A SAR-GMTI Approach Aided by Online Knowledge With an Airborne Multichannel Quad-Pol Radar System

In the complicated geographical environment, there will be a seriously deleterious effect to the performance of synthetic aperture radar (SAR)-ground moving target indication (SAR-GMTI) system, because it is difficult to obtain the homogeneous training samples to accurately estimate the clutter covariance matrix (CCM) without prior information of the observed scene. To this end, this article proposes a SAR-GMTI approach aided by online knowledge with an airborne multichannel quadrature-polarimetric (quad-pol) radar system. Generally, this article can be divided into two parts: online knowledge acquisition and polarization knowledge-aided (Pol-KA) SAR-GMTI processing. First, based on the similarity of pixels from the multichannel and multipolarization information, a weighed estimation method of polarimetric coherency matrix is proposed, which can overcome the over-smoothing problem and increase the estimation accuracy of coherency matrix. Furthermore, a hybrid weighted local K-means based on geodesic distance (GD-HWLKM) clustering algorithm is proposed to achieve the aim of unsupervised classification. Here, GD is exploited to measure the distance between multifeature region covariance matrixes (MFRCMs) and a hybrid weight from different scales (including local cover class distribution, region, and pixel) is calculated to automatically update the cluster centroid, which can make full use of the local spatial information by taking the interclass samples’ similarity and the diversity of different classes into consideration. Second, with the assistance of the previous polarization SAR (PolSAR) image classification result, a Pol-KA SAR-GMTI method is developed. For each ground cover category, an accurate CCM can be estimated with the independent and identically distributed (IID) training samples. Then, the multichannel clutter suppression and preliminary constant false alarm rate (CFAR) detection are performed. Finally, with an airborne multichannel quad-pol radar system, the experimental results on real measured data demonstrate that the proposed method can efficiently improve the clutter suppression preformation and moving-target detection preformation.


I. INTRODUCTION
W ITH ensuring high-resolution, synthetic aperture radar (SAR)-ground moving target indication (SAR-GMTI) technology is capable of velocity estimation and localization of ground moving targets [1], [2]. In the actual environments, the ground moving targets usually are submerged in strong clutter. In order to reliably detect the moving targets, the azimuth multichannel methods are proposed, such as imaging spacetime adaptive processing (ISTAP) [3], extended displaced phase center antenna (EDPCA) [4], and joint pixel vector processing [5]. The above-mentioned methods can obtain a satisfactory clutter suppression performance with a relatively accurate clutter covariance matrix (CCM) for the homogeneous clutter scene. But, in the actual work, the geographical environment observed is often complicated and the clutter distribution is inhomogeneous, such as sea-land, suburban or shadow-covered mountains. The traditional sample covariance matrix (SCM) method [6] ordinarily estimates the CCM by using the clutter samples from all distances. The power and distribution characteristics of training samples are always different with the range cell under test, which causes a significant reduction for the clutter suppression performance. The power selection training (PST) method [7] can solve the problem that the clutter suppression performance is insufficient by selecting the high power samples and can effectively suppress the strong clutter. However, this method certainly broadens the clutter adapting notch with the high-power samples, which will reduce the detection capability of the moving targets.
The prior knowledge-aided (KA) method is a possible way to circumvent this drawback by the prior environment information surrounding the radar [8], [9], [10], which selects representative training samples to construct the CCM aided by the digital terrain elevation models (DTEMs), the geographic information systems (GISs) and the land cover categories map. It can effectively suit inhomogeneous environment, but the timeliness of the prior This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ knowledge limits its widely application. In addition, system errors such as the platform's position error, attitude error, or beam pointing error may lead to the clutter mode of the recorded data mismatch with the prior knowledge, which will greatly reduce the estimation accuracy of CCM. However, the classification technology of polarimetric synthetic aperture radar (PolSAR) images can provide the online geographic or land cover information to help to robustly estimate CCMs in SAR-GMTI processing for the relatively identical clutter categories.
The clutter classification techniques contain supervised classification [11], [12], [13], [14], semisupervised classification [15], [16], [17], [18], and unsupervised classification [19], [20], [21], [22], [23], [24]. However, for SAR-GMTI processing, the mismatch between the land cover categories of the recorded data and the prior knowledge usually occurs because of the inevitable system errors, which make it difficult to obtain the labeled training samples. Therefore, the unsupervised classification methods are the most appropriate approaches to help to improve the SAR-GMTI processing performance. The unsupervised methods usually represent the inner structure information of unlabeled data by designing a function and make a decision rule to cluster samples into several different groups. Without any prior knowledge, the unsupervised methods have been widely considered, such as Wishart classifier [19], Markov random field methods [20], [21], polarimetric scattering characteristic-preserved method [22], K-means cluster classifier [23], fuzzy c-means cluster classifier [24], and so on. Owing to the simplicity of operation and small computational complexity, K-means clustering approach has become one of the most popular PolSAR image classification methods, which can be applied to improve the preformation of the actual SAR-GMTI systems in the increasingly diverse surveillance tasks.
The K-means approach is an iterative optimization algorithm, which mainly is related to three aspects: the representation of samples, the distance metric between unlabeled samples and the cluster centroids, and the updated mode of cluster centroids, as follows: 1) is the representation of samples. In fact, not only for the K-means approaches, the representation of samples is usually an important factor to the great mass of PolSAR classification methods. The classical representation way is based on pixels. Lee et al. [25] classify the PolSAR image pixels using K-means algorithm with the multivariate complex Wishart distribution of sample coherency matrices. Chen et al. [26] achieve PolSAR image unsupervised classification based on the scattering similarity between pixels. Bi et al. [27] also propose an unsupervised Pol-SAR image classification using discriminative clustering. But, the pixel-based classification methods are easily affected by the speckle noise. Recently, region covariance-based classification has been widely applied to PolSAR image classification [28], [29]. As one type of representation of samples, region covariance matrix (RCM) considers the relation between neighbor pixels and can efficiently suppress the speckle noise. Besides, compared with the pixel vector, RCM-based representation can take the correlated properties of features into account by providing both the variance of each feature (i.e., matrix's diagonal entries) as well as the joint distribution between them (i.e., matrix's off-diagonal entries). 2) is the distance metric between unlabeled samples and the cluster centroids. The classical or standard distance metric is norm-based metric, including Euclidean distance [30], Manhattan distance [31], log-Euclidean distance [32], and so on. Furthermore, the Wishart-based [23] dissimilarity measure has become a popular distance metric. Song et al. [33] adopt the Bartlett distance to replace the similarity measure in PolSAR classification tasks. Kersten et al. [24] have identified that the distance measures based on the Wishart distribution have a better unsupervised classification performance than those based on the norms. In the past few years, scholars find that the covariance matrices with semipositive definite structure do not lie on the Euclidean space but on the Riemannian space. So, the affine invariant Riemannian (AIR) distance [29], [34] is proposed to define the distance between covariance matrices. But, the AIR distance requires eigenvalue decomposition operation and has a high computation complexity. 3) is the updated mode of cluster centroids. The standard K-means algorithm equally treats all variables and do not select samples in the clustering process, which usually cannot obtain a satisfactory result. Many works about variable weighted clustering have been progressed to improve the performance by identifying important variables with variable weights [35], [36]. This updated mode of cluster centroids can be seen as feature extraction. Xiang et al. [37] update the cluster centroids with a weighted local K-means algorithm by measuring the similarity of pixels, which can improve the clustering result but it is easily affected by the speckle noise. In addition, the aforementioned PolSAR image classification methods are based on the PolSAR systems with one azimuth receive channel.
This article mainly pays attention to improve SAR-GMTI performance aided by online PolSAR image classification knowledge with the multipolarization and multichannel SAR-GMTI system. As shown in Fig. 1, this article mainly can be divided into two parts: online knowledge acquisition by the proposed hybrid weighted local K-means based on geodesic distance (GD-HWLKM) clustering algorithm and online polarization knowledge-aided (Pol-KA) SAR-GMTI processing. In part I, we first propose a weighed estimation method of coherency matrix for multichannel PolSAR system to robustly estimate the polarimetric coherency matrix, which can overcome the over-smoothing problem and increase the estimation accuracy of coherency matrix by effectively taking advantage of the multichannel and multipolarization information. Then, the significant polarimetric features belonging to different categories are extracted with multifeature RCM (MFRCM), including the polarimetric decomposition-based, color, and texture features. Different from the existing feature selection methods, we extract the texture features from the scattering model-based decomposition features map, which can represent the structure of different physical scattering components better. Due to the MFRCMs lie on the Riemannian space, GD [38], [39] is exploited to describe the similarity distance between them, which can effectively tolerate the scaling and rotation between MFRCMs. And, we also discuss other distance metrics. Furthermore, to achieve the aim of unsupervised classification, a GD-HWLKM clustering algorithm is proposed within the K-means algorithm framework. Where the hybrid weight is calculated to describe the interclass samples' similarity and the diversity of different classes from different scales, including local cover class distribution, region, and pixel, which can make full use of the local spatial information and can preserve the edges as many as possible. In part II, an online Pol-KA SAR-GMTI processing method is developed with the assistance of the previous Pol-SAR image classification results. For each single-polarimetric (S-pol) component, the independent and identically distributed (IID) clutter training samples are selected to accurately estimate the CCM of each land cover. Then, the clutter cancellation in image domain for each land cover and preliminary constant false alarm rate (CFAR) detection are performed. Besides, the noncoherence integration detection (NCID) technique is exploited to reduce the false alarms, further. Finally, the parameters estimation step is carried out to obtain the velocity of moving targets.
The main contributions of this article are listed as follows.
1) The proposed weighed estimation method of coherency matrix not only can reduce the effect of speckle noise when solving the over-smoothing problem, but also can increase the estimation accuracy of coherency matrix by the local weight based on the similarity of pixels from the multichannel and multipolarimetric information.
2) The proposed unsupervised GD-HWLKM PolSAR image clustering approach can achieve a better classification performance while preserving the edges. This algorithm first utilizes GD to measure the distance between MFR-CMs, which is more effective for classification. Meanwhile, considering the interclass samples' similarity and the diversity of different classes, a hybrid local weight is constructed to revise the KM algorithm from different scales, which can preserve the image edges and obtain the better classification result. 3) An online Pol-KA SAR-GMTI method is developed based on the airborne multichannel quad-pol radar system. Under the guidance of the online knowledge about the land scene, the IID clutter samples are selected to accurately estimate CCM for each land cover, which can improve clutter suppression and moving-target detection performance. The rest of this article is organized as follows. Section II gives an introduction about the operation of airborne multichannel quad-PolSAR systems. In Section III, the proposed method is described in detail. And the experiment results based on the real measured data are presented in Section IV. Finally, Section V concludes this article.

II. OPERATION OF AIRBORNE MULTICHANNEL QUAD-POLSAR SYSTEM
Since the multichannel and multipolarization data are used in this article, the quad-pol airborne radar system with multiple azimuth receiving channels is brief introduced as follows: As shown in Fig. 2, the airborne quad-pol radar system [40] operates in side-looking mode with platform velocity V p , platform height h, including three azimuth receiving channels. The H-and V-polarization echoes can be simultaneously received for each azimuth receiving channel. H-polarization channel transmits the horizontally polarized radar pulses with azimuth full aperture D a and three spatial receiving channels receive the returns with H polarization and V polarization, simultaneously, which can form the HH and HV polarization components. Similarly, the VH and VV components are formed when V-polarization channel transmits the vertically polarized radar pulses. Then, HH-pol, HV-pol, VH-pol, and VV-pol components of each azimuth receiving channels are performed SAR image processing to obtain the multichannel and multipolarization SAR images.

III. METHODOLOGY
The framework of the proposed method can be seen in Fig. 1. This article mainly can be divided into two parts: online knowledge acquisition and Pol-KA GMTI processing. First, the online Pol-KA GMTI processing aided by polarimetric classification result is developed. Second, online knowledge acquisition is given in detail, which contains three main stages including the weighed estimation of coherency matrix with the multichannel SAR system, the multiple feature extraction with RCMs, and the unsupervised classification by the proposed GD-HWLKM clustering algorithm. We now describe them in detail.

A. Multichannel SAR-GMTI Processing Aided by Online Polarimetric Classification Knowledge 1) Signal Model and Traditional CCM Estimation Method:
In multichannel GMTI processing, the detection for the pth pixel in image domain can be expressed as follows: , and n(p) ∈ C M ×1 represent the pixel data vector, moving-target signal vector, clutter vector, and noise vector, respectively. γ p denotes the complex amplitude of target and M is the number of spatial receiving channel.
To suppress the clutter, the optimal weight calculation is given in image domain by the linear constraint minimum variance (LCMV) criterion where [·] H represents the conjugate transpose operation, s(v r ) is the multichannel spatial steering vectors of moving-target with the radial velocity v r , w is the optimum weight, and R cn is the ideal CCM of the RCUT. Obviously, the CCM R cn of the RCUT only can be obtained by estimating. And, the well-known Reed-Mallett-Brennan (RMB) rule [6] has proved that it requires the IID training samples and can achieve an averaging performance loss within 3 decibels (dB). In the traditional CCM estimation method [3], [4], [5], the CCM is estimated by the pixel data vectors of adjoining range cells, which are regarded as training samples, as follows: whereR cn is the estimation of R cn with traditional methods and K is the number of training samples.

2) Proposed Online Pol-KA SAR-GMTI Method:
In the complicated observation environment, the various land covers usually have different clutter scattering and statistic characteristics, which often result in severe performance degradation of the traditional methods. To this end, it is a core issue to estimate CCMs by selecting the homogeneous training samples for achieving a great GMTI performance.
In this article, an online Pol-KA SAR-GMTI method is developed with the guidance of polarimetric classification. The accurate CCM for each land cover can be estimated with the clutter samples belonging to the same cover category, which have the same scattering characteristic and can be regarded as homogeneous clutter and generally satisfy IID, as follows: whereR cn (l) is the estimation value of R cn by the proposed method for the lth (l = 1, 2, · · · , L) land cover category, C l is the lth clutter samples dataset from the online polarimetric classification result, k ∈ C l represents the training samples belonging to C l . η l denotes a weighting real coefficient of samples data for the lth land cover from polarimetric classification result, which satisfies η l = 1/N l and N l is the number of lth land cover clutter samples dataset.
After obtaining the CCM of each land cover, the optimum can be employed to suppress the clutter. The detection performance is directly proportional to the output signal-to-clutter-plus-noise ratio (SCNR) at the output of the adaptive matched filter for clutter suppression, which can be calculated as follows: where R s is the multichannel data matrix of the moving target. Different from the KA methods [8], [9], [10], the proposed method is more suitable to the inhomogeneous environment without any prior knowledge. In the KA methods, CCM is estimated withR KA = ζ · R 0 + (1 − ζ) ·R cn , whereR cn is the estimation value of CCM in (3), R 0 is the constructed CCM by exploited the knowledge of radar, and ζ is the weight coefficient. Evidently, the proposed method is entirely based on the receiving data, which can absolutely suit the system working state and avoid the mismatch between the prior knowledge and the receiving data. And, it can be widely applied to the complicated environment, such as sea-land, suburban, urban, shadow-covered mountains, and so on. In summary, the proposed method can obtain the accurate estimation values of CCM for each ground cover under the guidance of the obtained online classification knowledge. Therefore, the online knowledge acquisition suiting the multichannel and multipolarimetric system is the research focus for the next article.

B. Weighed Estimation Method of Coherency Matrix
As a major polarimetric carrier, the estimation of coherency matrix is a fatal step for the PolSAR images classification. Traditionally, an averaging operator is used to perform the spatial multilook processing to reduce the effect of speckle noise. But, this averaging filter will unavoidably cause the over-smoothing problem and debase the accuracy of the estimated coherency matrix. To address this problem, in this article, we propose a weighed estimation method of coherency matrix for multichannel PolSAR system.
As we all know, the multichannel data are the reflection echoes obtained from the radar at the same time and the same scene. So, there is only a little difference for an arbitrary receiving channel relative to the reference channel, which is mainly reflected in correlation coefficient (or interferometric phase). The bigger correlation coefficient is; the more similar coherence matrixes are. In this article, we propose to fuse the coherency matrixes from different receiving channels by weighted estimation with correlation coefficients between channels. This step can be seen as a multilook operation among channels, as follows: where (T p ) m is the estimated coherency matrix of the pth pixel with mth receiving channel, ρ r,m represents the correlation coefficient between rth and mth channels, which can be calculated in [41]. Here, it should be noted that we generally assume the rth receiving channel is the reference channel and the correlation coefficient is equal to 1 when m = r. For coherency matrix of each receiving channel, we not only consider the pixel similarity of polarimetric, but also take the interferometric phase information between channels into consideration by a pixel similarity weight during multilook procedure. If the pixels are more similar, their similarity weight will be bigger. For the mth spatial receiving channel, the weighted coherency matrix for the pth pixel can be estimated as follows: where 1) N p represents the neighborhood around pixel p.
2) k is the complex polarinetric scattering vector in Pauli representation in the mth receiving channel, as follows: 3) ω p (q) is the similarity weight of qth pixel inside the neighborhood N p , which is defined as follows: where d(p, q) = g p − g q 2 involves a dissimilarity measure between pixel p and q. and T is the multichannel and multipolarimetric united data vector. It should be noted that the interferometric phase information between channels also is taken into account different with the traditional method, which can robustly measure the similarity of pixels in neighborhood.σ p = π/2MAD is the scale parameter, which can be estimated by the mean absolute deviation (MAD) [42] of all the distance value d(p, q), ∀q ∈ N p .

C. Feature Extraction With Multifeature RCM
It is obvious that the features based on the decomposition theories are just linear transformations from the original polarimetric features (including the complex scattering matrix, covariance matrix, coherency matrix elements, and so on). So, for reducing feature redundancies, the features based on the different decomposition theories are only considered to extract the polarimetric scattering information. Besides, different from the existing methods, the texture features are extracted from the scattering model-based decomposition features map in this article, which can describe the structure of different physical scattering components better. As shown in Fig. 3, the extracted PolSAR features can generally be divided into two different categories: polarimetric features and the spatial structure features.
1) belongs to the polarimetric features usually based on the different decomposition theories, which try to describe the average scattering power by some independent components. In this article, scattering model-based decomposition (Yamaguchi decomposition) [43] and eigenvaluebased decomposition (H/A/α decomposition) [44]   They are calculated using mean ratio operator [46]. It should be noted that the gradients are extracted with scattering model-based decomposition feature images of PolSAR but not with original images, which can take advantage of spatial and polarimetric features as much as possible. In addition, the color features are sufficiently represented by the color histograms to describe the color distributions for each pixel, which can commendably adapt to the human eye system. Detailedly, for each pixel, we use a local sliding window to select the neighborhood pixels and count the number of pixel belonging to each color level. In this article, the RGB image is first composited with the double-bounce scattering (red), volume scattering (green) and surface scattering (blue). Then, according to the human vision, this RGB image is converted to the hue-saturation-value (HSV) images, which contain 7, 3 and 3 color levels of the hue, saturation and value, respectively. Therefore, the color features contain 13 bins, as f Color where h i , s i and v i is the hue, saturation, value bin, respectively. For normalization, all features can be linearly into [−1, 1]. In a word, this article extracts multiple significant features are listed in Table I.
As a local feature descriptor, RCM considers the relation between neighbor pixels and can efficiently suppress the speckle noise. Besides, compared with the pixel feature vector, RCM can take into account their correlated properties by providing both the variance of each feature and their joint distribution. We first construct the feature vector with earlier multiple features to describe the scattering characteristics and local spatial structure of each pixel, as follows: Assuming that a local image region is constructed by w × w pixels around a studied pixel p ∈ I, which is denoted by N (p) and built from a sliding spatial window. The MFRCM is defined as follows: where f q ∈ R 36 is the multifeature vector at pixel q.µ = (1/w 2 ) q∈N (p) f q represents the estimated mean value of the multifeature vector with local region N (p). Obviously, the RCM is a local polarimetric-geometric descriptor including multiple features. In addition, it should be noted that the used feature descriptor is not limited to only contain the above-mentioned four different features and the other useful features also can be incorporated into this descriptor.

D. Proposed Hybrid Weighted Local K-Means Clustering Algorithm Based on GD of MFRCMs
The weighted K-means clustering is an effective unsupervised algorithm, which has a relatively small computational complexity and has been widely applied in PolSAR image classification tasks. Traditionally, the weighted K-means clustering algorithms mainly focus on the weight with pixel feature vector representation, which cannot capture the local spatial information. Although the RCM-based methods partly overcome this shortcoming by using the RCM representation, there are still two important issues. On the one hand, the traditional methods usually utilize norm distance or Wishart distances to describe the similarity of MFRCMs, which cannot tolerate the scaling and rotation of MFRCMs. On the other hand, they only calculate the weights with single scale (region) samples and completely ignore the diversity among pixels with the assumption that all pixels from each local region belong to the same class, which will lose a great deal of structure information, such as edges. To address these issues, a GD-HWLKM PolSAR image clustering algorithm is proposed in this article. First, GD is exploited to describe the similarity of MFRCMs because they lie on the Riemannian space, which can effectively tolerate the scaling and rotation of MFRCMs. Furthermore, it is clearly that the homogeneous points should be the more believable than the inhomogeneous points that are generally located in the edge areas or heterogeneous areas. As shown in Fig. 4, three different types of weights from different scales are taken into consideration to measure the local distribution characteristics, including: 1) local class distribution-based weight (large scale), 2) region-based weight (medium scale), and 3) pixel-based weight (small scale). The constructed hybrid weight can preserve the local structure information as much as possible by the restriction in different scales. The details of the proposed GD-HWLKM algorithm are presented as follows.
Let c l be the lth (l = 1, 2, . . . , L) cluster and L be the total class number. The MFRCM of point p is R p , then the objective function F KM of the proposed weighted K-means is defined as follows: where R l represents the lth cluster centroid, which can enhance the discrimination by using RCM descriptor and GD metric. c l is the lth class dataset. D GD (R p , R l ) is the GD between the pth point and the lth cluster centroid, which is defined as follows [38], [39]: (14) where (·) T and T r(·) donate the transpose and trace operators, respectively. W (p) is the assigned hybrid weight for the pth point, as shown in Fig. 4, which is composed of three types of weights from different scales. The construction details of them are presented as follows.

1) Local Class Distribution-Based Weight:
In the homogeneous areas, the pixel classes are usually alike. On the contrary, the pixel classes should be diverse in the heterogeneous areas. Therefore, in each iteration, the local class (or clutter type) diversity from the temporary classification can represent the local spatial information, which can be regarded as the largescale information. In this article, we use the class distribution probability to represent this local spatial information. Assume that the class of current pixel is l p ∈ 1, 2, . . . , L, the local class distribution-based weight is defined as follows: where Num(class = l) represents the total number of pixelbased points belonging to the lth clutter type. Obviously, in this large scale, the local class distribution is more concentrated on the class of current pixel, the weighted value will be bigger and the current pixel is more believable to update the cluster centroids.

2) Region-Based Weight:
The region-based weight is constructed by the spatial adjacency regions from the local image patches. It can be regarded as a medium-scale weight. The pth region is represented by the MFRCM R p . Let R N = {R 1 , R 2 , . . . , R q , . . . , R Q } denote the spatial adjacency regions. And, the region-based weight W region (p) is defined as follows: where R q is MFRCM of the qth adjacency region and Q is the total number of the spatial adjacency regions. It should be noted that R q and R p are adjoining but they do not have the same pixel, as shown in Fig. 4. For the same classes, the local regions are more similar, the region-based weight is bigger. And this local region will be more probably regarded as a homogeneous region. Conversely, the region-based weight will be close to 0. It should be noted that we define the first term or the second term is equal to zero in the above-mentioned formula when N R l q =l p = 0 or N R l q =l p = 0, respectively.

3) Pixel-Based Weight:
For each MFRCM, all pixels belonging to a local patch are regarded as the same class, which ignore the diversity among pixels and might smooth the image edges. To address this issue, we construct a pixel-based weight to measure this diversity characteristic for each patch, which can be regarded as the least scale weight. From Section III.B, a local region N (p) contains w × w pixels around a studied pixel p ∈ I. The pixel-based weight is constructed as follows: where f q is the multifeature vector at pixel q. d pixel (p, q) is a similarity metric between two vectors, which is defined with the cosine value of vectors angle, as follows: where 2 is norm operation for vectors, is the operation for calculating the inner product between vectors. Obviously, the defined similarity metric also can tolerate the scaling and rotation of vectors, which is more reasonable and effective for measuring the similarity of different land covers owning various polarimetric information. Similarly, the first term or the second term is equal to zero in the above-mentioned formula when N R l q =l p = 0 or N R l q =l p = 0, respectively. In summary, the local hybrid weight can be calculated with the above-mentioned three type weights, as follows: where it should be note that the above-mentioned weights are all performed by normalization operation in advance.

Algorithm 1: Proposed SAR-GMTI Approach.
Inputs: Multichannel and multipolarization image data. The maximum number T, total class number L of covers and clustering end marker ε. 1: Online knowledge acquisition.
1.1: Calculate the coherence matrixes by (6) and (7). 1.2: Extract multiple features listed in Table I  We first adopt the polarization-space classification technique [47] to obtain the initial cluster centroids. By alternately computing cluster centroid and performing the clustering steps, the above-mentioned function can be minimized with the iterative processing. In each iterative procedure, the weight value of each point will be updated with new cluster result. The proposed SAR-GMTI approach is summarized in Algorithm 1.

IV. EXPERIMENTAL RESULT AND ANALYSIS
In this section, the experiments on real measured data are performed to evaluate the proposed method. Section IV.A gives a brief description of the experimental data and Section IV.B is the experimental results, which is composed of the weighed estimation result of coherency matrix, the classification result of PolSAR image, and the GMTI processing result.

A. Data Description
The experimental multichannel and multipolarization data used in this article are taken by the quad-pol radar system with three azimuth receiving channels which operates at X-band (10  GHz), platform height is 8000 m, antenna azimuth aperture is 1.056 m, the platform velocity is 150 m/s, and PRF is 800 Hz. The tested scene is located in Weinan, Shaanxi Province, China, in April 2017. In Fig. 5, we can see that this experimental scene is situated in the northeast of Weinan city and adjacent to the Weihe river. The single-look slant-range and the azimuth resolutions are about 0.5 m for the experimental data. For reducing processing time, the azimuth resolution is reduced to 2 m by multilook processing. In the top of image shown in Fig. 5, the grayscale HH-polarization (HH-pol) SAR image overlaps on an optics image from Google Earth. It is obvious that this illuminated scene contains a part of urban region, river, farm land, asphalt roads, etc.

1) Weighed Estimation Result of Coherency Matrix:
An accurate false color image is an intuitionistic expression for different ground covers with polarimetric information, which is consist of multiple scattering components from different scattering  Fig. 6(a) and (b), which are from the traditional estimated coherency matrix (we call it as "the traditional estimated method") and the proposed weighed estimated coherency matrix (we call it as "the proposed estimated method"), respectively. Due to the use of the simple averaging filter in one receiving channel, a large number of vegetation is mistakenly recognized as surface scatterings with the traditional estimated method. However, the proposed estimated method can increase the estimation accuracy of the coherency matrixes by the two steps, including a weighted multilook operator with the multipolarimetric data and the multichannel coherency matrix fusion using correlation coefficient. This method not only can take the similarity of pixels into consideration to avoid the over-smoothing problem, but also can fuse the coherency matrixes of different receiving channels to obtain a robust estimation value. To further quantitatively analyze the performance of the proposed method, the power contributions of different scatterings are shown in Fig. 6(c). It is obvious that the volume scattering component is enhanced from 30% to 51% and the double-bounce scattering component has a few decreases in a certain extent, which are the mistaken scatterings of the traditional estimated method.
2) Classification Result of PolSAR Image: To demonstrate the effectiveness of the proposed multichannel GD-HWLKM PolSAR classification approach, several classification methods within the K-mean cluster algorithm framework are shown in Fig. 7, which are based on different metrics, including Manhattan metric, Euclidean metric, Wishart metric, AIR metric, and GD metric. It should be noted that the different classifiers have the same inputs (coherency matrices, which are obtained  by the proposed weighted estimation method). The ground truth is shown in Fig. 7(a), which is from the manual segmentation by the "Image Labeler" toolbox of MATLAB under the guidance of the Google Earth© optics image. There are five ground covers including buildings, high-density vegetation, low-density vegetation, water, and bare soil. Here, it should be specially pointed out that the roads will be divided into the water or bare soil category because they have similar scattering characteristics. In addition, some land covers located in the image border at the range direction will be also classified as the water, since they have the very low intensities with the influence of antenna pattern. Table II, we can see that the proposed GD-HWLKM method shows superior classification results than other methods. It has the highest overall accuracy (OA) of 89.78% among all the methods. Fig. 7(c)-(h) show that the methods based on the MFRCMs can effectively suppress the speckle noise. Comparing the methods based on the MFRCMs, the method based on the pixel feature vectors has lots of small misclassified fragments on account of ignoring the spatial information, which results in the decrease of accuracy for all covers, as shown in Fig. 7(b). And, the classical methods based on norm-based metric have a relatively weak distinguishing ability for high-, or low-density vegetation, as shown in Fig. 7(c) and (d). The Wishart-based method can separate the different vegetation, but the classification accuracy of buildings is greatly reduced, as shown in Fig. 7(e). For the methods (AIR, GD) based on Riemannian space, as shown in Fig. 7(f) and (g), they can obtain a remarkable and great classification accuracy, since they can effectively tolerate the scaling or rotation between two MFRCMs and are more suitable to measure the similarity of matrixes. But, those methods usually cause smoothness of cover edges, which will lose a lot of edge information. Under this circumstance, it is very difficult that the roads are correctly classified. However, the proposed GD-HWLKM method not only takes the local spatial information into account by using GD to measure the similarity of MFRCMs, but also exploits three different types of weights (region-based weight, pixel-based weight, and local class distribution-based weigh) to describe the diversity among pixels from different scales, which can    preserve the edges as many as possible. Furthermore, with the K-means algorithm framework, the proposed method can obtain unsupervised classification results with higher overall accuracy.

As shown in
3) GMTI Processing Result: Under the guidance of polarimetric classification results (online knowledge), the accurate CCMs can be estimated for different ground covers, which can greatly improve the clutter suppression performance, especially for the strong clutter. Fig. 8 gives the suppression results of different polarimetric channels without classification (by SCM method) or with classification and the corresponding average power of all range bins. Intuitively, it can be found that the residual power of the urban area is less when using the classification information. To further quantitatively analyze the performance of the proposed method, the averaging residual CNR of different areas are calculated as listed in Table III, including extreme strong clutter, buildings, and natural scene such vegetation and bare soil. It is shown that the extreme strong clutter can be better suppression up to 11-17 dB and the averaging residual CNR of buildings is reduced by 4-10 dB. Furthermore, an urban area marked by red rectangle is selected in Fig. 9 to compare clutter suppression performance with different CCM estimation methods for extreme strong clutter, as shown in Fig. 10. Here, SCM and PST methods are performed without any knowledge. As we can see, comparing the SCM method, the clutter suppression performance of the proposed method can be improved about 10-18 dB for different polarimetric channels. In addition, there also is a certain improvement about clutter cancellation ability besides the PST method because the used training samples are from the same land cover.
To further demonstrate the improvement of GMTI performance of the proposed method, we carry out the moving targets detection and parameter estimation for four polarimetric channels with or without classification result assistance, respectively. The GMTI results are shown in Fig. 11(a)-(d) and (f)-(i) with PFA = 10 −5 for different S-pol components, including HH-pol, HV-pol, VH-pol, and VV-pol channel. From Fig. 11(a)-(d), without classification results, it can be found that there are a large number of false alarms in urban area because of the residual power of strong clutter after cancellation, which are marked with yellow circles. Obviously, the proposed method can greatly reduce the false alarms and reserve the truth moving targets as much as possible, as shown in Fig. 11(f)-(i). Fig. 12 gives the number of detected targets (all, false alarm (FA), and true) from different processing methods. It should be noted that there are no cooperative targets in the experimental scene, so we distinguish the truth targets from all detected targets by the estimated velocity of targets and others are recognized as false alarm. In this article, the velocity threshold is set as 1.78 m/s, which is corroding to the 1/2 antenna beam main lobe.
In addition, to further improve the GMTI performance, we take advantage of NCID technique to combine the targetdetection results of four S-pol channels. Although the scattering characteristics and power of targets and clutter are different between different polarimetric channels, they still can be considered as independent observations at the same time and the targets are more believable after NCID processing. In this article, the NCID technique is achieved by M/N rule (M = 2, N = 4) and the final PFA = 6 × 10 −10 . The detection results of 2/4 rule without or with classification information are demonstrated in Fig. 11(d) and (j), respectively. And the number of detected targets is shown in Fig. 12. Evidently, the NCID technique can improve the detection performance, further. Specifically, the most of true moving targets are effectively preserved when remarkably reducing the false alarms in urban area. Additionally, 15 moving targets are selected to compare the improvement of output SCNR. The locations of selected targets are marked in Fig. 9 and their SCNRs are depicted in Fig. 13 for different polarimetric channels with different methods, including SCM, PST, and the proposed method. As it can be seen more clearly, comparing the results with SCM and PST methods, the output SCNRs for most of the targets are improved, because the overestimation problem of CCM almost is almost avoided under the guidance of the previous classification results. In summary, the proposed method can improve the GMTI performance in evident reduction of the false alarms and the improvement of moving-target detection performance.

V. CONCLUSION
An online Pol-KA SAR-GMTI approach has been proposed in this article to improve the GMTI performance by online polarization classification result, which need not any prior knowledge and can suit the complicated geographical environment. For obtaining the online knowledge, the weighed estimation of coherency matrix with the multichannel SAR system, the multiple feature extraction with RCMs, and the unsupervised classification by the proposed GD-HWLKM clustering algorithm is performed in sequence. Typically, the proposed weighed estimation method of coherency matrix makes full use of the spatial information by fusing the coherency matrixes from different azimuth receive channels in multichannel. And the proposed unsupervised GD-HWLKM PolSAR image clustering approach can achieve a better classification performance when preserving the edges. Under the guidance of the online knowledge, the IID clutter training samples are selected to accurately estimate the CCMs of different land covers, which can improve suppression performance about 10-18 dB and detection performance moving targets on the decrease of the false alarms and the increase of output SCNR of moving targets. In addition, we perform the SAR-GMTI processing for HH-, HV-, VH-, and VV-pol components, respectively. How to joint polarimetric and spatial adaptive processing for GMTI with the guidance of online knowledge will be under investigation in our future work to achieve the better improvement performance of clutter suppression and moving-target detection.