A ground feature classification method based on dual-band dual-polarization SAR image
By registering dual-band dual-polarization SAR images and classifying them using the H/αWishart algorithm, combined with polarization despotting and Gaussian filtering, the problem of multi-band image data fusion was solved, and the accuracy and consistency of ground feature classification were improved.
Patent Information
- Application Number
- CN202310686826.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-11
- Publication Date
- 2025-12-09
- Estimated Expiration
- 2043-06-11
AI Technical Summary
Existing polarimetric SAR image land cover classification methods are mostly based on single bands, failing to effectively utilize the complementary information of dual-polarimetric SAR images. Furthermore, multi-band image data fusion is difficult to achieve, and existing methods have failed to effectively fuse classification results in the image domain.
After registration of dual-band dual-polarization SAR images, the H/α Wishart algorithm is used for classification. The classification results are then fused in the image domain through steps such as polarization despecciation processing, connected component analysis, and Gaussian filtering. The removal methods include: decomposing the polarization entropy H and polarization scattering angle α parameters, and combining the Wishart category decision criterion and Gaussian filtering to enhance image uniformity.
It achieves effective fusion of classification results of multi-band dual-polarization SAR images in the image domain, improving the accuracy and consistency of ground feature classification, and is applicable to multi-band images that cannot be fused at the data level.
Smart Images
Figure CN116682025B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of ground object classification, and particularly relates to a ground object classification method based on dual-band dual-polarization SAR images. BACKGROUND
[0002] Ground object classification is an important application direction of polarimetric SAR images, and has wide application prospects and importance in land survey, environmental protection, geology and agriculture. In particular in the field of agriculture, ground object classification can quickly evaluate crop growth conditions and help fine agricultural management. At present, the theory and method of ground object classification of polarimetric SAR images are relatively mature and have been put into practical use.
[0003] However, many polarimetric SAR ground object classification methods are based on single-band full-polarization SAR images, and many dual-polarization SAR images have not been effectively utilized. In addition, due to the differences in scattering characteristics of actual ground objects under different wave bands, the information contained in different wave band polarimetric SAR images is not completely the same. These polarimetric SAR images may have complementary effects on ground object classification, but the processing method of multi-band polarimetric SAR image fusion is still not very mature.
[0004] Considering that most dual-band polarimetric SAR images of different wave bands are collected by different devices at different times and places, it is difficult to fuse them at the data level, so it is necessary to fuse the classification results in the image domain based on the results of single-band dual-polarization SAR ground object classification. This has not been mentioned in the existing various multi-band dual-polarization SAR image ground object classification methods. SUMMARY
[0005] To solve the above problems, the application provides a ground object classification method based on dual-band dual-polarization SAR images, which can be used for overall ground object classification and distinguish different vegetation regions. After the registration of two wave band images, H / alpha Wishart algorithm is used for classification respectively, and finally the classification results are fused in the image domain.
[0006] The dual-band dual-polarization SAR image classification method provided by the application includes:
[0007] Step 1, register two dual-band dual-polarization SAR images of different wave bands in the same area, so that they have the same size and resolution. Before registration, the dual-band dual-polarization SAR images need to be processed into the form of polarimetric covariance matrix.
[0008] Step 2, perform polarimetric speckle reduction on the two images. Methods such as multi-view or fine Lee filtering can be used to suppress the coherent noise of the images.
[0009] Step 3, classify the two images respectively using dual-polarization H / a Wishart algorithm. The detailed steps of H / a Wishart algorithm are as follows.
[0010] Step 3-1, perform Cloude-Pottier decomposition on the polarimetric SAR image to obtain two parameters of polarization entropy H and polarization scattering angle a.
[0011] Step 3-2, divide the initial classes for all pixel points in the image according to the dual-polarization H-a plane.
[0012] Step 3-3, based on the initial classes, traverse each pixel point, and determine the new class of the pixel point according to the Wishart class decision criterion (the class of the pixel point may not change).
[0013] Step 3-4, repeat the operation of 3-3, and iterate until the proportion of pixel points with class changes in the traversal process is less than 10%. The final result is the classification result of the dual-polarization H / a Wishart algorithm.
[0014] Step 4, extract each class in the two images, determine the background class, and exclude it.
[0015] Step 5, for all classes (non-background classes) remaining in the two images, remove the fine parts in the classes, and the removal method is as follows.
[0016] For a class of an image, count the number of connected regions in the class and the size of each connected region, and arrange them in descending order according to the latter. Then, according to the window size used in polarization speckle reduction, discard the connected regions with an area smaller than the window size.
[0017] Step 6, merge the corresponding classes in the two images.
[0018] Step 7, perform multiple Gaussian filtering on the non-background classes of the two images to grow the connected regions and fill some small gaps to enhance the uniformity of the image. After each Gaussian filtering, the non-zero part of the image is considered as 1 to preserve the binary nature of the image. The number of Gaussian filtering can be around 2-3 times.
[0019] Step 8, superimpose all non-background classes. For pixels not belonging to any non-background class, directly consider them as the background class. In addition, some pixel points may have multiple non-background classes, and one needs to be selected as the final class, and the selection method is as follows.
[0020] Examine the area size of the local connected regions of each class in which the pixel point is located, and select the class with the largest area as the final class of the pixel point.
[0021] At this point, all steps are completed.
[0022] The present application has the advantages of:
[0023] The present application provides a new method of fusing the ground feature classification results of two bands of dual-polarized SAR images, which is suitable for ground feature classification of multi-band dual-polarized SAR images that cannot be fused at the data level. BRIEF DESCRIPTION OF DRAWINGS
[0024] Figure 1 The flow chart of the ground feature classification method based on dual-band dual-polarized SAR images is shown in
[0025] Figure 2 The dual-polarized H-alpha plane is shown in
[0026] Figure 3 The optical photo of the example area is shown in
[0027] Figure 4 The classification results of the two bands of images are shown in
[0028] Figure 5 The results of removing the fine parts of the L-band categories are shown in
[0029] Figure 6 The results of removing the fine parts of the C-band categories are shown in
[0030] Figure 7 The results of Gaussian filtering of the merged non-background categories are shown in
[0031] Figure 8 The final fused classification results are shown in DETAILED DESCRIPTION
[0032] The present application will be described in detail below in conjunction with the drawings and examples.
[0033] The flow chart of the present application is shown in Figure 1 The multi-band dual-polarized SAR images can be acquired by the same or different SAR satellites. The specific steps of the present application are as follows:
[0034] Step 1, register the two dual-polarized SAR images of the same area and different bands so that they have the same size and resolution. Before registration, the dual-polarized SAR images need to be processed into the form of polarization covariance matrix.
[0035] Step 2, perform polarization speckle reduction on the two images. Methods such as multi-view or refined Lee filtering can be used to suppress the coherent noise of the images.
[0036] Step 3, use dual-polarized H / alpha Wishart algorithm to classify the two images respectively. The specific steps of the H / alpha Wishart algorithm are as follows.
[0037] Step 3-1, Cloude-Pottier decomposition is applied to the polarimetric SAR image to obtain two parameters, the entropy H and the alpha angle a.
[0038] Step 3-2, the initial classes are assigned to all the pixels in the image according to the H-a plane, as shown in Figure 2 .
[0039] Step 3-3, based on the initial classes, the new classes of all the pixels are determined according to the Wishart classification criterion. The Wishart classification criterion is as follows.
[0040] First, the statistical average of the polarimetric covariance matrix of each class is calculated as the respective class center.
[0041] For any pixel, the distance between the pixel and each class is calculated using the following formula.
[0042]
[0043] where ω m denotes the class m, ∑ m represents the class center, and Tr(·) represents the trace of the matrix. After obtaining the distance between the pixel and each class, the following decision criterion is applied to determine the new class of the pixel: for all j≠m, if d(C, ω m )≤d(C, ω j ), then the pixel belongs to class m.
[0044] Step 3-4, repeat the operation of 3-3, and iterate until the proportion of pixels that change classes in the iteration process is less than 10%. The final result is the classification result of the dual-polarization H / a Wishart algorithm.
[0045] Step 4, extract each class in the two images, determine the background class, and exclude it.
[0046] Step 5, for all the remaining classes (non-background classes) in the two images, remove the fine parts in the classes, and the removal method is as follows.
[0047] For a class in an image, count the number of connected regions in the class and the size of each connected region, and arrange them in descending order according to the latter. Then, according to the window size used in polarization speckle reduction, discard the connected regions with an area smaller than the window size.
[0048] Step 6, merge the corresponding classes in the two images.
[0049] Step 7, for the non-background classes of the two images, do several times of Gaussian filtering, grow the connected regions, fill some small gaps, to enhance the image uniformity. After each time of Gaussian filtering, all the non-zero parts of the image are regarded as 1, to keep the binary nature of the image. The number of times of Gaussian filtering can be around 2-3.
[0050] Step 8, superimpose all the non-background classes. For the pixels not belonging to any non-background class, directly regard them as the background class. In addition, some pixels can have multiple non-background classes, and one needs to be selected as the final class, the selection method is as follows.
[0051] Investigate the area size of the local connected region of each class of the pixel, and select the class with the largest area as the final class of the pixel.
[0052] At this point, all the steps are completed.
[0053] Next, a specific example is given.
[0054] This example selects a farmland area in Perth, Australia. The optical photo of the area is shown in Figure 3 The area mainly has three different characteristics of ground objects, i.e. dense plantation area (mainly Figure 3 triangular dark area in the middle), sparse vegetation area, and empty land (including roads). This example uses two spaceborne dual-polarized SAR images of different bands, and the specific data parameters are shown in Table 1.
[0055] Table 1
[0056] First image Second image Acquisition platform ALOS-2 satellite Sentinel-1 satellite Acquisition time 23 August 2019 24 August 2019 Polarization mode HH + HV VV + VH Band L C
[0057] Step 1 is performed, and the two images are registered.
[0058] Step 2 is performed, and speckle reduction is performed on the two images, and the speckle reduction method is to use Lee filtering algorithm with a 7x7 window.
[0059] Step 3 is performed, and the classification results of the two images are obtained, as shown in Figure 4 The number of classes of the two images is 4, of which the light blue is the background class.
[0060] Step 4 is performed, and the non-background classes of the two images are extracted.
[0061] Step 5 is performed, and the fine parts of the non-background classes are removed. Figure 5 and Figure 6 show the results of the L-band and C-band classes after the fine parts are removed, respectively.
[0062] Step 6 is performed, and the corresponding classes in the two images are merged.
[0063] Perform step 7, Gaussian filter the merged classes. Figure 7 The result of Gaussian filtering.
[0064] Perform step 8, combine all the classes to get the final classification result. Figure 8 The final classification result is shown. Overall, different plots can be well distinguished, indicating the effectiveness of the method.
[0065] Of course, the present application can have other various embodiments, and those skilled in the art can make various corresponding changes and modifications according to the present application without departing from the spirit and essence of the present application, and these corresponding changes and modifications shall all belong to the protection scope of the claims attached to the present application.
Claims
1. A method for ground object classification based on dual-band dual-polarization SAR images, characterized in that The method comprises the following steps: Step 1, two dual-polarization SAR images of the same area and different wave bands are registered so as to have the same size and resolution, and before registration, the dual-polarization SAR images are processed into the form of polarization covariance matrix; Step 2, the polarization speckle reduction is performed on the two images; Step 3, use dual polarization on both images separately Wishart algorithm for classification; Step 4, each class in the two images is extracted, the background class is judged and excluded; Step 5, for all the classes of the two images, i.e. all the non-background classes, the number of connected regions in a class and the size of each connected region are counted for a class of one image, and the latter is arranged from large to small; then, according to the window size used in polarization speckle reduction, the connected regions with an area smaller than the window size are discarded; Step 6, the corresponding classes in the two images are merged; Step 7, the non-background classes of the two images are subjected to multiple Gaussian filtering to grow the connected regions and fill some small gaps, so as to enhance the image uniformity; Step 8, all the non-background classes are superimposed and combined.
2. The method of classifying ground objects based on dual-band dual-polarization SAR images according to claim 1, characterized in that In step 3, the following is used Wishart algorithm for classification.
3. The method of classifying ground objects based on dual-band dual-polarization SAR images according to claim 1, characterized in that In step 3, the polarimetric SAR image is decomposed by Cloude-Pottier to obtain the polarimetric entropy and the polarimetric scattering angle Two parameters, according to dual-polarization The plane is used to divide the initial class of all pixel points in the image.
4. The method of ground object classification based on dual-band dual-polarization SAR images according to claim 1, characterized in that In the step 3, the distance between any pixel and each class is calculated by using the following formula: ； wherein denotes the class m, denotes the center of its class, denotes the trace of the matrix; after obtaining the distance of the point to each class, the following decision criterion is applied to determine its new class: for all , if then the point belongs to class m.
5. The method of classifying ground objects based on dual-band dual-polarization SAR images according to claim 1, characterized in that In the step 7, after the classification of the two images, the non-background classes of the two images are subjected to multiple Gaussian filtering to enhance the image uniformity.
6. The method of classifying ground objects based on dual-band dual-polarization SAR images according to claim 1, characterized in that In the step 8, the largest class is selected as the final class of the pixel.