Search

CN-122024011-A - InSAR point target selection method and device based on double-branch fusion

CN122024011ACN 122024011 ACN122024011 ACN 122024011ACN-122024011-A

Abstract

The invention discloses an InSAR point target selection method based on double-branch fusion, which comprises the steps of constructing a space branch label and a time branch label based on a time sequence SLC image, constructing a DSTFF-PTNet model, wherein the DSTFF-PTNet model comprises a space branch module and a time branch module, respectively obtaining a point target selection result diagram of the space branch and a point target selection result diagram of the time branch, and finally fusing the two point target selection result diagrams to obtain a final selection result of an interference point target.

Inventors

  • LIAO HAISHENG
  • FAN JINYONG
  • LI XIAOLIANG
  • HUANG RONG
  • LI LINZE
  • YANG LEI
  • Luo Huiheng
  • WEN FAN
  • JIANG LIMING

Assignees

  • 三峡金沙江云川水电开发有限公司

Dates

Publication Date
20260512
Application Date
20260312

Claims (10)

  1. 1. The InSAR point target selection method based on the double-branch fusion is characterized by comprising the following steps of: step 1, acquiring SAR images with high track precision of a plurality of time nodes in a research area, respectively performing registration processing to obtain SLC images of the plurality of time nodes, and arranging all the SLC images in time sequence to form a time sequence SLC image; step 2, based on the time sequence SLC image, constructing a space branch label and a time branch label, aligning and slicing the time sequence SLC image, the space branch label and the time branch label to obtain a plurality of time sequence SLC sub-images and corresponding space branch sub-labels and time branch sub-labels, and constructing a training set, a verification set and a test set; Step 3, constructing DSTFF-PTNet model, wherein the DSTFF-PTNet model comprises a space branching module and a time branching module, a time sequence SLC sub-image is input into the space branching module to obtain a point target selection result diagram of the space branching, and a time sequence SLC sub-image is input into the time branching module to obtain a point target selection result diagram of the time branching; Step 4, setting a loss function, training the DSTFF-PTNet model through a training set, calculating loss by using a point target selection result diagram of the space branch and a corresponding space branch sub-label, calculating loss by using a point target selection result diagram of the time branch and a corresponding time branch sub-label, and storing model parameters after training is finished to obtain a DSTFF-PTNet model with optimal weight; And 5, inputting the time sequence SLC image to be processed into a DSTFF-PTNet model with optimal weight to obtain a point target selection result diagram of the space branch and a point target selection result diagram of the time branch, and finally fusing the two point target selection result diagrams to obtain a final selection result of the interference point target.
  2. 2. The method for selecting the target point of the InSAR based on the double-branch fusion according to claim 1, wherein the SAR images with high orbit precision of the plurality of time nodes of the investigation region in the target region in the step 1 are obtained by: Step 1.1.1, downloading SAR images, digital elevation models and precise orbit determination ephemeris data of a plurality of time nodes of a target area; Step 1.1.2, extracting SAR images at corresponding positions according to the spatial positions of a research area in a target area in each SAR image to obtain SAR images of a plurality of time nodes of the research area; And 1.1.3, replacing original incidental low-precision orbit information in each SAR image of the research area by using the precise orbit determination ephemeris data to obtain high-orbit precision SAR images of a plurality of time nodes of the research area.
  3. 3. The method for selecting the InSAR point target based on the double-branch fusion according to claim 1, wherein the spatial branch labels in the step 2 are constructed by selecting point targets with average coherence larger than a first set value from each SLC image in a spatial dimension to construct high-coherence point target images, and taking each high-coherence point target image as the spatial branch label of the corresponding SLC image.
  4. 4. The InSAR point target selection method based on the double-branch fusion according to claim 1, wherein the time-branch label is constructed by screening point targets with a ratio of a time sequence amplitude mean value to a standard deviation larger than a second set value in a time dimension.
  5. 5. The method for selecting the InSAR point target based on the double-branch fusion according to claim 1, wherein the training set, the verification set and the test set are constructed by the following modes: The method comprises the steps of aligning a space branch label and a time branch label with a time sequence SLC image, slicing the time sequence SLC image, the space branch label and the time branch label in a space dimension by adopting a sliding window with a set size to obtain a plurality of time sequence SLC sub-images and corresponding space branch sub-labels and time branch sub-labels, wherein each time sequence SLC sub-image comprises SLC sub-images of a plurality of time nodes, each time sequence SLC sub-image and corresponding space branch sub-label and time branch sub-label are used as one sample, the time sequence SLC sub-images in each sample are used as input samples, and all samples are proportionally divided into a training set, a verification set and a test set.
  6. 6. The method for selecting the target of the InSAR point based on the double-branch fusion according to claim 1, wherein the space branch module is constructed by the following steps: The time sequence SLC sub-images are input into a space branching module, and firstly, the space branching module expands the time sequence SLC sub-images into SLC sub-images of all time nodes along the time dimension; the space branch module also comprises a space branch processing module and a space branch integration module; The spatial branch processing module comprises a first type convolution module, a first residual error module and a second residual error module, wherein the SLC sub-images are input into the first type convolution module, the output characteristic images of the first type convolution module are input into the first residual error module, the output characteristic images of the first residual error module are input into the second residual error module, the second residual error module outputs the output characteristic images of the spatial branch processing module, the SLC sub-images of all time nodes are sequentially input into the spatial branch processing module to respectively obtain the output characteristic images of the corresponding spatial branch processing module, the spatial branch integration module comprises mean value pooling processing, and the output characteristic images of the spatial branch processing module corresponding to the SLC images of all time nodes are subjected to mean fusion to obtain a point target selection result image of the spatial branch corresponding to the time sequence SLC sub-images; The method comprises the steps that a first type convolution module comprises a 3X 3 convolution layer, a batch normalization layer and a ReLU activation function which are sequentially connected, an SLC sub-image is input into the 3X 3 convolution layer, the batch normalization layer and the ReLU activation function are sequentially processed, an output feature image of the ReLU activation function of a first type convolution module is input into a first residual error module, the residual error module comprises a first type convolution module, a second type convolution module and the ReLU activation function, an input feature image of the residual error module is input into the first type convolution module of the residual error module, an output feature image of the first type convolution module of the residual error module is input into the second type convolution module, a residual error path is introduced through a skip connection mechanism, the input feature image of the residual error module is matched with the channel number of the output feature image of the second type convolution module through the 1X 1 convolution layer on the residual error path, and the input feature image of the second type convolution module is obtained after channel matching is added, and then the input feature image of the second type convolution module is input into the ReLU activation function of the residual error module; the input feature map of the first residual error module is an output feature map of a ReLU activation function of a first convolution module of the first residual error module, and the input feature map of the second residual error module is an output feature map of a ReLU activation function of the first residual error module; The second type of convolution module comprises a 3×3 convolution layer and a batch normalization layer which are connected in sequence.
  7. 7. The method for selecting the InSAR point target based on the double-branch fusion according to claim 1, wherein the time-branch module comprises the following processing steps of: The time branching module comprises a linear mapping module, a leachable position encoding module, a transducer encoder, a mean value pooling process and a1 multiplied by 1 convolution layer; Step 3.2.1, inputting a time sequence SLC sub-image into a time branching module, firstly expanding the time sequence into a pixel-level time sequence in a space dimension, wherein each space pixel position corresponds to a one-dimensional sequence, each one-dimensional sequence comprises pixel values of a plurality of time nodes, and the pixel values of each time node are respectively data of the SLC sub-image of the corresponding time node in the time sequence SLC sub-image at the corresponding pixel position; Step 3.2.2, mapping pixel values of all time nodes in each one-dimensional sequence to a feature space with the dimension of 32 through a linear mapping module to obtain 32-dimensional feature vectors of all the time nodes in each one-dimensional sequence, wherein the 32-dimensional feature vectors of all the time nodes of each one-dimensional sequence form sequence features of the corresponding one-dimensional sequence; Step 3.2.3, coding each spatial pixel position in the time sequence SLC sub-image through a learnable position coding module to obtain a position code of each spatial pixel position in the time sequence SLC sub-image, and then adding the position code and sequence features of a one-dimensional sequence of corresponding spatial pixel positions element by element to obtain a first sequence feature after fusion of the position codes; step 3.2.4, inputting the first sequence features of the one-dimensional sequences corresponding to the spatial pixel positions into a transducer encoder for modeling to obtain modeled second sequence features, and outputting the second sequence features of the one-dimensional sequences corresponding to the spatial pixel positions by the transducer encoder; And 3.2.5, respectively averaging the second sequence features of the one-dimensional sequence corresponding to each spatial pixel position output by the transducer encoder along the time dimension through mean value pooling treatment, remolding the one-dimensional sequence corresponding to each averaged spatial pixel position according to the spatial pixel position to obtain a time sequence SLC feature image, and finally adjusting the channel number of the time sequence SLC feature image through a 1X 1 convolution layer to obtain a point target selection result diagram of the time branch.
  8. 8. The method for selecting the target of the InSAR point based on the double-branch fusion as set forth in claim 1, wherein the DSTFF-PTNet model is trained by the following means: step 4.1, configuring training parameters, namely setting training rounds by adopting a AdamW optimizer and a dynamic learning rate adjustment mechanism, and setting a loss function as a binary cross entropy loss function with Logits; Step 4.2, carrying out random transformation on samples in the training set; And 4.3, training the DSTFF-PTNet model through a training set, in each round of training, firstly inputting an input sample in the training set into the DSTFF-PTNet model, carrying out iterative optimization on model parameters by means of a back propagation mechanism, recording an average loss value in the training process, then testing the DSTFF-PTNet model by utilizing a verification set, calculating the average verification loss value, and if the verification result shows that the model has abnormal fitting or is not converged in training, re-executing the training process by adjusting super parameters such as learning rate, batch size and the like until the configuration of optimal parameters meeting the training requirement is obtained.
  9. 9. The method for selecting the InSAR point target based on the double-branch fusion according to claim 1, wherein the point target selecting result diagram of the space branch and the point target selecting result diagram of the time branch are fused to obtain the final selecting result of the interference point target by the following modes: The joint strategy based on the consistency of the spatial position and the occurrence frequency is adopted, namely, a spatial tolerance range is firstly set, and points from different branches are regarded as points at the same position in the spatial tolerance range; Counting the occurrence frequency of points at each position in the point target selection result diagram of the space branch and the point target selection result diagram of the time branch, and reserving when a certain point occurs in the point target selection result diagram of the space branch and the point target selection result diagram of the time branch or the occurrence frequency reaches a preset threshold; and traversing the points at each position in sequence to obtain the selection result of the interference point target.
  10. 10. A computer device comprising a memory and a processor, the memory having stored therein a computer program, characterized in that the processor, when executing the computer program, performs the steps of the method for selecting an InSAR point target based on a double branch fusion as claimed in any one of claims 1-9.

Description

InSAR point target selection method and device based on double-branch fusion Technical Field The invention belongs to the technical field of remote sensing image processing and intelligent target recognition, and particularly relates to an InSAR point target selection method based on double-branch fusion and computer equipment. Background In recent years, synthetic aperture radar interferometry (SYNTHETIC APERTURE RADAR INTERFEROMETRY, inSAR) has found widespread use in surface deformation monitoring. In particular, the introduction of the multi-time-phase InSAR technology effectively overcomes the limitation of the traditional Differential InSAR (DINSAR) in the aspects of atmospheric disturbance and out-of-phase interference noise by carrying out time sequence analysis on the multi-phase differential interference image. However, not all scatterers are suitable for multi-phase analysis, and some are extremely sensitive to noise and coherence loss. Scatterers can be classified into three types, permanent scatterers (PERSISTENT SCATTERER, PS), distributed scatterers (Distributed Scatterer, DS) and Non-scatterers (Non-scatterer, NS), depending on the degree of influence of incoherence and noise in the phase time series. The PS point target has higher coherence, usually corresponds to strong reflectors such as buildings and rocks or slopes facing away from satellites, the DS point target has moderate coherence and is commonly found in farmland, sandy lands and other areas, and the NS point target cannot be used for deformation estimation due to interference of strong noise and interference loss and is commonly found in water and vegetation areas. Permanent scatterer InSAR (PERSISTENT SCATTERER INSAR, PSINSAR) is a common multi-temporal InSAR technique, and the core of the technique is the effective selection of high-quality PS points. Early algorithms were based on amplitude dispersion index (Amplitude Dispersion Index, ADI) in combination with a priori time models, performed well in urban areas, but had limited effectiveness in areas where signal-to-noise ratio was low or where strong reflectors were lacking. The StaMPS method proposed by Hooper does not depend on a priori model of deformation, but recognizes PS points by analyzing spatial correlation and phase noise estimation of deformation, and effectively overcomes the defects of the ADI method. However, staMPS still has the case of missing PS points in complex terrain. Ferretti et al, which also present difficulties in identifying in low reflectivity regions by analyzing the amplitude statistical differences of neighboring pixels. Currently, staMPS is the only publicly available algorithm, and the rest is mostly proprietary software. Despite the variety of algorithms available, there are still a large number of potential point targets in real-world applications that cannot be effectively identified, and available point targets remain scarce, especially in non-urban areas. In recent years, deep learning (DEEP LEARNING, DL) has achieved remarkable results in tasks such as image classification and time series prediction, and is gradually introduced into the remote sensing field, and particularly has good potential in aspects of feature extraction and classification of SAR images. Because the remote sensing data is often accompanied with the characteristics of high noise, complex geometric structures, irregular distribution of point targets and the like, the deep learning model needs to be optimized through targeted super parameters in practical application so as to improve the recognition precision and the generalization capability of the model. In a multi-temporal InSAR processing chain, the process of selecting the coherence points (e.g., PS and DS) typically takes up a significant amount of computation time. This process involves processing millions of points and requires multiple iterations to ensure convergence of the phase standard deviation. The processing time varies from hours to days, depending on the number of candidate points and the size of the multi-temporal SAR image. In a scenario oriented to near real-time surface deformation monitoring, multi-temporal InSAR processing needs to be run regularly, thus making the demand for computing resources a challenge. Disclosure of Invention The invention aims to solve the problems of the prior art, such as low efficiency, dependence on manual experience, strong subjectivity and the like, by constructing a Dual-branch deep neural network (Dual-branch Spatio-Temporal Feature Fusion Network for InSAR Point Target Selection, DSTFF-PTNet) which fuses time sequence information and spatial structural characteristics, the invention provides an InSAR point target selection method based on Dual-branch fusion, and provides computer equipment. The above object of the present invention is achieved by the following technical means: the InSAR point target selection method based on the double-branch fusion comprises the