 is to the semantic segmentation ground truth value Seg_GT. This represents the total number of pixels in a feature map; Indicates the total number of categories; Indicates the pixel index of the current summation; Indicates the category index of the current summation; The dual-branch multimodal feature fusion network is optimized based on this loss function.

6. The InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization according to claim 1, characterized in that: The dual-branch fusion network includes a dynamic fusion attention module; The fusion attention module uses a lightweight neural network in the channel attention part. The lightweight neural network dynamically predicts the compression ratio of the channel dimension features in the anomaly boundary features and InSAR surface deformation anomaly features, compressing the number of channels to a preset range. In the spatial attention section, the fusion attention module fuses the average pooling, max pooling, and mean structure of the original input features with the anomaly boundary features and InSAR surface deformation anomaly features to construct a comprehensive feature map, which is then convolved to generate a spatial attention map. The fusion attention module fuses the compressed channel attention and spatial attention maps to generate InSAR surface deformation anomaly information.

7. The InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization according to claim 1, characterized in that: The boundary features of the optimized segmentation result are characterized using linear features to represent the convolution, resulting in a boundary feature map; including: The horizontal and vertical boundary features of the optimized segmentation result are extracted by using two linear features to characterize the convolution kernel. The horizontal and vertical boundary features are fused to obtain a complete boundary feature map; Specifically, a cross-entropy loss function is established between the optimized segmentation result and the semantic segmentation ground value to constrain the segmentation accuracy; Establish a boundary loss function to constrain the consistency between the boundary feature map of the optimized segmentation result and the boundary feature map of the semantic segmentation ground truth.

8. The present invention also provides an InSAR deformation anomaly extraction system based on multimodal fusion and vector optimization, realizing the InSAR deformation anomaly extraction method based on multimodal fusion and vector optimization as described in any one of claims 1 to 7, characterized in that, include: Multi-source data acquisition module, model truth value acquisition module, surface deformation information extraction module, and surface deformation vector conversion module; The multi-source data acquisition module includes: Acquire regional multi-source remote sensing data, calculate surface deformation phase data, topographic relief, slope, vegetation index, water index and soil index, and resample them to obtain multi-source remote sensing full-band data; The model truth value acquisition module includes: Based on expert knowledge, the InSAR surface deformation anomaly range in this area was manually interpreted to obtain vector data and binarized mask data of the InSAR surface deformation anomaly range. The surface deformation information extraction module includes: A dual-branch multimodal feature fusion network is constructed, and the binarized mask data is used as the semantic segmentation ground truth. The dual-branch multimodal feature fusion network initially extracts InSAR surface deformation anomaly information based on the semantic segmentation ground truth and the multi-source remote sensing full-band data. The surface deformation vector conversion module includes: A dynamic vector optimization network is constructed to obtain shallow spatial features based on the surface deformation phase data in the multi-source remote sensing full-band data. Based on the shallow spatial features, the InSAR surface deformation anomaly information is optimized at the boundary and dynamically converted into vector results. Based on the vector data, vector boundary anchor points are generated. These vector boundary anchor points are used as the truth values ​​for dynamic vectorization operations, and their differences from the vector results are measured to optimize the dynamic vector optimization network.

Images