CN115546627B - Unsupervised change detection method for polarimetric SAR images based on regional relative peak and VIT - Google Patents

Abstract

Translated fromChinese

本发明公开了一种基于区域相对峰值与VIT的无监督极化SAR图像变化检测方法，通过对多时相的极化SAR图像进行配准并提取边缘信息和区域相对峰值，利用边缘信息对区域相对峰值进行修正，重构极化SAR图像对的差异图；该差异图根据边缘修正区域相对峰值以及多尺度超像素重构得到，避免现有技术导致的变化信息丢失以及无法准确检测变化结果的缺陷，之后再对差异图进行双阈值分割生成伪标记图；以伪标记图构建样本从而训练VIT网络以感知复杂后向散射特性的多时相极化SAR图像的变化特征，之后使用该网络进行变化检测，本发明提高了边界定位能力以及对变化信息的保持能力，有效地提升了极化SAR变化检测的总体精度与区域内部一致性。

The invention discloses an unsupervised polarimetric SAR image change detection method based on regional relative peak value and VIT. The method comprises the following steps: registering polarimetric SAR images of multiple time phases and extracting edge information and regional relative peak value, using edge information to correct regional relative peak value, and reconstructing a difference map of polarimetric SAR image pairs; the difference map is obtained according to the edge-corrected regional relative peak value and multi-scale superpixel reconstruction, thereby avoiding the defects of loss of change information and inability to accurately detect change results caused by the prior art, and then performing double threshold segmentation on the difference map to generate a pseudo-label map; using the pseudo-label map to construct samples, thereby training a VIT network to perceive the change characteristics of multi-time phase polarimetric SAR images with complex backscattering characteristics, and then using the network to perform change detection. The invention improves the boundary positioning capability and the ability to retain change information, and effectively improves the overall accuracy and regional internal consistency of polarimetric SAR change detection.

Description

Unsupervised polarization SAR image change detection method based on regional relative peak value and VIT

Technical Field

The invention belongs to the field of image processing and pattern recognition, and particularly relates to an unsupervised polarized SAR image change detection method based on regional relative peak values and VIT.

Background

Synthetic aperture radar (SYNTHETIC APERTURE RADAR, SAR) has all-day, all-weather properties, and can provide imaging at any time. As an expansion system of SAR, a polarized synthetic aperture radar (Polarimetric Synthetic Aperture Radar, polSAR) acquires four scattered echo signals by alternately transmitting and simultaneously receiving and vertically transmitting and receiving, thereby recording full polarization information of a target scattered echo. Therefore, compared with SAR, the polarized SAR can acquire abundant target information, can avoid the problem of uncertainty of the target information and enhance clutter suppression and anti-interference capability and the like. The polarized SAR image change detection refers to a technology for locating a change area caused by human or natural factors in multi-temporal polarized SAR data, and plays a great role in the fields of monitoring forestation, natural disaster assessment, land utilization, urban development, farmland detection and the like. The change detection process can be classified into two categories, supervised and unsupervised, according to whether a manually marked change pattern is used in the detection process. Although the supervision model can achieve satisfactory results, it is a time-consuming and laborious task to manually annotate the change situation of the earth's surface during the change detection process, subject to image noise and complex ground pattern. The unsupervised method is widely used because it can automatically detect a change occurring on the ground without relying on a manual marking of the change area. In recent years, the mainstream framework of unsupervised polarized SAR image change detection comprises three steps, namely 1) image preprocessing, 2) generating a difference map based on similarity measurement, and 3) analyzing the difference map to extract a change region. Wherein, the difference graph with difference characteristic information has great influence on the change detection performance. In the early stage of research, scholars proposed a difference map construction algorithm such as a polarization state construction method, a Wishart distance method, a Hotelling-LAWLEY TRACE (HLT) detection method, a maximum likelihood ratio method, and the like.

However, the above-described disparity map construction method ignores the adverse effects of inherent speckle noise in polarized SAR data, resulting in lower overall accuracy of detection and intra-region consistency for subsequent steps. In order to better preserve the edge information of the change region and suppress noise, mahdavi, S et al, in their published paper "APolSAR Change Detection Index Based on Neighborhood Information for Flood Mapping"(Remote Sensing,2019,11(16):1854-), disclose an unsupervised polarized SAR change detection method (PDI) based on image local region texture characteristics that utilizes a weight factor based on the total polarized scattered power ratio and the neighborhood variation coefficient to generate a difference map. The weight factors can measure the importance degree of the central pixel and other pixels in the neighborhood corresponding to the central pixel, so that the weights corresponding to the original value items and the filtering items in the difference map solving operator are adaptively adjusted. The method can inhibit noise and simultaneously retain edge detail information of the image, and improves the overall accuracy and internal consistency of change detection to a certain extent. But it has a drawback in that, first, the method only uses part of the polarization characteristics (i.e. total polarized scattering power) instead of the complete polarized scattering matrix to measure the difference between the two-phase polarized SAR images, which inevitably leads to loss of change information. In addition, the method uses the Otsu method to carry out threshold processing on the difference graph to extract the change region, and the global threshold method has simple thought, but the detection result is sensitive to the selection of the threshold value, and has poor robustness, so that the change detection result cannot be accurately obtained from the difference graph.

The unsupervised polarized SAR image change detection method based on the traditional algorithm can realize change detection to different degrees, but still has the following defects that 1) firstly, due to the coherent imaging characteristic of polarized SAR data, inherent coherent speckle noise limits the operability of using the polarized SAR data to carry out change detection, so that change information and edge information are lost, 2) manual image features with weak semantics are difficult to describe, so that changed pixels and unchanged pixels cannot be accurately distinguished, and 3) the robustness of a change detection result obtained from a difference image through algorithms such as a global threshold value, simple clustering and the like is low, and parameters are required to be adjusted to adapt to different data.

In recent years, a series of deep learning algorithms typified by Convolutional Neural Networks (CNNs) have been vigorously developed. Because the CNN network has the mapping capability between input and output pairs, the extraction of high-level semantic features is automatically performed. Currently, CNN-based methods have been proposed and used for unsupervised change detection of polarized SAR images. In its published paper "TCD-Net:A Novel Deep Learning Framework for Fully Polarimetric Change Detection Using Transfer Learning"(Remote Sensing,2022,14(3):438), habibollahi et al discloses an adaptive end-to-end three-channel deep neural network (TCD-Net) based on transfer learning for implementing unsupervised polarized SAR image change detection. First, the method combines the results of the pre-training model with the results of the FCM algorithm in parallel, generating high reliability samples for TCD-Net training. Then, a three-channel TCD-Net network based on the self-adaptive multi-scale shallow block and the self-adaptive multi-scale residual block is constructed, the network is sensitive to objects with different sizes, and the accurate classification of the variable class and the non-variable class can be realized, so that a change detection result is obtained.

Although the deep learning-based method can realize higher change detection precision than the traditional algorithm by automatically extracting deep semantic features of an image, the method is limited by limited effective receptive fields, CNN can only capture short-range contexts in polarized SAR images, and is insufficient for fully exploring data correlation among different slices, so that change perception features cannot be effectively extracted from multi-temporal polarized SAR images with complex backscattering characteristics.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides an unsupervised polarized SAR image change detection method based on regional relative peak values and VIT. The technical problems to be solved by the invention are realized by the following technical scheme:

the invention provides an unsupervised polarization SAR image change detection method based on regional relative peak value and VIT, which comprises the following steps:

Step 1, acquiring dual-phase original polarized SAR images of the same region, and registering the dual-phase original polarized SAR images to obtain registered polarized SAR image pairs;

step 2, extracting edge information and regional relative peak values of the polarized SAR image pair;

step 3, correcting the relative peak value of each polarized SAR image by utilizing the edge information of each polarized SAR image to obtain an edge corrected relative peak value of the polarized SAR image pair;

step 4, reconstructing the relative peak value by utilizing the edge correction to obtain a multi-scale super-pixel reconstruction difference map of the polarized SAR image pair;

step 5, performing double-threshold segmentation marking on the multi-scale super-pixel reconstruction difference map of each polarized SAR image to obtain a pseudo-marking map;

the pseudo mark graph comprises a changed area, an unchanged area and an unknown area;

step 6, constructing a training sample marked with labels according to the changed area and the unchanged area in the pseudo-label graph and constructing a prediction sample without labels according to the unknown area in the pseudo-label graph;

Step 7, taking the training sample as input of a preset VIT network, taking a changed area and an unchanged area in the pseudo-marker graph as pixel class labels, and iteratively training the preset VIT network until reaching a training cut-off condition to obtain a trained VIT network;

And 8, inputting the prediction sample into a trained VIT network to obtain a detection result for predicting whether the original polarized SAR image of the double time phases changes.

Optionally, extracting edge information of the polarized SAR image pair in step 2 includes:

Step 2-11, calculating polarization covariance matrixes corresponding to all pixel points of each polarized SAR image in the polarized SAR image pair;

step 2-12, extracting the amplitude values of diagonal elements and upper triangle elements in each polarization covariance matrix to obtain a channel image of each polarization SAR image in the polarization SAR image pair;

step 2-13, calculating the edge value of each channel image by using an edge detection method to obtain multi-channel edge information of each channel image;

Step 2-14, calculating the average value of the multi-channel edge information of each channel image to obtain the channel average value edge information of each polarized SAR image;

And 2-15, averaging the channel mean value edge information of each polarized SAR image to obtain the edge information of the polarized SAR image pair.

Optionally, extracting the relative peak value of the polarized SAR image pair in step 2 includes:

Step 2-21, calculating the relative peak value of the region of each polarized SAR image in the polarized SAR image pair according to the rule that the polarized SAR image obeys K-wishart distribution, and obtaining a relative peak value set of each polarized SAR image;

And 2-22, averaging each relative peak value set to obtain the regional relative peak value of the polarized SAR image pair.

Optionally, the step 3 includes:

step 3-1, performing binarization segmentation on the edge information of the polarized SAR image pair by using a threshold binarization method to obtain a threshold corresponding to the edge information of the polarized SAR image pair;

and 3-2, correcting the relative peak value of the polarized SAR image pair by utilizing the edge information of each pixel point in the polarized SAR image pair to obtain the edge corrected relative peak value of the polarized SAR image pair after correction.

Optionally, the step4 includes:

step 4-1, calculating an edge correction relative peak value difference map by using a polarization covariance matrix corresponding to each polarized SAR image and the edge correction relative peak value;

step 4-2, performing super-pixel segmentation on the edge correction relative peak difference graph on a plurality of scales by using a super-pixel segmentation method to obtain a super-pixel set;

Step 4-3, counting the average value and the median value of all pixels of the edge correction relative peak difference map in the super pixel set range, and obtaining a mean value set and a median value set of the edge correction relative peak difference map corresponding to the super pixel set range;

And 4-4, reconstructing the multi-scale super-pixel reconstruction difference map of the polarized SAR image pair by utilizing the edge correction relative peak value of the polarized SAR image pair, the edge correction relative peak value difference map, the mean value set and the median value set.

Optionally, the step 6 includes:

Step 6-1, stacking pixel-by-pixel neighborhood of the value corresponding to the marked region in the polarization covariance matrix in the pseudo-marker graph and the difference value corresponding to the marked region in the multi-scale super-pixel reconstruction difference graph in the pseudo-marker graph, and taking the stacking result as a training sample;

and 6-2, stacking the values of the unknown areas in the pseudo-marker map in the polarization covariance matrix and the difference values of the unknown areas in the pseudo-marker map in the multi-scale super-pixel reconstruction difference map in a pixel-by-pixel neighborhood mode, and taking the stacking result as a prediction sample.

1. According to the invention, the edge correction area relative peak value and the multi-scale superpixel segmentation of the polarized SAR image of any two time phases are used for constructing the difference image between any two time phases, so that the defect that the change information is lost due to the fact that the difference image between the polarized SAR images of the two time phases is only measured by using partial polarization characteristics instead of a complete polarization scattering matrix, and the change area cannot be accurately obtained from the difference image due to the fact that the global threshold Otsu method is used for carrying out threshold processing on the difference image is avoided, the boundary positioning capability and the capability of keeping the change information are improved, and the overall accuracy of polarized SAR change detection and the internal consistency of the area are effectively improved.

2. According to the invention, by utilizing the VIT network to automatically extract the deep semantic features of the images from the multi-temporal polarized data and the difference map for recognition, the defect that the CNN adopted in the traditional TCD-Net technology can only capture short-range contexts in the polarized SAR image and is insufficient for fully exploring the data correlation between different slices, so that the change perception features cannot be effectively extracted from the multi-temporal polarized SAR image with complex backscattering characteristics is avoided, and the accuracy rate of polarized SAR change detection and the internal consistency of the region are effectively improved. Experimental results show that the method can obtain higher overall precision and regional internal consistency in the detection of the polarization SAR variation.

The present invention will be described in further detail with reference to the accompanying drawings and examples.

Drawings

FIG. 1 is a flowchart of the overall implementation of an unsupervised polarized SAR image change detection method based on regional relative peak and VIT;

FIG. 2 is a schematic diagram of a VIT network structure based on a difference-guided global variation feature provided by the present invention;

FIG. 3 is a schematic diagram of an extraction model of the VIT network of the present invention;

FIG. 4 is a graph comparing the results of the detection of changes in polarized SAR images of the present invention and prior art applied to two-phase airport 1 data;

fig. 5 is a comparison graph of the results of the change detection of polarized SAR images of the present invention and prior art applied to two-phase airport 2 data.

Detailed Description

The present invention will be described in further detail with reference to specific examples, but embodiments of the present invention are not limited thereto.

Before describing the present invention, the origin of the idea of the present invention will be described first.

The tranformer model based on attention mechanism proposed by google has been used in fields of natural language processing, image processing and the like and has obtained classical deep learning networks such as RNN, CNN and the like, in particular Vision Transformer (VIT) suitable for the field of computer vision.

Referring to fig. 1, the method for detecting the change of the unsupervised polarized SAR image based on the regional relative peak value and the VIT provided by the invention comprises the following steps:

Step 1, acquiring dual-phase original polarized SAR images of the same region, and registering the dual-phase original polarized SAR images to obtain registered polarized SAR image pairs;

It is worth to say that the polarized SAR image T₁ of the first time phase in the same area is taken as a reference image, and the polarized SAR image T₂ of the second time phase is registered with the reference image to obtain a registered polarized SAR image set T₁'、T₂', wherein the sizes of the polarized SAR images of the two time phases are H multiplied by L, H is more than or equal to 200, L is more than or equal to 200;

step 2, extracting edge information and relative peak values of the polarized SAR image pair;

As an optional embodiment of the present invention, extracting edge information of the polarized SAR image pair in the step 2 includes:

Step 2-11, calculating polarization covariance matrixes corresponding to all pixel points of each polarized SAR image in the polarized SAR image pair;

Extracting the amplitudes of the diagonal elements and the upper triangle elements of the polarization covariance matrix C₁、C₂ corresponding to all pixel points in the registered T₁'、T₂ 'to obtain a 6-channel image corresponding to T₁'、T₂'

In the above step, the polarization covariance matrix C may be expressed as:

Wherein, (·)^* represents complex conjugate, ·represents ensemble averaging, S_hh represents horizontally transmitted and horizontally received echo data, S_hv represents horizontally transmitted and vertically received echo data, and S_vv represents vertically transmitted and vertically received echo data;

step 2-12, extracting the amplitude values of diagonal elements and upper triangle elements in each polarization covariance matrix to obtain a channel image of each polarization SAR image in the polarization SAR image pair;

step 2-13, calculating the edge value of each channel image by using an edge detection method to obtain multi-channel edge information of each channel image;

Step 2-14, calculating the average value of the multi-channel edge information of each channel image to obtain the channel average value edge information of each polarized SAR image;

And 2-15, averaging the channel mean value edge information of each polarized SAR image to obtain the edge information of the polarized SAR image pair.

It is worth to say that the edge value of each channel in I₁、I₂ is calculated by an edge detection method to obtain the 6-channel edge information corresponding to I₁、I₂Calculating the average value of the edge information of each channel in E₁、E₂ to obtain the average value edge information of the channel corresponding to E₁、E₂Calculation ofAnd (3) withTo obtain the edge information corresponding to T₁'、T₂The invention can select ROEWA operators suitable for polarized SAR images as an edge detection method.

As an alternative embodiment of the present invention, the extracting the relative peak value of the polarized SAR image pair in step 2 includes:

Step 2-21, calculating a relative peak value of each polarized SAR image in the polarized SAR image pair according to the rule that the polarized SAR images obey K-wishart distribution, and obtaining a relative peak value set of each polarized SAR image;

And 2-22, calculating an average value of each relative peak value set to obtain the relative peak value of the polarized SAR image pair.

According to the polarized SAR image obeying K-wishart distribution, the relative peak value R₁(x,y)、R₂ (x, y) of the pixel point demarcation region with the position (x, y) in T₁'、T₂ 'is calculated, and a relative peak value set R₁、R₂ corresponding to T₁'、T₂' is obtained. Calculating the average value of R₁ and R₂ to obtain the relative peak value corresponding to R₁ and R₂Wherein T_m' defines the relative peak value of the region for all pixel points with (x, y)The calculation formula of (2) is as follows:

Wherein,Represents the horizontally transmitted and horizontally received echo data at the (x, y) position in the mth phase image,Representing the horizontally transmitted and vertically received echo data at the (x, y) position in the mth phase image,Representing the echo data of the vertical transmission and the vertical reception at the (x, y) position in the mth time phase image, & represents the modulus taking this number, E {. Cndot. }, represents the mean taking this number, in the present invention the window size of the delimited area may be 3.

Step 3, correcting the relative peak value of each polarized SAR image by utilizing the edge information of each polarized SAR image to obtain an edge corrected relative peak value of the polarized SAR image pair;

as an alternative embodiment of the present invention, the step 3 includes:

step 3-1, performing binarization segmentation on the edge information of the polarized SAR image pair by using a threshold binarization method to obtain a threshold corresponding to the edge information of the polarized SAR image pair;

and 3-2, correcting the relative peak value of the polarized SAR image pair by utilizing the edge information of each pixel point in the polarized SAR image pair to obtain the edge corrected relative peak value of the polarized SAR image pair after correction.

The invention utilizes the threshold value binarization method pairPerforming binarization segmentation to obtainCorresponding threshold valueThe threshold value binarization method of the invention can adopt an OSTU binarization method byEdge information of all the pixels at (x, y)For a pair ofThe relative peak value of the pixel at (x, y) in all positionsCorrecting to obtain corrected edge correction relative peak valueEdge corrected relative peak value of pixel with position (x, y)The correction formula of (2) is:

where α is an empirical constant, and it is desirable in practice that α=100.

Step 4, reconstructing the relative peak value by utilizing the edge correction to obtain a multi-scale super-pixel reconstruction difference map of the polarized SAR image pair;

As an alternative embodiment of the present invention, the step 4 includes:

step 4-1, calculating an edge correction relative peak value difference map by using a polarization covariance matrix corresponding to each polarized SAR image and the edge correction relative peak value;

step 4-2, performing super-pixel segmentation on the edge correction relative peak difference graph on a plurality of scales by using a super-pixel segmentation method to obtain a super-pixel set;

Step 4-3, counting the average value and the median value of all pixels of the edge correction relative peak difference map in the super pixel set range, and obtaining a mean value set and a median value set of the edge correction relative peak difference map corresponding to the super pixel set range;

And 4-4, reconstructing the multi-scale super-pixel reconstruction difference map of the polarized SAR image pair by utilizing the edge correction relative peak value of the polarized SAR image pair, the edge correction relative peak value difference map, the mean value set and the median value set.

The specific process for obtaining the multi-scale super-pixel reconstruction difference map is as follows:

(1) Relative peak correction by two-phase polarization SAR covariance matrix C₁、C₂ and edge correctionObtaining an edge correction relative peak difference map DI, wherein a difference value calculation formula of pixel points with positions (x, y) in DI is as follows:

Wherein d_W {.cndot }, represents the modified Wishart distance satisfying symmetry,Representing calculating an average value in a pixel neighborhood omega_(x,y) at (x, y), wherein in the embodiment, the window size of the pixel neighborhood is 3;

In the above step, the calculation formula of the corrected Wishart distance d_W {. Cndot } satisfying symmetry between the polarization matrices C₁(x,y)、C₂ (x, y) is:

wherein p=3 in the full polarization data;

in the above steps, the average value in neighborhood Ω_(x,y) at position (x, y) in any m-phase polarization covariance matrix C_m

Wherein N is the number of pixel points in the neighborhood omega_(x,y);

(2) Super-pixel segmentation is carried out on DI on L scales by utilizing a super-pixel segmentation method to obtain a super-pixel setIn this embodiment, the super pixel dividing method adopts an SLIC method;

(3) Counting average mean_l of all pixels within each super pixel S_l to obtain DI corresponding toMean value set of (a)

(4) Statistical DI corresponds to average value media_l of all pixels within each super-pixel S_l to obtain DI corresponds toMean value set of (a)

(5) By passing throughDI、AndObtaining a multi-scale super-pixel reconstruction difference map DI ', wherein a difference value calculation formula of pixel points with positions (x, y) in DI' is as follows:

Step 5, performing double-threshold segmentation marking on the difference map of each polarized SAR image to obtain a pseudo-marking map;

the pseudo mark graph comprises a changed area, an unchanged area and an unknown area;

The method can divide DI' by using a double-threshold method to obtain the pseudo-marker diagram PI containing the changed region, the unchanged region and the unknown region, wherein in the embodiment, the double-threshold method adopts an OSTU double-threshold method.

Step 6, constructing a training sample marked with labels according to the changed area and the unchanged area in the pseudo-label graph and constructing a prediction sample without labels according to the unknown area in the pseudo-label graph;

As an alternative embodiment of the present invention, the step 6 includes:

Step 6-1, stacking the value of the marked region in the pseudo-marker graph corresponding to the polarization covariance matrix and the difference value of the marked region in the pseudo-marker graph corresponding to the multi-scale super-pixel reconstruction difference graph in a pixel-by-pixel neighborhood manner, and taking the stacking result as a training sample;

And 6-2, stacking the values of the unknown areas in the pseudo-marker image corresponding to the polarization covariance matrix and the difference values of the unknown areas in the pseudo-marker image corresponding to the multi-scale super-pixel reconstruction difference image in a pixel-by-pixel neighborhood manner, and taking the stacking result as a prediction sample.

It should be noted that the training samples and the prediction samples are identical in construction process, and different parameters are utilized, namely, polarization data and difference values at the same position (x, y) in the phase polarization SAR covariance matrix C₁、C₂ and DI' are stacked pixel by pixel neighborhood, and the stacked polarization data and difference values are used as input patches of the preset VIT network, wherein the input patches are expressed as:

patch={vec₁(x,y)||vec₂(x,y)||DI'(x,y),1≤x≤H,1≤y≤L}

Wherein,The representation takes the real part of the component,Representing the imaginary part, vec₁(x,y)、vec₂ (x, y) and DI '(x, y) respectively represent the corresponding vectors vec₁、vec₂ of the polarized SAR images of the two time phases and the polarization data of the (x, y) position in DI' pixel by pixel neighborhood, obtaining training samples and input samples, wherein both types of samples are input pixel blocks, the number of which is h×l, and the input size in the step can be set to be 13×13.

Step 7, taking the training sample as input of a preset VIT network, taking a changed area and an unchanged area in the pseudo-marker graph as pixel class labels, and iteratively training the preset VIT network until reaching a training cut-off condition to obtain a trained VIT network;

It is worth to describe that referring to fig. 2 and 3, input patches are used for VIT network training, taking training a tensor tensor as an example, the process is that each patch in tensor is flattened and dimension reduced through a Linear Layer, the tensor position is encoded by position encoding (pos_ embedding) to obtain a training tensor, classification class_token is newly built and connected in parallel with tensor, the result is added with the position encoding, dropout Layer is connected, a Transformer structure is used for extracting features, classification results are obtained through a Pool Layer, a Layer Norm Layer and a Linear Layer in sequence until the network converges, a classifier for extracting change perception features is obtained, in the embodiment, the training network adopts an Adam optimizer with a learning rate of 0.001, the training iteration number is 50 epoch, and the loss function is cross entropy loss.

The training cut-off condition is that the VIT network converges or reaches iteration times.

And 8, inputting the prediction sample into a trained VIT network to obtain a detection result for predicting whether the original polarized SAR image of the double time phases changes.

And inputting a corresponding prediction sample of an unknown region in the pseudo-marker diagram PI into the trained network and predicting to obtain a final change detection result.

The effect of the invention can be further confirmed by the following experiments:

1. simulation conditions and content:

The simulation experiment environment is ：MATLAB R2016a,Intel(R)Core(TM)i7-8700 CPU 3.20GHz,Window 10,PyTorch1.5.1,NVIDIA GeForce RTX 1660Ti GPU.

The invention verifies the overall accuracy and internal consistency of the change detection of the invention and the existing unsupervised polarization SAR change detection method (PDI) based on the image local area texture characteristics and the self-adaptive end-to-end three-channel deep neural network (TCD-Net) based on transfer learning by adopting two-time phase different data, and the result is shown in a table 1 in simulation result analysis, wherein the internal consistency is evaluated by Kappa coefficient.

Simulation 1, comparing and simulating the change detection result of the polarized SAR image of the data of the two-time-phase airport 1, which is applied to the invention and the prior art, wherein the result is shown in figure 4.

Simulation 2, comparing and simulating the change detection result of the polarized SAR image of the data of the two-time-phase airport 2, which is applied to the invention and the prior art, and the result is shown in figure 5.

2. Simulation result analysis:

Referring to fig. 4, wherein (a) is a detection result of the existing method PDI of airport 1 data, (b) is a detection result of the existing method TCD-Net of airport 1 data, (c) is a change detection result of the method of the present invention of airport 1 data, (d) is a reference image of the change detection result of airport 1 data, white areas in the image represent change areas, and black areas represent unchanged areas, and the (a) image, (b) image, (c) image and (d) image are respectively compared, so that the retention capability of the (c) image on local edge detail information and the suppression capability of the (c) on local micro noise are stronger than those of the (a) image and the (b) image, and the results given by the reference image are closer;

Referring to fig. 5, where (a) is a detection result of the existing method PDI of airport 2 data, (b) is a detection result of the existing method TCD-Net of airport 2 data, (c) is a change detection result of the method of the present invention of airport 2 data, and (d) is a reference image of the change detection result of airport 2 data, white areas in the image indicate change areas, and black areas indicate unchanged areas, and comparing (a) with (b) and (c) with (d) respectively, it can be seen that the retention capability of (c) on local edge detail information and the suppression capability of local micro noise are stronger than those of (a) with (b) and are closer to the results given by the reference image.

TABLE 1

Referring to table 1, the present invention is superior to the prior art in detection accuracy. The invention reconstructs a difference image between any two time phases by using the relative peak value of the edge correction area and multi-scale superpixel segmentation of the polarized SAR image of any two time phases, the difference image is obtained by using the relative peak value of the edge correction area and multi-scale superpixel reconstruction, the defects that the change information is lost due to the fact that the difference image obtained by the prior PDI technology only utilizes partial polarization characteristics instead of a complete polarization scattering matrix to measure the difference between two-time-phase polarized SAR images and the change area cannot be accurately obtained due to the fact that the difference image is subjected to threshold processing by using a global threshold Otsu method are avoided, and in addition, the invention constructs a VIT network based on the difference guiding global change characteristics by using Vision Transformer (VIT) to automatically extract the deep semantic characteristics of the image from the multi-time-phase polarization data and the difference image, thereby avoiding that CNN adopted in the prior art can only capture the context in the polarized image, and the situation that the data association between different slices is insufficient, and further, the change detection result cannot be effectively extracted from the multi-time-phase polarized SAR image with complex backward scattering characteristics is obviously consistent with the prior art.

The invention realizes the unsupervised change detection of the polarized SAR image, is used for solving the technical problem of lower detection overall accuracy and region internal consistency in the prior art, and can be applied to the fields of land utilization, urban construction, disaster evaluation, forest monitoring, ice field monitoring and the like.

Furthermore, the terms "first," "second," and the like, are used for descriptive purposes only and are not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include one or more such feature. In the description of the present invention, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.

Although the application is described herein in connection with various embodiments, other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed application, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the "a" or "an" does not exclude a plurality.

The foregoing is a further detailed description of the invention in connection with the preferred embodiments, and it is not intended that the invention be limited to the specific embodiments described. It will be apparent to those skilled in the art that several simple deductions or substitutions may be made without departing from the spirit of the invention, and these should be considered to be within the scope of the invention.

Claims (6)

1. An unsupervised polarized SAR image change detection method based on regional relative peak value and VIT is characterized by comprising the following steps:

Step 1, acquiring dual-phase original polarized SAR images of the same region, and registering the dual-phase original polarized SAR images to obtain registered polarized SAR image pairs;

step 2, extracting edge information and regional relative peak values of the polarized SAR image pair;

step 3, correcting the relative peak value of each polarized SAR image by utilizing the edge information of each polarized SAR image to obtain an edge corrected relative peak value of the polarized SAR image pair;

step 4, reconstructing the relative peak value by utilizing the edge correction to obtain a multi-scale super-pixel reconstruction difference map of the polarized SAR image pair;

step 5, performing double-threshold segmentation marking on the multi-scale super-pixel reconstruction difference map of each polarized SAR image to obtain a pseudo-marking map;

the pseudo mark graph comprises a changed area, an unchanged area and an unknown area;

step 6, constructing a training sample marked with labels according to the changed area and the unchanged area in the pseudo-label graph and constructing a prediction sample without labels according to the unknown area in the pseudo-label graph;

Step 7, taking the training sample as input of a preset VIT network, taking a changed area and an unchanged area in the pseudo-marker graph as pixel class labels, and iteratively training the preset VIT network until reaching a training cut-off condition to obtain a trained VIT network;

And 8, inputting the prediction sample into a trained VIT network to obtain a detection result for predicting whether the original polarized SAR image of the double time phases changes.

2. The method for detecting changes in an unsupervised polarized SAR image based on regional relative peak and VIT according to claim 1, wherein extracting edge information of the pair of polarized SAR images in step 2 comprises:

Step 2-11, calculating polarization covariance matrixes corresponding to all pixel points of each polarized SAR image in the polarized SAR image pair;

step 2-12, extracting the amplitude values of diagonal elements and upper triangle elements in each polarization covariance matrix to obtain a channel image of each polarization SAR image in the polarization SAR image pair;

step 2-13, calculating the edge value of each channel image by using an edge detection method to obtain multi-channel edge information of each channel image;

Step 2-14, calculating the average value of the multi-channel edge information of each channel image to obtain the channel average value edge information of each polarized SAR image;

And 2-15, averaging the channel mean value edge information of each polarized SAR image to obtain the edge information of the polarized SAR image pair.

3. The method for detecting changes in an unsupervised polarized SAR image based on regional relative peak and VIT according to claim 1, wherein said extracting the relative peak of the pair of polarized SAR images in step 2 comprises:

Step 2-21, calculating the relative peak value of the region of each polarized SAR image in the polarized SAR image pair according to the rule that the polarized SAR image obeys K-wishart distribution, and obtaining a relative peak value set of each polarized SAR image;

And 2-22, averaging each relative peak value set to obtain the regional relative peak value of the polarized SAR image pair.

4. The method for detecting the change of the unsupervised polarized SAR image based on the relative peak value and the VIT of the area according to claim 1, wherein said step 3 comprises:

step 3-1, performing binarization segmentation on the edge information of the polarized SAR image pair by using a threshold binarization method to obtain a threshold corresponding to the edge information of the polarized SAR image pair;

and 3-2, correcting the relative peak value of the polarized SAR image pair by utilizing the edge information of each pixel point in the polarized SAR image pair to obtain the edge corrected relative peak value of the polarized SAR image pair after correction.

5. The method for detecting the change of the unsupervised polarized SAR image based on the relative peak value and the VIT of the area according to claim 2, wherein said step 4 comprises:

step 4-1, calculating an edge correction relative peak value difference map by using a polarization covariance matrix corresponding to each polarized SAR image and the edge correction relative peak value;

step 4-2, performing super-pixel segmentation on the edge correction relative peak difference graph on a plurality of scales by using a super-pixel segmentation method to obtain a super-pixel set;

Step 4-3, counting the average value and the median value of all pixels of the edge correction relative peak difference map in the super pixel set range, and obtaining a mean value set and a median value set of the edge correction relative peak difference map corresponding to the super pixel set range;

And 4-4, reconstructing the multi-scale super-pixel reconstruction difference map of the polarized SAR image pair by utilizing the edge correction relative peak value of the polarized SAR image pair, the edge correction relative peak value difference map, the mean value set and the median value set.

6. The method for detecting the change of the unsupervised polarized SAR image based on the relative peak value and the VIT of the area according to claim 2, wherein said step 6 comprises:

Step 6-1, stacking pixel-by-pixel neighborhood of the value corresponding to the marked region in the polarization covariance matrix in the pseudo-marker graph and the difference value corresponding to the marked region in the multi-scale super-pixel reconstruction difference graph in the pseudo-marker graph, and taking the stacking result as a training sample;

and 6-2, stacking the values of the unknown areas in the pseudo-marker map in the polarization covariance matrix and the difference values of the unknown areas in the pseudo-marker map in the multi-scale super-pixel reconstruction difference map in a pixel-by-pixel neighborhood mode, and taking the stacking result as a prediction sample.