Unsupervised change detection method for polarimetric SAR images based on regional relative peak and VIT

Abstract

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

Technical Field

[0001] The invention belongs to the field of image processing and pattern recognition, and in particular relates to an unsupervised polarimetric SAR image change detection method based on regional relative peak value and VIT. Background Art

[0002] Synthetic Aperture Radar (SAR) has the characteristics of all-day and all-weather, and can provide imaging at any time. As an extended system of SAR, polarimetric synthetic aperture radar (PolSAR) obtains four scattered echo signals by alternating transmission and simultaneous reception and vertical transmission and reception to record the full polarization information of the target scattered echo. Therefore, compared with SAR, polarimetric SAR can obtain rich target information, avoid the uncertainty of target information, and enhance clutter suppression and anti-interference capabilities. Polarimetric SAR image change detection refers to the technology of locating the change area caused by human or natural factors in multi-phase polarimetric SAR data. It plays a huge role in monitoring deforestation, natural disaster assessment, land use, urban development, farmland detection and other fields. According to whether the manually marked change map is used in the detection process, the change detection process can be divided into supervised and unsupervised categories. Although the supervised model can achieve satisfactory results, it is a time-consuming and laborious task to manually mark the changes in the surface during the change detection process due to the influence of image noise and complex ground object types. Unsupervised methods are widely used because they can automatically detect changes on the ground without relying on manual marking of change areas. In recent years, the mainstream framework of unsupervised polarimetric SAR image change detection includes three steps: 1) image preprocessing; 2) generating difference maps based on similarity metrics; 3) analyzing difference maps to extract change areas. Among them, difference maps with difference feature information have a great influence on change detection performance. In the early stage of research, scholars proposed difference map construction algorithms such as polarization state construction method, Wishart distance method, Hotelling-Lawley Trace (HLT) detection method, and maximum likelihood ratio method.

[0003] However, the above difference map construction method ignores the adverse effects of inherent coherent speckle noise in polarimetric SAR data, resulting in low overall detection accuracy and regional internal consistency in subsequent steps. In order to better retain the edge information of the change area and suppress noise, Mahdavi, S et al. published an unsupervised polarimetric SAR change detection method (PDI) based on the texture characteristics of the local area of ​​the image in their paper "APolSAR Change Detection Index Based on Neighborhood Information for Flood Mapping" (Remote Sensing, 2019, 11 (16): 1854-). This method uses a weight factor based on the total power ratio of polarimetric scattering and the neighborhood coefficient of variation to generate a difference map. The weight factor can measure the importance of the central pixel and other pixels in its corresponding neighborhood, so the weights corresponding to the original value item and the filter item in the difference map solution operator are adaptively adjusted. This method can suppress noise while retaining the edge detail information of the image, and to a certain extent improves the overall accuracy and internal consistency of change detection. However, its defects are that, firstly, this method only uses partial polarization features (i.e., total polarization scattering power) instead of the complete polarization scattering matrix to measure the difference between two-phase polarization SAR images, which will inevitably lead to the loss of change information. In addition, this method uses the Otsu method to perform threshold processing on the difference map to extract the change area. This global threshold method has a simple idea, but the detection result is sensitive to the selection of the threshold and has poor robustness, which makes it impossible to accurately obtain the change detection result from the difference map.

[0004] Although the above unsupervised polarimetric SAR image change detection methods based on traditional algorithms can achieve change detection to varying degrees, they still have the following defects: 1) First, due to the coherent imaging characteristics of polarimetric SAR data, its inherent coherent speckle noise limits the operability of using polarimetric SAR data for change detection, resulting in the loss of change information and edge information; 2) Manual image features with weak semantics are difficult to describe, resulting in the inability to accurately distinguish between changed and unchanged pixels; 3) The robustness of the change detection results obtained from the difference map through algorithms such as global thresholding and simple clustering is low, and parameters need to be adjusted to adapt to different data.

[0005] In recent years, a series of deep learning algorithms represented by convolutional neural networks (CNNs) have flourished. Since CNN networks have the ability to map between input and output pairs, they can automatically extract high-level semantic features. At present, CNN-based methods have been proposed and used for unsupervised change detection in polarimetric SAR images. R.Habibollahi et al. published a transfer learning-based adaptive end-to-end three-channel deep neural network (TCD-Net) for unsupervised polarimetric SAR image change detection in their paper "TCD-Net: A Novel Deep Learning Framework for Fully Polarimetric Change Detection Using Transfer Learning" (Remote Sensing, 2022, 14(3): 438). First, the method combines the results of the pre-trained model with the results of the FCM algorithm in parallel to generate high-reliability samples for TCD-Net training. Afterwards, a three-channel TCD-Net network based on adaptive multi-scale shallow blocks and adaptive multi-scale residual blocks is constructed. The network is sensitive to objects of different sizes and can accurately classify change classes and non-change classes, thereby obtaining change detection results.

[0006] Although the above-mentioned deep learning-based method can achieve higher change detection accuracy than traditional algorithms by automatically extracting deep semantic features of images, it is limited by the limited effective receptive field. CNN can only capture the short-range context in polarimetric SAR images, which is insufficient to fully explore the data correlation between different slices, and thus makes it impossible to effectively extract change perception features from multi-phase polarimetric SAR images with complex backscattering characteristics. Summary of the invention

[0007] In order to solve the above problems existing in the prior art, the present invention provides an unsupervised polarimetric SAR image change detection method based on regional relative peak and VIT. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0008] The present invention provides an unsupervised polarimetric SAR image change detection method based on regional relative peak value and VIT, comprising:

[0009] Step 1: acquiring dual-phase original polarization SAR images of the same area, and registering the dual-phase original polarization SAR images to obtain a registered polarization SAR image pair;

[0010] Step 2: extract edge information and regional relative peak value of the polarimetric SAR image pair;

[0011] Step 3: correcting the relative peak value of each polarization SAR image using the edge information of each polarization SAR image to obtain the edge-corrected relative peak value of the polarization SAR image pair;

[0012] Step 4: obtaining a multi-scale super-pixel reconstruction difference map of the polarimetric SAR image pair by using the edge-corrected relative peak reconstruction;

[0013] Step 5: performing double threshold segmentation and labeling on the multi-scale super-pixel reconstruction difference map of each polarimetric SAR image to obtain a pseudo-labeling map;

[0014] Wherein, the pseudo-labeled image includes a changed area, an unchanged area and an unknown area;

[0015] Step 6: Construct labeled training samples based on the changed areas and unchanged areas in the pseudo-labeled image, and construct unlabeled prediction samples based on the unknown areas in the pseudo-labeled image;

[0016] Step 7: Using the training sample as the input of the preset VIT network, using the changed area and the unchanged area in the pseudo-labeled image as pixel category labels, iteratively training the preset VIT network until the training cutoff condition is reached, and obtaining a trained VIT network;

[0017] Step 8: Input the predicted sample into the trained VIT network to obtain a detection result of predicting whether the dual-phase original polarization SAR image has changed.

[0018] Optionally, extracting edge information of the polarization SAR image pair in step 2 includes:

[0019] Step 2-11: Calculate the polarization covariance matrix corresponding to all pixel points of each polarization SAR image in the polarization SAR image pair;

[0020] Step 2-12: extracting the amplitudes of the diagonal elements and the upper triangular elements in each polarization covariance matrix to obtain a channel image of each polarization SAR image in the polarization SAR image pair;

[0021] Step 2-13: using an edge detection method to calculate the edge value of each of the channel images to obtain multi-channel edge information of each of the channel images;

[0022] Step 2-14: Calculate the average value of the multi-channel edge information of each of the channel images to obtain the channel mean edge information of each of the polarimetric SAR images;

[0023] Step 2-15: averaging the channel mean edge information of each polarimetric SAR image to obtain edge information of the polarimetric SAR image pair.

[0024] Optionally, extracting the relative peak value of the polarization SAR image pair in step 2 includes:

[0025] Step 2-21: according to the principle that the polarimetric SAR image obeys the K-wishart distribution, calculating the regional relative peak value of each polarimetric SAR image in the polarimetric SAR image pair, and obtaining a relative peak value set of each polarimetric SAR image;

[0026] Step 2-22: averaging each relative peak value set to obtain the regional relative peak value of the polarization SAR image pair.

[0027] Optionally, step 3 includes:

[0028] Step 3-1: performing binary segmentation on edge information of the polarimetric SAR image pair using a threshold binarization method to obtain a threshold value corresponding to the edge information of the polarimetric SAR image pair;

[0029] Step 3-2: using edge information of each pixel in the polarimetric SAR image pair, correct the relative peak value of the polarimetric SAR image pair to obtain a corrected edge-corrected relative peak value of the polarimetric SAR image pair.

[0030] Optionally, step 4 includes:

[0031] Step 4-1: using the polarization covariance matrix corresponding to each of the polarization SAR images and the edge-corrected relative peak value, calculate an edge-corrected relative peak difference map;

[0032] Step 4-2: performing superpixel segmentation on the edge-corrected relative peak difference map at multiple scales using a superpixel segmentation method to obtain a superpixel set;

[0033] Step 4-3: Count the average values ​​and medians of all pixels in the superpixel set of the edge-corrected relative peak difference map to obtain a mean value set and a median value set of the edge-corrected relative peak difference map corresponding to the superpixel set;

[0034] Step 4-4: reconstruct a multi-scale superpixel reconstruction difference map of the polarimetric SAR image pair using the edge-corrected relative peak of the polarimetric SAR image pair, the edge-corrected relative peak difference map, the mean value set, and the median value set.

[0035] Optionally, step 6 includes:

[0036] Step 6-1: stacking the values ​​of the marked areas in the pseudo-label image corresponding to the polarization covariance matrix and the difference values ​​of the marked areas in the pseudo-label image corresponding to the multi-scale superpixel reconstructed difference image pixel by pixel, and using the stacking result as a training sample;

[0037] Step 6-2: Stack the values ​​of the unknown areas in the pseudo-labeled image in the polarization covariance matrix and the difference values ​​of the unknown areas in the pseudo-labeled image in the multi-scale superpixel reconstructed difference map in a row-by-pixel neighborhood, and use the stacking result as a prediction sample.

[0038] 1. The present invention constructs a difference map between any two phases through edge correction region relative peak and multi-scale superpixel segmentation of polarimetric SAR images of any two phases, avoiding the defects of difference map obtained by existing PDI technology that only uses partial polarization features instead of the complete polarization scattering matrix to measure the difference between polarimetric SAR images of two phases, resulting in loss of change information, and uses the global threshold Otsu method to perform threshold processing on the difference map to extract the change area, resulting in the inability to accurately obtain change detection results from the difference map. The boundary positioning capability and the ability to retain change information are improved, and the overall accuracy and regional internal consistency of polarimetric SAR change detection are effectively improved.

[0039] 2. The present invention uses the VIT network to automatically extract deep semantic features of images from multi-phase polarimetric data and difference maps for identification, avoiding the defect that the CNN used in the existing TCD-Net technology can only capture short-range context in polarimetric SAR images, which is insufficient to fully explore the data correlation between different slices, and thus cannot effectively extract change perception features from multi-phase polarimetric SAR images with complex backscattering characteristics, effectively improving the accuracy and regional internal consistency of polarimetric SAR change detection. Experimental results show that the present invention can achieve higher overall accuracy and regional internal consistency in polarimetric SAR change detection.

[0040] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is an overall implementation flow chart of the unsupervised polarimetric SAR image change detection method based on regional relative peak value and VIT of the present invention;

[0042] Figure 2 A schematic diagram of the VIT network structure based on the difference-guided global change feature provided by the present invention;

[0043] Figure 3 It is a schematic diagram of the structure of the extraction model of the VIT network of the present invention;

[0044] Figure 4 A comparison diagram of the change detection results of the polarization SAR image applied to the field 1 data of two phases between the present invention and the prior art;

[0045] Figure 5 The figure is a comparison diagram of the change detection results of the polarization SAR image applied to the field 2 data of two phases of the present invention and the prior art. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0047] Before introducing the present invention, the origin of the concept of the present invention is firstly introduced.

[0048] The Transformer model based on the attention mechanism proposed by Google has been used in fields such as natural language processing and image processing, and has achieved results that surpass traditional classic deep learning networks such as RNN and CNN, especially the Vision Transformer (VIT) suitable for the field of computer vision. This model overcomes the shortcomings of the limited receptive field of traditional CNN, can effectively model and explore the data correlation between different slices, and can achieve better results than CNN.

[0049] refer to Figure 1 As shown, the present invention provides an unsupervised polarimetric SAR image change detection method based on regional relative peak and VIT, which includes:

[0050] Step 1: acquiring dual-phase original polarization SAR images of the same area, and registering the dual-phase original polarization SAR images to obtain a registered polarization SAR image pair;

[0051] It is worth noting that: the polarimetric SAR image T1 of the first phase of the same area is used as the reference image, and the polarimetric SAR image T2 of the second phase is registered with the reference image to obtain the registered polarimetric SAR image sets T1' and T2', wherein the size of the polarimetric SAR images of the two phases is H×L, H≥200, L≥200; taking the two-phase airport 1 data as an example, in this embodiment, H=360, L=324;

[0052] Step 2: extracting edge information and relative peak values ​​of the polarimetric SAR image pair;

[0053] As an optional implementation manner of the present invention, extracting edge information of the polarization SAR image pair in step 2 includes:

[0054] Step 2-11: Calculate the polarization covariance matrix corresponding to all pixel points of each polarization SAR image in the polarization SAR image pair;

[0055] Extract the diagonal elements and the amplitude of the upper triangular elements of the polarization covariance matrices C1 and C2 corresponding to all pixels in T1' and T2' after registration to obtain the 6-channel images corresponding to T1' and T2'

[0056] In the above steps, the polarization covariance matrix C can be expressed as:

[0057]

[0058] in,(·) * represents the complex conjugate, represents the ensemble mean, S hh Indicates the echo data of horizontal transmission and horizontal reception, S hv Indicates the echo data transmitted horizontally and received vertically, S vv Indicates the echo data of vertical transmission and vertical reception;

[0059] Step 2-12: extracting the amplitudes of the diagonal elements and the upper triangular elements in each polarization covariance matrix to obtain a channel image of each polarization SAR image in the polarization SAR image pair;

[0060] Step 2-13: using an edge detection method to calculate the edge value of each of the channel images to obtain multi-channel edge information of each of the channel images;

[0061] Step 2-14: Calculate the average value of the multi-channel edge information of each of the channel images to obtain the channel mean edge information of each of the polarimetric SAR images;

[0062] Step 2-15: averaging the channel mean edge information of each polarimetric SAR image to obtain edge information of the polarimetric SAR image pair.

[0063] It is worth noting that the edge value of each channel in I1 and I2 is calculated by the edge detection method to obtain the 6-channel edge information corresponding to I1 and I2 Calculate the average value of each channel edge information in E1 and E2 to obtain the channel mean edge information corresponding to E1 and E2 calculate and The mean of T1' and T2' is obtained. The present invention can select the ROEWA operator suitable for polarization SAR images as the edge detection method.

[0064] As an optional implementation manner of the present invention, extracting the relative peak value of the polarization SAR image pair in step 2 includes:

[0065] Step 2-21: according to the principle that the polarimetric SAR image obeys the K-wishart distribution, the relative peak value of each polarimetric SAR image in the polarimetric SAR image pair is calculated to obtain a relative peak value set of each polarimetric SAR image;

[0066] Step 2-22: averaging each of the relative peak value sets to obtain the relative peak values ​​of the polarization SAR image pair.

[0067] In this step, according to the K-wishart distribution of the polarimetric SAR image, the relative peak values ​​R1(x,y) and R2(x,y) of the areas defined by all pixels at positions (x,y) in T1' and T2' are calculated to obtain the relative peak value sets R1 and R2 corresponding to T1' and T2'. The mean of R1 and R2 is calculated to obtain the relative peak values ​​corresponding to R1 and R2. Where T m The relative peak value of the area defined by all pixels at position (x,y) in ' The calculation formula is:

[0068]

[0069] in, Represents the echo data of horizontal transmission and horizontal reception at the position (x, y) in the mth phase image, Represents the echo data emitted horizontally and received vertically at the position (x, y) in the mth phase image. represents the echo data of vertical transmission and vertical reception at the (x, y) position in the mth phase image, · represents the modulus of this number, and E{·} represents the mean of this number; in the present invention, the window size of the defined area can be 3.

[0070] Step 3: correcting the relative peak value of each polarization SAR image using the edge information of each polarization SAR image to obtain the edge-corrected relative peak value of the polarization SAR image pair;

[0071] As an optional implementation manner of the present invention, step 3 includes:

[0072] Step 3-1: performing binary segmentation on edge information of the polarimetric SAR image pair using a threshold binarization method to obtain a threshold value corresponding to the edge information of the polarimetric SAR image pair;

[0073] Step 3-2: using edge information of each pixel in the polarimetric SAR image pair, correct the relative peak value of the polarimetric SAR image pair to obtain a corrected edge-corrected relative peak value of the polarimetric SAR image pair.

[0074] The present invention uses a threshold binarization method to Perform binary segmentation and obtain The corresponding threshold The threshold binarization method of the present invention can adopt the OSTU binarization method; The edge information of all pixels at position (x,y) in right The relative peak value of all pixels at position (x,y) in Make corrections to obtain the corrected edge correction relative peak value The edge correction relative peak value of the pixel at position (x, y) The correction formula is:

[0075]

[0076]

[0077] Among them, α is an empirical constant; in the actual process, α=100 can be taken.

[0078] Step 4: obtaining a multi-scale super-pixel reconstruction difference map of the polarimetric SAR image pair by using the edge-corrected relative peak reconstruction;

[0079] As an optional implementation manner of the present invention, step 4 includes:

[0080] Step 4-1: using the polarization covariance matrix corresponding to each of the polarization SAR images and the edge-corrected relative peak value, calculate an edge-corrected relative peak difference map;

[0081] Step 4-2: performing superpixel segmentation on the edge-corrected relative peak difference map at multiple scales using a superpixel segmentation method to obtain a superpixel set;

[0082] Step 4-3: Count the average values ​​and medians of all pixels in the superpixel set of the edge-corrected relative peak difference map to obtain a mean value set and a median value set of the edge-corrected relative peak difference map corresponding to the superpixel set;

[0083] Step 4-4: reconstruct a multi-scale superpixel reconstruction difference map of the polarimetric SAR image pair using the edge-corrected relative peak of the polarimetric SAR image pair, the edge-corrected relative peak difference map, the mean value set, and the median value set.

[0084] The specific process of obtaining the multi-scale super-pixel reconstruction difference map in the present invention is as follows:

[0085] (1) Relative peak value of two-phase polarimetric SAR covariance matrix C1, C2 and edge correction The edge-corrected relative peak difference map DI is obtained, where the difference value calculation formula of the pixel point at position (x, y) in DI is:

[0086]

[0087] Among them, d W {·} represents the modified Wishart distance that satisfies symmetry, Indicates the calculation of the neighborhood Ω of the pixel at (x, y) (x,y) In this embodiment, the window size of the pixel neighborhood is 3;

[0088] In the above steps, the modified Wishart distance d between the polarization matrices C1(x,y) and C2(x,y) satisfies the symmetry W The calculation formula of {·} is:

[0089]

[0090] Wherein, p = 3 in the full polarization data;

[0091] In the above steps, the polarization covariance matrix C of any m-phase m The neighborhood Ω at position (x,y) (x,y) The average value within

[0092]

[0093] Where N is the neighborhood Ω (x,y) The number of pixels in the

[0094] (2) Use the superpixel segmentation method to perform superpixel segmentation on DI at L scales to obtain a superpixel set In this embodiment, the superpixel segmentation method adopts the SLIC method;

[0095] (3) The statistical DI corresponds to each superpixel S l The average value of all pixels in the range l , we get DI corresponding to The mean set of

[0096] (4) Statistical DI corresponds to each superpixel S l The median value is the average value of all pixels in the range. l , we get DI corresponding to The mean set of

[0097] (5) Pass DI, as well as The multi-scale super-pixel reconstruction difference map DI' is obtained, where the difference value calculation formula of the pixel point at position (x, y) in DI' is:

[0098]

[0099]

[0100] Step 5: performing double threshold segmentation and labeling on the difference map of each polarimetric SAR image to obtain a pseudo-labeled map;

[0101] Wherein, the pseudo-labeled image includes a changed area, an unchanged area and an unknown area;

[0102] The present invention can use a dual threshold method to segment DI' to obtain a pseudo-labeled image PI including a changed area, an unchanged area and an unknown area; in this embodiment, the dual threshold method adopts the OSTU dual threshold method.

[0103] Step 6: Construct labeled training samples based on the changed areas and unchanged areas in the pseudo-labeled image, and construct unlabeled prediction samples based on the unknown areas in the pseudo-labeled image;

[0104] As an optional implementation manner of the present invention, step 6 includes:

[0105] Step 6-1: stacking the values ​​of the labeled areas in the pseudo-labeled image corresponding to the polarization covariance matrix and the difference values ​​of the labeled areas in the pseudo-labeled image corresponding to the multi-scale superpixel reconstructed difference map pixel by pixel, and using the stacking result as a training sample;

[0106] Step 6-2: stack the values ​​of the unknown regions in the pseudo-labeled image in the polarization covariance matrix and the difference values ​​of the unknown regions in the pseudo-labeled image in the multi-scale superpixel reconstructed difference map pixel by pixel, and use the stacking result as the prediction sample.

[0107] It is worth noting that the construction process of the training samples and the prediction samples is the same, but the parameters used are different. The 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 as the input patch of the VIT network preset in the present invention, where the input patch is expressed as:

[0108] patch={vec1(x,y)||vec2(x,y)||DI'(x,y),1≤x≤H,1≤y≤L}

[0109]

[0110]

[0111] in, represents the real part, Indicates taking the imaginary part, vec1(x,y), vec2(x,y) and DI'(x,y) respectively represent the pixel-by-pixel neighborhood of the polarization data at the (x,y) position in the corresponding vectors vec1, vec2 and DI' of the polarization SAR images of the two phases, and obtains training samples and input samples. Both types of samples are input pixel blocks, and their number is H×L; the input size can be set to 13×13 in this step.

[0112] Step 7: Using the training sample as the input of the preset VIT network, using the changed area and the unchanged area in the pseudo-labeled image as pixel category labels, iteratively training the preset VIT network until the training cutoff condition is reached, and obtaining a trained VIT network;

[0113] It is worth mentioning that: Reference Figure 2 as well as Figure 3 , input patch for VIT network training, taking the training of a tensor tensor as an example, the process is: first, flatten each patch in tensor and reduce the dimension through the Linear layer; secondly, use the position encoding (Pos_embedding) to encode the position of tensor to obtain the training tensor; then create a new classification Class_token and connect it in parallel with tensor, and add this result to the position encoding, followed by a Dropout layer; use the Transformer structure to extract features, and pass the Pool layer, Layer Norm layer and Linear layer in sequence to obtain the classification result until the network converges and obtains a classifier for change-aware feature extraction; in this embodiment, the training network uses the Adam optimizer with a learning rate of 0.001, the number of training iterations is 50 epochs, and the loss function is the cross entropy loss.

[0114] The training cutoff condition is that the VIT network converges or reaches the number of iterations.

[0115] Step 8: Input the predicted sample into the trained VIT network to obtain a detection result of predicting whether the dual-phase original polarization SAR image has changed.

[0116] The corresponding predicted samples of the unknown area in the pseudo-labeled image PI are input into the trained network and predicted to obtain the final change detection result.

[0117] The effect of the present invention can be further confirmed by the following experiments:

[0118] 1. Simulation conditions and contents:

[0119] 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.

[0120] The overall accuracy and internal consistency of change detection of the present invention, the existing unsupervised polarimetric SAR change detection method (PDI) based on texture characteristics of local areas of images, and the adaptive end-to-end three-channel deep neural network (TCD-Net) based on transfer learning are verified using data from two different time phases. The results are shown in Table 1 in the simulation result analysis, where the internal consistency is evaluated by the Kappa coefficient.

[0121] Simulation 1 compares the change detection results of polarization SAR images of the present invention and the prior art applied to two-phase field 1 data. The results are as follows: Figure 4 shown.

[0122] Simulation 2 compares the change detection results of polarization SAR images of field 2 data in two phases with those of the prior art. Figure 5 shown.

[0123] 2. Analysis of simulation results:

[0124] Reference Figure 4 , where (a) is the detection result of the existing method PDI for Airport 1 data, (b) is the detection result of the existing method TCD-Net for Airport 1 data, (c) is the change detection result of the method of the present invention for Airport 1 data, and (d) is the reference diagram of the change detection result of Airport 1 data. The white area in the diagram represents the changed area, and the black area represents the unchanged area. By comparing (a), (b), (c) and (d) respectively, it can be seen that (c) is stronger in retaining local edge detail information and suppressing local tiny noise than (a) and (b), and is closer to the result given by the reference diagram;

[0125] Reference Figure 5, where (a) is the detection result of the existing method PDI for Airport 2 data, (b) is the detection result of the existing method TCD-Net for Airport 2 data, (c) is the change detection result of the method of the present invention for Airport 2 data, and (d) is the reference diagram of the change detection result of Airport 2 data. The white area in the diagram represents the changed area, and the black area represents the unchanged area. By comparing (a), (b), (c) and (d) respectively, it can be seen that (c) is stronger than (a) and (b) in retaining local edge detail information and suppressing local tiny noise, and is closer to the result given by the reference diagram.

[0126] Table 1

[0127]

[0128]

[0129] Referring to Table 1, the present invention is superior to the prior art in detection accuracy. This is because the present invention reconstructs the difference map between any two phases through the relative peak value of the edge-corrected region and multi-scale superpixel segmentation of the polarimetric SAR images of any two phases. The difference map is obtained by using the relative peak value of the edge-corrected region and multi-scale superpixel reconstruction, which avoids the defects that the difference map obtained by the existing PDI technology only uses partial polarization features instead of the complete polarization scattering matrix to measure the difference between the two-phase polarimetric SAR images, resulting in loss of change information, and uses the global threshold Otsu method to perform threshold processing on the difference map to extract the change area, resulting in the inability to accurately obtain the change detection result from the difference map. In addition, the present invention automatically extracts the deep semantic features of the image from multi-phase polarimetric data and the difference map by using Vision Transformer (VIT), and constructs a VIT network based on difference-guided global change features, avoiding the defect that the CNN used in the prior art can only capture the short-range context in the polarimetric SAR image, which is insufficient to fully explore the data correlation between different slices, and thus cannot effectively extract the change perception features from the multi-phase polarimetric SAR image with complex backscattering characteristics. Therefore, the overall accuracy of change detection and the internal consistency of the region of the present invention are significantly higher than those of the prior art.

[0130] The present invention realizes unsupervised change detection of polarimetric SAR images, which is used to solve the technical problems of low overall detection accuracy and low internal consistency in the region in the prior art, and can be applied to land use, urbanization construction, disaster assessment, forest monitoring, ice field monitoring and other fields.

[0131] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0132] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of components or steps.

[0133] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. An unsupervised polarimetric SAR image change detection method based on regional relative peak and VIT, characterized in that: include: Step 1: acquiring dual-phase original polarization SAR images of the same area, and registering the dual-phase original polarization SAR images to obtain a registered polarization SAR image pair; Step 2: extract edge information and regional relative peak value of the polarimetric SAR image pair; Step 3: correcting the relative peak value of each polarization SAR image using the edge information of each polarization SAR image to obtain the edge-corrected relative peak value of the polarization SAR image pair; Step 4: obtaining a multi-scale super-pixel reconstruction difference map of the polarimetric SAR image pair by using the edge-corrected relative peak reconstruction; Step 5: performing double threshold segmentation and labeling on the multi-scale super-pixel reconstruction difference map of each polarimetric SAR image to obtain a pseudo-labeling map; Wherein, the pseudo-labeled image includes a changed area, an unchanged area and an unknown area; Step 6: Construct labeled training samples based on the changed areas and unchanged areas in the pseudo-labeled image, and construct unlabeled prediction samples based on the unknown areas in the pseudo-labeled image; Step 7: Using the training sample as the input of the preset VIT network, using the changed area and the unchanged area in the pseudo-labeled image as pixel category labels, iteratively training the preset VIT network until the training cutoff condition is reached, and obtaining a trained VIT network; Step 8: Input the predicted sample into the trained VIT network to obtain a detection result of predicting whether the dual-phase original polarization SAR image has changed.

2. The unsupervised polarimetric SAR image change detection method based on regional relative peak and VIT according to claim 1 is characterized in that: Extracting edge information of the polarimetric SAR image pair in step 2 includes: Step 2-11: Calculate the polarization covariance matrix corresponding to all pixel points of each polarization SAR image in the polarization SAR image pair; Step 2-12: extracting the amplitudes of the diagonal elements and the upper triangular 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: using an edge detection method to calculate the edge value of each of the channel images to obtain multi-channel edge information of each of the channel images; Step 2-14: Calculate the average value of the multi-channel edge information of each of the channel images to obtain the channel mean edge information of each of the polarimetric SAR images; Step 2-15: averaging the channel mean edge information of each polarimetric SAR image to obtain edge information of the polarimetric SAR image pair.

3. The unsupervised polarimetric SAR image change detection method based on regional relative peak and VIT according to claim 1 is characterized in that: Extracting the relative peak value of the polarization SAR image pair in step 2 includes: Step 2-21: according to the principle that the polarimetric SAR image obeys the K-wishart distribution, calculating the regional relative peak value of each polarimetric SAR image in the polarimetric SAR image pair, and obtaining a relative peak value set of each polarimetric SAR image; Step 2-22: averaging each relative peak value set to obtain the regional relative peak value of the polarization SAR image pair.

4. The unsupervised polarimetric SAR image change detection method based on regional relative peak and VIT according to claim 1 is characterized in that: The step 3 comprises: Step 3-1: performing binary segmentation on edge information of the polarimetric SAR image pair using a threshold binarization method to obtain a threshold value corresponding to the edge information of the polarimetric SAR image pair; Step 3-2: using edge information of each pixel in the polarimetric SAR image pair, correct the relative peak value of the polarimetric SAR image pair to obtain a corrected edge-corrected relative peak value of the polarimetric SAR image pair.

5. The unsupervised polarimetric SAR image change detection method based on regional relative peak and VIT according to claim 2 is characterized in that: The step 4 comprises: Step 4-1: using the polarization covariance matrix corresponding to each of the polarization SAR images and the edge-corrected relative peak value, calculate an edge-corrected relative peak difference map; Step 4-2: performing superpixel segmentation on the edge-corrected relative peak difference map at multiple scales using a superpixel segmentation method to obtain a superpixel set; Step 4-3: Count the average values ​​and medians of all pixels in the superpixel set of the edge-corrected relative peak difference map to obtain a mean value set and a median value set of the edge-corrected relative peak difference map corresponding to the superpixel set; Step 4-4: reconstruct a multi-scale superpixel reconstruction difference map of the polarimetric SAR image pair using the edge-corrected relative peak of the polarimetric SAR image pair, the edge-corrected relative peak difference map, the mean value set, and the median value set.

6. The unsupervised polarimetric SAR image change detection method based on regional relative peak and VIT according to claim 2 is characterized in that: The step 6 comprises: Step 6-1: stacking the values ​​of the marked areas in the pseudo-label image corresponding to the polarization covariance matrix and the difference values ​​of the marked areas in the pseudo-label image corresponding to the multi-scale superpixel reconstructed difference image pixel by pixel, and using the stacking result as a training sample; Step 6-2: Stack the values ​​of the unknown areas in the pseudo-labeled image in the polarization covariance matrix and the difference values ​​of the unknown areas in the pseudo-labeled image in the multi-scale superpixel reconstructed difference map in a row-by-pixel neighborhood, and use the stacking result as a prediction sample.

Images