An image classification method based on multi-baseline tomographic polarization target decomposition

Abstract

The application provides an image classification method based on multi-baseline tomographic polarimetric target decomposition, which is used for expanding the image classification application based on scattering characteristics and height information.The method comprises the following steps: performing coherent speckle filtering and other pretreatments on an image, and screening smooth areas; performing polarimetric tomographic decomposition on polarimetric tomographic data, and dividing the image pixels into three categories of surface scattering, even scattering and volume scattering; performing in-class expansion and merging on the pixels in the three categories according to scattering weights to realize initial classification; and performing optimization iteration on the initial classification until a predetermined iteration number is reached or the change value of the category center is small.The application further expands the application of polarimetric interferometric tomographic SAR target information extraction and image classification.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of Tomographic Synthetic Aperture Radar (TomoSAR), and particularly relates to an image classification method based on multi-baseline tomographic polarimetric target decomposition. BACKGROUND

[0002] The polarimetric interferometric SAR tomography technology is a combination of SAR polarization and interferometric tomography technology, so that the high-resolution imaging radar has the ability of target electromagnetic feature detection and target space structure and environment perception, which makes the polarimetric interferometric SAR tomography technology have further exploration space in target detection, identification, feature parameter extraction and other aspects. The polarimetric tomographic target decomposition can obtain multi-dimensional feature information of the target, and provides key information support for radar target identification and image classification. At present, the image classification method combined with the polarimetric tomographic target decomposition algorithm is still relatively few, and it is urgent to further explore the multi-dimensional information obtained by the polarimetric tomographic target decomposition, construct an efficient image classifier, and further expand the application of polarimetric interferometric SAR tomographic target information extraction and image classification.

[0003] With the massive data acquisition of repeated flight of airborne and spaceborne polarimetric SAR systems, the hard condition for realizing vertical range synthesis aperture technology has been met. Deeply mining the target characteristics in the polarimetric interferometric SAR tomographic image and extracting the target information meeting the application requirements from the image data are the key steps of radar target identification and image classification. However, the research on the image classification algorithm based on the polarimetric interferometric SAR tomography is still relatively few. SUMMARY

[0004] In view of this, the present application aims to provide an image classification method based on multi-baseline tomographic polarimetric target decomposition, and expand the application of polarimetric tomographic multi-information features to image classification.

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows:

[0006] An image classification method based on multi-baseline tomographic polarimetric target decomposition comprises the following steps:

[0007] Step 1, pre-processing the image by coherent speckle filtering and the like, and screening the smooth area.

[0008] Step 2, performing polarimetric tomographic decomposition on the polarimetric tomographic data, and dividing the image pixels into three categories of surface scattering, even scattering and volume scattering.

[0009] Step 3, performing intra-class expansion and merging in the three categories to realize initial classification.

[0010] Step 4, optimize iteration of initial classification until a predetermined number of iterations or a pixel number change rate of class center is less than 5%.

[0011] Further, the step 1 comprises:

[0012] (1) pre-process the polarimetric tomography SAR image to be processed, such as coherent speckle filtering.

[0013] (2) judge the smooth area based on the total power data of the image, and divide the smooth area into surface scattering class.

[0014] Further, the step 2 comprises:

[0015] (1) target decomposition of single phase center is performed on the polarimetric tomography data to obtain the weight and corresponding phase center height of surface scattering, double scattering and volume scattering components.

[0016] (2) the height of the target is estimated by using the phase center height of the surface scattering, double scattering and volume scattering components.

[0017] (3) for the pixel points judged as non-smooth area, they are divided into three categories of surface scattering, double scattering and volume scattering according to their dominant scattering type.

[0018] Further, the step 3 comprises:

[0019] (1) according to the power value of the three scattering components of the pixel points, the pixels in each scattering class are divided into at least 30 small classes with similar number of pixels.

[0020] (2) according to the principle of minimum average height difference of initial small class, a plurality of small classes are clustered and merged into a predetermined number of initial classification classes.

[0021] Further, the step 4 comprises:

[0022] (1) construct the probability density function of the Wishart mixed distribution model (GWMM) based on polarimetric tomography.

[0023] (2) reclassify the sample points according to the principle of maximum similarity.

[0024] (3) repeat steps (1)-(2) until a predetermined number of iterations or a pixel number change rate of class center is less than 5%.

[0025] Beneficial effects:

[0026] An image classification method based on multi-baseline tomographic polarimetric decomposition technique is proposed in the paper, which can effectively improve the image classification effect by increasing the target height information during the polarimetric feature extraction, and further expand the application of polarimetric interferometric tomography SAR target information extraction and image classification. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a flow diagram of an image classification method based on multi-baseline tomographic polarimetric target decomposition of the present application;

[0028] Figure 2a Figure 2b is an experimental data diagram; wherein, Figure 2a is an optical control diagram of experimental data, Figure 2b is a synthesis diagram of experimental data;

[0029] Figure 3 is an initial classification result;

[0030] Figure 4 is the classification result after iteration optimization. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0032] The present application performs polarimetric target decomposition and phase center height estimation on polarimetric interferometric tomography data, so as to realize image classification. The image is preprocessed by coherent speckle filtering and the like, and the smooth area is screened. The polarimetric tomography data is subjected to polarimetric tomography decomposition, and the image pixels are divided into three categories of surface scattering, even scattering and volume scattering. The three categories are expanded and merged within the class according to the scattering weight to realize initial classification. The initial classification is iteratively optimized until a predetermined number of iterations is reached or the variation of the class center is small.

[0033] As shown in Figure 1 , the image classification method based on multi-baseline tomographic polarimetric target decomposition of the present application specifically comprises the following steps:

[0034] Step 1, the image is preprocessed by coherent speckle filtering and the like, and the smooth area is screened. Specifically, it includes:

[0035] (1) the polarimetric tomography SAR image to be processed is preprocessed by coherent speckle filtering and the like.

[0036] ​(2) Based on the total power data of the image to determine the smooth area, and divide the smooth area into surface scattering type.

[0037] In order to more accurately distinguish the surface scattering type target, Harris feature detection based on image feature detection is adopted to solve the image pixel gray gradient and calculate the x and y direction gradients of each pixel point in the image. The calculation method is as follows:

[0038]

[0039] Where I(x, y) is the image amplitude value at pixel position (x, y).

[0040] The matrix is constructed, Where I x , I y is the x and y direction gradient of each pixel point.

[0041] Calculate the parameter R to determine the smooth and uniform area:

[0042] R = detM - k (traceM) 2 ,

[0043] Where detM represents the determinant of matrix M, traceM represents the trace of matrix M, and k is an empirical value, which is in the range of 0.04-0.06.

[0044] When the absolute value of R is very small, that is, when the following formula is satisfied, the pixel is determined to be a smooth area.

[0045] | R | ≤ threshold (2)

[0046] Where threshold is the threshold for determining the smooth area.

[0047] Step 2, polarized tomographic data is decomposed by polarized tomography, and the image pixels are divided into three categories of surface scattering, even scattering and body scattering. Specifically including:

[0048] (1) Single phase center target decomposition is performed on the polarized tomographic data to obtain the weight and corresponding phase center height of the surface scattering component, even scattering component and body scattering component.

[0049] (2) The height of the target is estimated by using the phase center height of the surface scattering component, even scattering component and body scattering component.

[0050] (3) For the pixel points determined to be non-smooth areas, they are divided into three categories of surface scattering, even scattering and body scattering according to their dominant scattering type.

[0051] Step 3, intra-class expansion and merging of surface scattering, multiple scattering and volume scattering to achieve initial classification. Specifically includes:

[0052] (1) According to the three scattering component power values of the pixel points, the pixels in each scattering category are divided into 30 (or more) small classes with similar number of pixels.

[0053] (2) According to the principle of minimum average height difference of initial small class, multiple small classes are clustered and merged into the pre-set initial classification category. That is, by comparing the average height of the initial small class, the two small classes with the minimum average height difference are merged into a new small class, and the process is repeated until the remaining class number reaches the pre-set initial classification category.

[0054] Step 4, optimization iteration of initial classification using Wishart classifier until the predetermined iteration number or the pixel number change rate of the class center is less than 5%. Specifically includes:

[0055] (1) Construct the probability density function of the Wishart mixed distribution model (GWMM) based on polarization tomography.

[0056]

[0057] Where G represents any sample point, subscript K represents the Kth class, Y i k and π i represent the class center and weight coefficient of the kth class of the i-th image, and L is the number of multi-view. A norm is a normalization factor used to ensure that the probability density function in the above formula is integrated to 1. It can be found that when L is 1, it is the ordinary Wishart distribution probability density function. Where, the similarity parameter η β is as follows:

[0058] η β (T, V m ) = exp{-β·d(T, V m )} (4)

[0059] Where β is a positive hyperparameter, d(T, V m ) is a distance measure, which can use the Wishart distance proposed by Lee et al. or any other effective distance measure, T is the polarization coherence matrix of the pixel point to be classified, and V m is the average polarization coherence matrix of the classified class. In this way, the distance between the two samples can be mapped to 0 to 1 through the kernel function to represent the similarity between the two.

[0060] (2) According to the maximum similarity principle, the class of each sample point is re-divided. That is, the similarity probability of the pixel to a certain class is selected as the class of the pixel:

[0061] G∈K|p(G|K) max =p(G|K=1,2,...M) (5)

[0062] where M is the total number of classes.

[0063] (3) Repeat steps (1)-(2) until a predetermined number of iterations is reached or the values of the class centers change very little.

[0064] The multi-baseline polarimetric SAR data used in the experiment verification of the present application is from the European Space Agency (ESA) 2009 airborne SAR remote sensing experiment in tropical forest (TropiSAR 2009), which is composed of P-band full-polarimetric SAR data obtained by 6-track repeated flight. The airborne data was obtained by the SETHI radar system developed by the French National Aerospace Research Center (ONERA) at the Parakou Research Base in French Guiana in August 2009. The data has been calibrated and registered, and the right lower part of the sub-image in the panoramic image is intercepted for polarimetric tomography experiment, as shown in Figure 2a , Figure 2b The area contains typical features such as forest land (A area), flat land (B area), building targets (C area), etc., as shown in the following figure. It should be noted that the imaging time of the Google earth optical image is different from that of the SAR image, and the SAR image was obtained in 2009, while the main part of the Google earth image was obtained in 2012, although it can be determined that area C is a small building area with relatively concentrated buildings, but the location and details of the buildings cannot be better corresponded.

[0065] The polarimetric tomography data of Figure 2a , Figure 2b are initially classified according to the above principle, and the even scattering and volume scattering are each divided into 4 classes, and the surface scattering is divided into 8 classes, and the results are shown in Figure 3 The experimental results show that the initial classification method can clearly distinguish the location and details of the forest area, flat area and building area, and especially the different height features of the forest area can also be effectively identified. The classification results using scattering feature weight analysis can not only effectively identify the scattering mechanism of different targets, but also can distinguish the height structure features of the same large class of targets according to the height estimation value, and can be refined into different tone subclasses, and the overall classification effect is good. Using the initial results of polarimetric tomography classification shown in Figure 3 , the measured polarimetric tomography data is processed by combining the above similarity classifier, and the data classification results after cyclic iteration are as shown in Figure 4The experimental results show that the categories of the forest target, the building target and the flat region in the image after the iteration optimization are more abundant, and the texture is more clear, and the classification result is more complex optical image features.

[0066] Those skilled in the art will easily understand that the above description is only the preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An image classification method based on multi-baseline tomographic polarization target decomposition, characterized in that, Specifically comprising the following steps: Step 1, coherent speckle filtering preprocessing is performed on the image, and a smooth region is screened out; Step 2, polarimetric decomposition is performed on the polarimetric tomography data, and the image pixels are divided into three categories of surface scattering, even scattering and volume scattering, including: (1) single phase center target decomposition is performed on the polarimetric tomography data to obtain the weight and corresponding phase center height of the surface scattering, even scattering and volume scattering components; (2) the height of the target is estimated by using the phase center height of the surface scattering, even scattering and volume scattering components; (3) for the pixel points determined as non-smooth regions, they are divided into three categories of surface scattering, even scattering and volume scattering according to their dominant scattering type; Step 3, the pixels in the three categories of surface scattering, even scattering and volume scattering are expanded and merged within the class to realize initial classification, including: (1) the pixels in each scattering category are divided into at least 30 initial small categories with similar pixel numbers according to the power values of the three scattering components of the pixel points; (2) multiple small categories are clustered and merged into the initial classification category number according to the minimum average height difference principle of the initial small categories; Step 4, the initial classification is optimized and iterated until the predetermined iteration number or the pixel number change rate of the category center is less than 5%.

2. The image classification method based on multi-baseline tomographic polarization target decomposition according to claim 1, characterized in that, The step 1 comprises: (1) coherent speckle filtering preprocessing is performed on the polarimetric tomography SAR image to be processed; (2) the smooth region is divided into the surface scattering category based on the total power data of the image.

3. The image classification method based on multi-baseline tomographic polarization target decomposition according to claim 2, characterized in that, The step 4 comprises: (1) the probability density function of the Wishart mixed distribution model based on polarimetric tomography is constructed; (2) the categories of the sample points are redivided according to the maximum similarity principle; (3) steps (1)-(2) are repeated until the predetermined iteration number or the pixel number change rate of the category center is less than 5%.

Images