Review of Ship Detection in Polarimetric Synthetic Aperture Imagery （in English）

1. Introduction

Synthetic Aperture Radar (SAR) is an active microwave remote sensing device. SAR uses virtual array and pulse compression technologies to obtain two-dimensional high-resolution images of ground objects. Polarization is an important attribute of electromagnetic waves, and the polarization characteristics of a target are one of its important inherent attributes, which can be used in target detection and recognition. Polarimetric SAR (PolSAR) is generated by inputting polarimetric information into the SAR system. The PolSAR system, which has retained complete target electromagnetic scattering characteristics, has become an important tool of modern radar imaging technology, and it has been widely used in military reconnaissance terrain mapping, environmental and natural disaster monitoring, ship detection, and other fields^[1]. In the past 30 years, many scholars have performed considerable work in the field of polarimetric radar remote sensing, which has laid a solid foundation for the application of polarimetric radar remote sensing and made significant achievements. Notably, Boerner and his international collaborators have made remarkable contributions. He is not only an active promoter of the polarimetric radar theory and its application but also an active disseminator of the polarimetric radar theory. The first practical orthogonal PolSAR system was Airborne SAR (AIRSAR) developed by the National Aeronautics and Space Administration Jet Propulsion Laboratory (NASA JPL)^[2] in 1985 based on time-sharing polarimetric measurement technology. The successful acquisition of data has stimulated the analysis and application of PolSAR data, and since then, the PolSAR system has entered a period of rapid development. Thus far, the typical PolSAR systems at home and abroad mainly include^[3]: (1) Airborne SAR: USA’s AIRSAR system, Australia’s AuSAR system, Denmark’s EMISAR system, New Zealand’s PHARUS system, Japan’s PISAR system, Canada’s CV-580 system, Germany’s DOSAR MEMPHIS ESAR system, and France’s STORM system and (2) Spaceborne SAR: Japan’s ALOS/PALSAR system, Germany’s Terrasar-X system, and Canada’s RADARSAT-1/2 system. On August 10, 2016, China’s first civilian C-band PolSAR satellite, Gaofen-3 (GF-3), was successfully launched from the Taiyuan Satellite Launch Center, reaching the advanced international level. GF-3 has been widely used in marine disaster reduction, water conservation, meteorology, and other fields.

With the rapid development of the PolSAR system, PolSAR image interpretation technology has been the focus of considerable attention. Target detection methods based on PolSAR images have become a frontier topic in the field of SAR information processing. Many research institutions at home and abroad have conducted fruitful work in the field of PolSAR target detection. Lincoln Laboratory, USA; Canada Centre for Remote Sensing, Canada; University of Niigata, Japan; German Aerospace Center (DLR), Germany; Defence Science and Technology Organisation (DSTO), Australia; Stirling University, UK; Trumsø University, Norway; Chinese Academy of Sciences, China; Tsinghua University, China; National University of Defense Technology, China; University of Electronic Science and Technology, China; Xi’an University of Electronic Science and Technology, China; Naval University of Engineering, China; and other research institutions have made outstanding contributions^[3]. As the representative research institutions, NASA JPL and Lincoln Laboratory have conducted pioneering works on PolSAR image target detection, classification, and recognition and obtained a large number of experimental and theoretical results. Research results show that polarimetric information will further improve the radar target detection, classification, and recognition performance^[4]. Novak of Lincoln Laboratory has made a fundamental contribution to the field of target polarimetric characteristic detection. The most common method of ship target detection in PolSAR image is Constant False Alarm Rate (CFAR) because the ship target has a stronger scattering echo than the sea clutter. In the case that the prior information of the target is unknown, the CFAR detection method can achieve a better performance than other detection methods. However, with the diversity of radar platform parameters and the complexity of sea surface environments, addressing the problems associated with sea clutter modeling and parameter estimation, slow and small target detection, dense target detection, and other issues remain a challenge in PolSAR ship detection.

By summarizing the well-known PolSAR ship detection algorithms, this paper provides the development history and future trend of PolSAR ship detection and some valuable suggestions for subsequent research. The framework of this overview is shown in Fig. 1. At present, PolSAR ship detection can be divided into four categories, i.e., (1) target polarimetric characteristic detection method, (2) slow-moving target detection method, (3) ship wake detection method, and (4) target detection algorithm based on deep learning. The target polarimetric characteristic detection method is mainly applied to ship detection when there is a certain difference between polarimetric characteristics of ship and clutter. The slow-moving target detection method is mainly applied to the situation in which the polarimetric characteristic difference between the target and clutter is small, but there is a certain velocity difference between the target and clutter. Meanwhile, the target motion information can be extracted to obtain the real-time sea surface situation. The ship wake detection method is not used to detect the ship target itself but to detect the wake generated by ship motion, which is mainly used for small and stealth target detection. The data-driven deep learning method is an important tool of Artificial Intelligence (AI) and does not require manual target feature extraction. This method has achieved great success in the intelligent interpretation of PolSAR images. CFAR detection is essentially statistical signal processing, which can be regarded as the follow-up process of the aforementioned methods. Thus, a series relationship is observed between the CFAR detection method and the aforementioned methods. LIU et al.^[5] have done considerable research on this, and only a representative introduction is listed in this paper.

Figure  1.  Classification of PolSAR ship detection methods

DownLoad: Full-Size Img PowerPoint

2. Target Polarimetric Characteristic Detection Method

The most common method of PolSAR image detection is based on the statistical characteristics of SAR pixels^[6]. To utilize the statistical similarity and spatial correlation of pixels, the pixel block detection method was developed^[7]. PolSAR ship characteristic detection methods can be divided into two categories, i.e., pixel level and pixel block^[7]. Pixel level detection methods mainly include simple polarimetric channel synthesis technology, polarimetric optimization detection technology, and detection technology based on the scattering mechanism^[8]. Pixel block detection methods, also known as spatial neighborhood detection methods, mainly include superpixel and neighborhood matrix detection algorithms.

2.1 Simple polarimetric channel synthesis technology

As the number of polarimetric channels increased, simple polarimetric channel synthesis techniques have emerged^[9]. Simple polarimetric channel synthesis has two schemes. The first scheme is to detect and process each polarimetric channel separately and to fuse the final detection results, which is called decision-level fusion, using a Single Channel Detector (SCD)^[9]. The second scheme, which is based on pixels, is to synthesize the data of different polarimetric channels and perform CFAR detection on the synthesized images in a single-channel manner, which is called signal-level fusion and can be regarded as the detection quantity obtained by the linear weighting of the scattering vector^[10] using a Span Detector (SD) and Optimal Span Detector (OSD)^[10]. A large number of experimental results show that the detection performance of polarimetric channel synthesis is better than that of single-polarization SAR^[10]. This method is relatively simple; however, it is no longer used, except for simple scenarios and situations with low performance requirements.

2.2 Polarimetric optimization detection technology

Polarimetric optimization refers to the information processing technology of extracting object characteristics from PolSAR images by comprehensively utilizing the information of different polarimetric channels according to a certain optimization objective function. Polarimetric optimization is an upgrade of simple polarimetric channel synthesis. In 1988, Boerner^[11] proposed the Polarimetric Matched Filter (PMF) based on the maximum Signal-to-Noise Ratio (SNR) criterion. The PMF is an optimal linear filter, which can suppress clutter and improve the signal-to-clutter ratio (SCR) by weighting the polarimetric scattering vector^[11]. In 1989, Novak et al.^[12] proposed an Optimal Polarimetric Detector (OPD) based on the principle of the likelihood ratio test. The proposed method reaches the theoretical upper limit of the polarization detector (with the prior information of the target required). The object covariance matrix in the OPD is replaced by the product of the unit matrix and the polarimetric channel power, and the identity likelihood ratio test detector can be obtained^[13]. In 1990, Novak and Burl^[14] proposed a Polarimetric Whitening Filter (PWF) to compensate for the defect that the OPD requires prior information by taking the clutter fluctuation as the optimization objective. YANG et al.^[15] proposed the Optimization of Polarimetric Contrast Enhancement (OPCE) based on the maximum SNR criterion by optimizing the polarimetric mode of transmitting and receiving and achieved a detection effect that is close to the matched filter. YANG et al.^[16] proposed a method based on the Minimal Clutter-to-Signal Ratio (MCSR) to enhance the polarimetric contrast and concluded that the PMF was a special case of the MCSR. The MCSR method has a good target detection performance in a specific environment; however, it cannot identify the optimal detection dimension of the subspace. Novak^[17] noted that detection performance depends not only on SNR but also on clutter fluctuation. The performance comparison of polarimetric optimization detection technology is shown in Fig. 2. The experimental results show that the OPD generally has the best performance, and the OPCE is generally less effective than the PWF^[18].

Figure  2.  Performance comparison of common polarimetric optimization detection methods^[18]

DownLoad: Full-Size Img PowerPoint

To extend the work of Novak et al. to the multi-view situation, Lopes and LIU^[19,20] extended the PWF to the multi-look PWF. Based on PWF statistical modeling and parameter estimation, LIU et al.^[21] conducted in-depth research on the CFAR characteristics of PolSAR ship detection. According to the sea clutter under different conditions, the CFAR analytical expressions of different statistical distributions are deduced, and the ideal CFAR results are obtained. To some extent, the problems of CFAR statistical modeling and parameter estimation under the influence of varying radar platform parameters and complex sea surface environments are solved. Fig. 3 shows the PWF-based CFAR detection results with clutter background characterized by the IBeta distribution product model. Notably, when the statistical model is consistent with the measured data, CFAR has the best performance, whereas other statistical models (e.g., K-Wishart, G0-Wishart, and lognormal models) have many false alarms. Inspired by the aforementioned works, GAO et al.^[22–25] also investigated the statistical modeling and parameter estimation of single PolSAR images.

Figure  3.  PWF processing results with different textural distributions^[21]

DownLoad: Full-Size Img PowerPoint

For a dense target environment, TAO^[27] proposed an adaptive truncation method to estimate the statistical model parameters of SAR images and achieved ideal detection results. LIU^[28] extended the method to PolSAR images, deduced the probability density function of the adaptive truncated polarimetric covariance matrix, and achieved ideal results in the dense target environment. The PWF detection results on the GF-3 image are shown in Fig. 4, where Fig. 4(a)–Fig. 4(e) represent the different parameter estimation methods and (f) represent the Pauli image of the RADARSAT-2 data, with the yellow rectangle denoting the false alarm and the red rectangle denoting the missing alarm. Notably, the effect of Fig. 4(a) on the accuracy of parameter estimation after truncation is significantly improved, which also indicates the necessity of accurate parameter estimation for clutter in dense areas.

Figure  4.  Dense ship detection in PolSAR images based on the adaptive truncation method^[28]

DownLoad: Full-Size Img PowerPoint

At present, the polarimetric optimization method is gradually combined with the interpretable machine learning method, and target detection is conducted by adopting classification optimization criteria^[29], which is a new method to improve the detection performance of small targets and can be used as the development direction of subsequent polarization optimization detection technology.

2.3 Detection technology based on the scattering mechanism

Given the different scattering mechanisms between ship target and ocean surface, the difference between polarimetric characteristics can distinguish the ship from the clutter. Thus, the detection method based on the polarimetric scattering characteristics has strong physical interpretability. Because sea clutter has reflection symmetry, Nunziata et al.^[30] proposed a Reflection Symmetry (RS) detector, which uses the correlation between co-polarimetric and cross-polarimetric channels to detect targets. WANG et al.^[31] conducted a further theoretical analysis of the RS detector and deduced the CFAR analytical expression and parameter estimation method of the detector. From the perspective of comparing the polarimetric scattering energy of clutter and ship target, Marino et al.^[32,33] deduced the Polarimetric Notch Filter (PNF) based on geometric disturbance and obtained the corresponding analytical expression of statistical distribution^[33]. Gao et al.^[34–36] applied the PNF to high sea state conditions and mixed polarimetric and interferometric conditions and achieved ideal results. In 2020, LIU et al.^[8] modified the PNF expression, called it the new PNF (NPNF), considering the real physical significance of energy in electromagnetic scattering, and achieved a better detection performance than the original PNF. The experimental comparison results are shown in Fig. 5(a). The effect of NPNF is similar to that of PWF. The comparison of the detection effects of PNF and NPNF is shown in Fig. 5(b).

Figure  5.  Performance comparison of PNF and NPNF^[8]

DownLoad: Full-Size Img PowerPoint

Another detection method based on the scattering mechanism is to use the polarimetric decomposition theory. Polarimetric decomposition is an effective tool to analyze the scattering mechanism of targets. Polarimetric decomposition can decompose the scattering or covariance matrix according to different physical scattering types and can be effectively applied to ship target detection in PolSAR images^[37]. In 1999, Cameron decomposition was first applied to SIR-C data from spaceborne imaging radar for wide-area surface ship target detection by Ringrose and Harris^[38]. In 2000, based on the Cloude-Pottier decomposition, Touzi et al.^[39] introduced the detection amount of polarimetric entropy derived from the polarimetric covariance matrix to detect the target through the eigenvalue of polarimetric entropy and scattering angle, which effectively improved the difference between the target and the background and achieved a better ship detection effect^[40]. In 2009, CHEN et al.^[41] proposed the concept of polarimetric cross-entropy and proved the effectiveness of feature detection using AIRSAR data. In 2013, Sugimoto et al.^[42] applied the decomposition of the four-component model proposed by Yamaguchi to ship target detection using ALOS PALSAR measured data. The advantage of this algorithm lies in the simplicity of interpretation and execution. YANG et al.^[43,44] added the scattering mechanism to OPCE and proposed a Generalized OPCE (GOPCE) method. Their results show that the addition of scattering similarity significantly improves the target detection capability.

In 2015, Touzi et al.^[45] used the optimal characteristics of depolarization to detect ships and proved that the algorithm could effectively detect ship targets using the RADARSAT-2 measured data. In 2018, GAO et al.^[36] combined polarimetric entropy and scattering energy to detect ships in PolSAR images and achieved good results, as shown in Fig. 6. This method reflects the advantages of the fusion of energy and polarimetric domain features and is more conducive to target detection. In 2019, Bordbari et al.^[46] divided the scattering mechanism into two categories, i.e., target scattering mechanism for the target and nontarget scattering mechanism for the clutter, and adopted the subspace projection method to suppress clutter and significantly improve the target detection performance. At present, a single polarimetric decomposition model cannot easily and effectively express the difference between target and clutter. Thus, new polarimetric decomposition models are investigated, and different component coefficients of different polarimetric decomposition models are reconstructed into new feature vectors. Moreover, the Fisher classification criterion or other feature vector space can improve the detection performance^[47], which is also the trend of ship target detection based on the polarimetric decomposition theory.

Figure  6.  Comparison of detection results among different methods for small targets^[36]

DownLoad: Full-Size Img PowerPoint

2.4 Spatial neighborhood detection method

Traditional PolSAR imagery processing is mostly based on a single pixel but ignores the relationship between neighborhood pixels. Therefore, its performance will be reduced when the clutter scattering characteristics are similar to the ships’ characteristics, such as ship detection in ports. To utilize the correlation between pixel and its neighborhood, the pixel block detection method has been developed. Recent studies show that the fusion of spatial and polarimetric information can considerably enhance the target detection capability, and the target detection effect can be further enhanced by extracting the spatial information^[6,48]. At present, detection based on superpixel and neighborhood matrices is mainly used^[7,48].

The essence of a superpixel is a set of points with similar scattering characteristics; thus, the definition of similarity is its core. WANG et al.^[49] proposed three superpixel-level similarity definitions to enhance the SCR and built an automatic detection algorithm based on this. Finally, using the RADARSAT-2 data, even at a low Signal Clutter Ratio (SCR), the method can still achieve a low Probability of False Alarm (PFA). Another study^[50] proposed a similar superpixel segmentation method based on the polarimetric scattering similarity of ship targets and sea clutter. CUI et al.^[51] determined the Similar Pixel Number (SPN) of PolSAR images by conducting the similarity test, and a target detection method based on this feature is proposed. The RADARSAT-2 and GF-3 measured data show that the introduction of airspace and polarimetric information can considerably enhance the detection capability of ship targets in PolSAR images. As shown in Fig. 7, the green rectangle denotes correct detection, the yellow rectangle denotes false alarm, and the red rectangle denotes missing alarm. Fig. 7(a)–Fig. 7(d) represent the different detection methods. The results show that the SPN method has obvious advantages.

Figure  7.  Comparison results among different ship detectors with SPN^[51]

DownLoad: Full-Size Img PowerPoint

Huang et al.^[52] proposed a PolSAR image significance detection model based on the distance between blocks. The polarimetric coherence matrix was used to measure the similarity between polarimetric pixel blocks at both local and global scales. Jäger et al.^[53] defined PolSAR image entropy according to polarimetric and spatial information and proposed a method to detect PolSAR image salient regions. Wang et al.^[54] proposed an image salient region detection method based on the similarity test of the pixel block polarimetric covariance matrix. LIN et al.^[55] first proposed the method of ship detection in single PolSAR images using the superpixel-level Fisher vector. Then, in their doctoral dissertation, they compared PolSAR images with RGB images and extended the edge sensing superpixel segmentation method to make the superpixel ship detection method applicable to PolSAR images.

To construct the neighborhood matrix, ZHANG et al.^[56,57] drew on the concept of optical local binary coding and proposed a polarimetric covariance matrix and a complete polarimetric covariance difference matrix using the amplitude, phase, and spatial information to distinguish ship target pixel and sea clutter pixel. However, the final matrix form lacks some good properties, such as it is not a positive semidefinite matrix and it is difficult to interpret as a practical target. Ships and clutter are both spatially correlated, and ships will change the spatial correlation of surrounding clutter. Sidelobes also affect the correlation of neighborhood pixels. Therefore, LIU et al.^[7] proposed the Neighborhood Polarimetric Covariance Matrix (NPCM) and a detection method based on the NPCM. Compared with the results of existing methods, their results show that, at present, the new method has outstanding performance and is nearly the best polarimetric detector, as shown in Fig. 8.

Figure  8.  Performance comparision among different ship detection methods with NPCM^[6]

DownLoad: Full-Size Img PowerPoint

The NPCM combines polarimetric and spatial characteristics and significantly increases the dissociability between ships and sea clutter. However, the best dimension of the neighborhood spatial correlation matrix is difficult to determine and the total dimension is high. In the future, the development direction of the neighborhood matrix is to reduce its dimension to avoid dimensional disaster while maintaining the detection performance. Simultaneously, the neighborhood matrix information can be combined with the superpixel segmentation, which is expected to improve the spatial detection performance.

3. Slow-moving Target Detection Method

With the continuous deterioration of the sea surface environments, the detection of the slow, small ship submerged in the sea clutter becomes more difficult. Ground Moving Target Indication (GMTI) has become one of the important functions that the PolSAR system needs to complete to more comprehensively grasp the situation of the naval battlefield and obtain the target motion information. Moving target detection algorithms are divided into two categories, i.e., range-Doppler detection algorithms and image detection algorithms^[58]. Compared with that in the range-Doppler domain, the operation in the SAR image domain has the advantages of low design requirements, small computation, and easy implementation. Thus, this paper focuses on detection in the image domain. GAO et al.^[59–61] conducted in-depth research on slow-moving target detection on the image domain in a single polarimetric situation. Here we mainly focus on the utilization of polarimetric information. Thus, the PolSAR-GMTI system is divided into single-channel PolSAR-GMTI and multichannel PolSAR-GMTI, where a single channel refers to a single radar and a multichannel refers to two or more radars.

3.1 Single-channel PolSAR-GMTI

In 1995, Park et al.^[62] conducted in-depth research on the polarimetric Space-Time Adaptive Processing (STAP) algorithm and proposed the joint processing method of space-time-polarization adaptive signal. Their research results show that the addition of polarimetric information can improve the moving target detection performance. In 2003, Park et al.^[63] further analyzed the potential of polarimetric information to improve the performance of the STAP algorithm. Wang et al.^[64] proposed a moving target detection algorithm based on polarimetric-interferometric optimization technology. On this basis, Friedlander et al.^[65] used the VSAR velocity estimation principle to estimate the moving target velocity based on the results after polarimetric-interferometric optimization. LIU et al.^[66] from NUDT analyzed the coherence of frequency domain residual interference in PolSAR images. For moving targets located on random rough surfaces, such as highways, Yeremy et al. in Canada, funded by the Ministry of National Defence of Canada, conducted a full PolSAR-GMTI test in Petawawa on June 8, 2002, and in Ottawa on September 24, 2002, using the CV-580 system and obtained a large amount of data. Using these data, Matter et al.^[67] published a special AD report in 2005 and investigated the feasibility of using PolSAR data for moving target detection. Two PolSAR moving target detection algorithms are mentioned in the report. One is based on single PolSAR data, and the other is the processing method of Along-track Interferometry (ATI) based on two cross-polarimetric channels. The cross-polarimetric ATI method uses the reciprocity of cross-polarimetric images of stationary targets through ATI processing of SAR images of two cross-polarimetric channels^[67] to achieve moving target detection. Of course, the cross-polarimetric ATI method can also use two cross-polarimetric channel SAR images for the Displaced Phase Center Antenna (DPCA) method to achieve moving object detection^[67]. ZOU et al.^[68] proposed an algorithm combining the PWF and time–frequency analysis method. This method enhanced the SNR of the moving target by utilizing polarimetric information, enhanced the visibility of the target image, and completed the detection of the linear contour of the moving target image by Radon transform.

LIU et al.^[69,70] analyzed the influence of motion on PolSAR images. Their results show that the phase difference between HV and VH images is 0 for the scatterer with reciprocity, whereas the phase difference between HV and VH sidelobes is $\pi$ . The motion will enhance the azimuth blur of the PolSAR image, and the detection effect of different polarimetric channels is shown in Fig. 9. However, this method also adds the blurring of stationary objects, which is also known as a ghost. How to use the cross-polarimetric information to achieve the suppression of stationary targets and the detection of slow-moving targets simultaneously is a problem to be solved in this method.

Figure  9.  Moving target detection results in different polarimetric channels in PolSAR images^[69]

DownLoad: Full-Size Img PowerPoint

3.2 Multichannel PolSAR-GMTI

Multichannel SAR increases the spatial information and can use both spatial and temporal information simultaneously, which can suppress the stationary clutter, detect the moving target more effectively, and solve the inherent blurring problem. Polarimetric information helps improve the interferometric quality, estimation accuracy of the interferometric phase^[58], and detection performance of slow-moving targets^[71]. To overcome the low performance limit of the Minimum Detectable Velocity (MDV) for single-platform PolSAR-GMTI, which is limited by size, LI et al.^[72] proposed the use of the virtual baseline in the mode of alternating transmitting and receiving by the fully polarimetric radar and combined three new virtual baselines to obtain a better MDV. Based on considerable research on single-polarimetric multichannel SAR-GMTI, Defence Research and Development Canada designed a dual-channel polarimetric beam mode for RS-2, which combines PolSAR imaging with GMTI functions using dual-channel and time-sharing polarimetric techniques. Chiu et al.^[73] proved that this method helps detect ship targets in the Arctic glacier region. Given that the RADARSAT Constellation Mission (RCM), which was launched in 2019, does not have a dedicated GMTI function, Chiu et al.^[74] proposed the moving target detection method of double PolSAR images based on cross-polarimetric information to generate shadows, thereby providing an auxiliary GMTI function.

Australian scholars Stacy and Preiss^[71] extended ATI amplitude-phase detection to the GLRT method similar to the single-channel situation but did not conduct in-depth formula derivation and analysis. To address the difficulties of GMTI in the current interferometric SAR (InSAR), Zhang Peng et al.^[75,76] used Polarimetric ATI SAR (PolATInSAR) to improve the detection performance of slow-moving targets. Their results show that the new method can significantly improve the detection probability and radial velocity estimation accuracy of the interference system for slow and small ship targets and can easily achieve CFAR detection. The simulation results are shown in Fig. 10. With the successful launch of the RCM and the subsequent development of China’s GF series constellation architecture, the need for multi-station polarimetric-interferometric processing has become more obvious.

Figure  10.  Simulation results of polarimetric along-track interferometry moving target detection^[76]

DownLoad: Full-Size Img PowerPoint

The current PolATInSAR studies lack the acquisition of measured data. To facilitate research on PolATInSAR and put it into practice as soon as possible, the subaperture decomposition method can be adopted to obtain the virtual interferometric effect. In 1985, Japanese scientist Ouchi^[77] first analyzed the characteristics of moving targets in multi-look sub-images in the time domain and indicated that there is a certain time delay between synthetic aperture centers of SAR subaperture images and the two-dimensional velocity of moving targets will produce imaging differences between multi-view image sequences. Based on Ouchi’s research, in 1996, German scientist Martin Kirscht^[78] first applied the multi-view processing method to SAR-GMTI and proposed the single-look image sequence method, which is also known as the subaperture decomposition method. This method achieves the goal of canceling clutter and retaining the moving target by subtracting multiple subaperture image sequences. Scholars from the Chinese Academy of Sciences and Beijing Aerospace University^[79–82] have made many improvements to the subaperture decomposition method and combined it with the multichannel SAR-GMTI method to achieve good results. In 2015, Marino et al.^[83] derived the target detection algorithm after subaperture decomposition and summarized the results of detection performance comparison. In the future, subaperture decomposition will be extended to the PolSAR direction by the dimension extension method to obtain a better PolSAR-GMTI effect.

4. Ship Wake Detection Method

Wake detection, which has an extensive development space, can acquire the course and speed of ship targets and other information and provide a new idea for the detection of PolSAR small targets and stealth targets. Wake detection can be summed up by the difference in energy between the wake and the background. Wake detection is mainly based on the linear feature extraction of the wake edge and the difference between the wake region and the background scattering characteristics. CHONG et al.^[84] considered the following issues: (1) wake is a linear feature body with a certain width; (2) the wake may be lighter or darker than the sea background; (3) wakes are not necessarily straight, but sometimes curved; (4) SAR images have inherent speckle noise; and (5) other non-wake linear structures can be detected in the image.

In 1992, Schuler et al.^[85] investigated the polarimetric characteristics of one arm of the P-band Kelvin wake. In 1999, Pottier et al.^[86] proposed the classification method of ship wakes using the feature vectors in the polarimetric eigenvalue decomposition. This method uses a neural network to classify the simulated fully polarimetric Kelvin wake data, and the classified samples are composed of the rotation invariants after polarimetric decomposition. However, the classification results do not specify the physical meaning of the wakes represented by these features, nor do they verify the measured data^[87]. In 1999, Hennings et al.^[88] analyzed the Kelvin arm model in SAR images and concluded that the visibility of the Kelvin arm on radar images largely depends on the radar polarization mode rather than the radar frequency. In HH polarization, the visibility of the Kelvin arm on radar images is higher than that of VV polarization. When the incident angle is small, the Kelvin arm has high visibility on radar images^[89]. In 2001, Imbo et al.^[87,90] investigated wake features in PolSAR images and conducted a Radon transform of the polarimetric covariance matrix to detect ship wakes. Two methods of wake type detection are discussed in this paper. One is to use the optimal channel combination after PWF processing for wake detection, and the other is to perform Radon transform in the six channels of the polarimetric covariance matrix, use the Mahalanobis distance to judge the peak value, and use the inverse transform method to detect the wake. The first method uses the polarimetric information incompletely, and the second method uses the polarimetric information of all channels, but the operation is extensive. The final test results are shown in Fig. 11, with the green line denoting the detected wakes.

Figure  11.  Wake detection results by PWF and Radon transform^[90]

DownLoad: Full-Size Img PowerPoint

In 2002, Kasilingam et al.^[91] investigated the ocean polarimetric correlation coefficient, mentioned that turbulent wakes could be observed in the image with the polarimetric correlation coefficient, and indicated that the change of surface roughness at a small scale would not modify the correlation coefficient with polarization; however, they did not state whether this parameter was beneficial to the detection of the Kelvin wake. In 2002, Morris et al.^[92] carefully investigated the spatial distribution of polarimetric characteristics inside and around the sea surface of a boat’s wake and determined that, because of the nonlinear effects, the polarimetric characteristics were distributed nonuniformly in the spatial domain. When moving from the wake region to the surrounding sea area, the corresponding eigenvector distribution produces a strong transition, which helps identify the wake and clutter regions. In 2003, Morris et al.^[93] proposed an entropy-based wake echo analysis method, which solved the problem of ship wake detection and analysis under low incident angle by analyzing the eigenvalue of the covariance matrix. Using the high-resolution radar system of the Defense Science and Technology Organization, the researchers obtained polarimetric measurements of wakes generated by an experiment boat operating in Port Phillip and decomposed the polarimetric covariance matrix data. The polarimetric entropy, scattering α, and anisotropy pattern figure are obtained. Then, the proper threshold is used to identify the wake and clutter regions. In 2004, Yang et al.^[94] used the fully polarimetric data obtained from AIRSAR and SIR-C/X-SAR to analyze the radar features of ocean features (such as wave, oil film, ship wake, oceanfront, and seabed terrain), and based on the aforementioned analysis, determined the optimal polarization mode for observing these features. They indicated that, in most cases, HH and RL polarization is better than VV polarization. In 2006, Arnold-Bos et al.^[95] proposed a multipurpose dual-base polarimetric radar simulator for sea use, including the generation of wakes. In 2011, Wu et al.^[87] extracted and investigated the polarimetric characteristics of Kelvin waves based on the difference in the scattering mechanism between Kelvin wave and sea clutter in SAR images. They proposed that a simple co-polarimetric ratio can enhance the Kelvin wave, focused on the application of polarimetric decomposition, and extracted the polarimetric entropy that can more completely represent the characteristics of the Kelvin wave and provide its physical meaning. In 2018, Xu et al.^[96] introduced a new two-step (rough and fine processing) detector to detect weak turbulent wakes in SAR images. Based on the polarimetric decomposition theory, this method proposed a new parameter, i.e., Surface Scattering Randomness (SSR), to enhance the difference between wake and sea surface. In the rough detection process, the potential ship wake area was extracted by the Digital Axoids Transform (DAT) of SSR. In the fine detection process, the regions obtained in the rough detection process are separated in the polarimetric domain and processed separately. Finally, regularized least squares are introduced to obtain the linear parameters of ship wakes, and prior information is obtained from the Radon transform. Experiments are conducted based on the RADARSAT-2 data, and the results show that, compared with the existing detectors based on Radon or Hough transform, the algorithm has a better capability to detect weak turbulence wakes from PolSAR images. In 2018, Biondi^[97] proposed a complete process for the automatic detection of ship target motion parameters through Kelvin wakes. This method achieves clutter suppression, sparse target detection, accurate wake tendency, and Kelvin spectrum analysis by two-stage Low-Rank plus Sparse Decomposition (LRSD) combined with Radon transform. Based on the robust principal component analysis, sparse targets of interest, including ship targets and Kelvin waves, can be inferred from invariant low-rank backgrounds. In 2019, Biondi^[98] considered the polarimetric information of SAR images and achieved the multichannel two-stage polarization LRSD. The upgraded algorithm achieves the extraction of polarimetric features of the sparse targets of interest and makes the detection of ship targets more accurate. The results are shown in Fig. 12. Notably, wake detection by combining sparse representation, polarization characteristics, and machine learning optimization theory will be the future trend.

Figure  12.  Ship wake detection results of low-rank and sparse decomposition combined with polarization^[98]

DownLoad: Full-Size Img PowerPoint

5. Target Detection Algorithm based on Deep Learning

In recent years, deep-learning-based algorithms have achieved great success in the field of optical target detection. The differences between the application of the deep Convolutional Neural Network (CNN) to SAR and optical images are mainly reflected by three points, i.e., the characteristic difference is significant, the data quantity is small, and the observation condition is sensitive, which are more obvious in PolSAR images. Scholars from various countries have extended these optical detection algorithms to the study of ship detection methods in SAR images, demonstrating the superior detection performance of deep learning in complex sea surface environments. The core of deep learning is to extract structure through hierarchical features, i.e., to characterize the original data with deep structures. The essence of deep learning can be regarded as a multilayer nested nonlinear fitting function, which skillfully trains the network to an appropriate fitting accuracy through a random gradient descent algorithm. The current great success of deep learning can be attributed to the emergence of the convolutional layer and the use of the Rectified Linear Unit (ReLU) activation function^[100]. The ReLU is the best activation function because it is similar to neurons, i.e., neurons do not respond to some inputs, the responses of neurons to some inputs are monotonically related, and the number of neurons in the active state at any moment is always sparse. Methods represented by CNN have a flexible structure and can automatically extract structured features, which can extract not only low-dimensional features of images but also high-dimensional features of images to better identify and classify targets^[100]. At present, CNN has become a research hotspot in the field of SAR ship target detection^[101]. GAO et al.^[102] combined binarized normed gradients and a faster region-based CNN (Faster R-CNN) to improve the detection performance. Xiaoling et al. proposed a detection method based on depthwise separable CNN for SAR ship detection with high real-time requirements, which has an obvious effect^[103].

However, the publicly available literature on PolSAR ship detection using CNN is relatively limited^[104–108]. In Ref. [104], the VGG network is adopted for ship detection. In Refs. [105,106], the improved Faster R-CNN is adopted. Similar to previous literature, they also adopted a network with many parameters, which easily overfits the training dataset. Cozzolino et al.^[108] proposed a small network for SAR target detection. This small network can be trained directly on a limited set of SAR and PolSAR data; however, it uses the max-pooling method, which will lead to the loss of the small targets. Of course, many ship detection methods based on CNN in SAR images can be extended to PolSAR images in theory. JIN et al.^[107] proposed a patch-to-pixel CNN (P2P-CNN) for small boat detection in PolSAR images, and the detection results are shown in Fig. 13. P2P-CNN shows a good performance. Fig. 13(a) illustrates the main structure of the network, which is a four-layer dense block structure with crops. The method adopts the multiplication operation combining the characteristics of each block. On the one hand, the combination yields the jump line connection effect, avoids the vanishing gradient problem, and makes the training more smooth. However, combined with multilayer semantics, the method uses high-level semantic features with powerful expression capability and the underlying semantic characteristics with only a slight change in sensitivity to jointly determine whether the center pixel is a ship and significantly reduce false alarms. Given the finiteness of PolSAR labeled data, the network parameters need to be reduced to decrease the risk of overfitting. Therefore, a dense connection module is introduced to reduce the width of the network using feature reuse, and hollow convolution is introduced to increase the sensing field and decrease the depth of the network without changing the parameters. The results show that the detection performance of P2P-CNN is considerably improved. The future trend of deep learning in PolSAR target detection is how to apply the current deep-learning-based single PolSAR image target detection methods to PolSAR images, utilize receptive field and polarimetric features, and build a new network structure.

Figure  13.  P2P-CNN network structure and detection performance^[107]

DownLoad: Full-Size Img PowerPoint

For deep learning recognition of ship wake, large amounts of ship wake of SAR images need to be obtained for training and practice. The measured SAR images of wake are few, which do not apply to the method using deep learning. Therefore, the use of deep learning for SAR images of ship wake recognition research is relatively limited^[109]. Fitch et al.^[110] used the artificial neural network method to help interpret the wake after Radon transform for NASA JPL AIRSAR images. Kang used the combination of CNN and SAR images to detect the azimuth offset between the ship and the wake and investigated the potential of using the offset to automatically detect the speed of ship targets^[111].

6. Summary and Scope

In this paper, the development and existing problems of different technical approaches for PolSAR ship detection are reviewed in detail. Table 1 lists the advantages and disadvantages of each method for the reference of interested researchers.

Table  1.  Applicable scenarios and pros & cons summary of PolSAR ship detection methods 

Detection method				Advantage		Disadvantage
Target polarimetric characteristic detection method (mainly applied to the detection of ship and clutter when the polarimetric characteri stics are different)		Simple polarimetric channel synthesis technology		The early PolSAR target dete ction method is simple and easy to realize, and the effect is better than the single PolSAR image target detection method		To some extent, the amplitude or phase information is used, but the polarimetric information is not fully utilized
		Polarimetric optimization detection technology		1. The information of each polari metric channel is fully utilized 2. PolSAR images are optimized to achieve the best contrast effect		Different optimization criteria should be selected according to the actual situation, and the selection of a sliding window also has a significant influence on the performance
		Detection technology based on the scattering mechanism		It has strong physical interpre tability		Adaptability to target types and sea state needs to be improved
		Spatial neighborhood detection method		The fusion of spatial and polarime tric information considerably enhances the target detection capability		Dimensions need to be reduced to avoid a dimensional disaster
Slow-moving target detection method (mainly applied to the situation where the polarimetric characteristic difference between target and clutter is small but thesre is a certain velocity diffe rence; simultaneously, it can also extract the target motion informa tion to obtain the real-time sea surface situation)		Single-channel PolSAR-GMTI		Cross-polarimetric information is used to enhance the detection effect of slow-moving targets		Suppression of stationary targets and detection of slow-moving targets are difficult to achieve simultaneously
		Virtual interferometric slow-moving target detection		The problem of the lack of actual interference data source is solved		It is difficult to choose the number of subaperture decompo sition and overlap degree
		Multichannel PolSAR-GMTI		Integrating the advantages of multi-platform and polarization information considerably enhan ces the ability to move target detection		Complex structure and cost limit engineering applications
Ship wake detection method (it is used not to detect the ship target body but to detect the wake gene rated by the movement, mainly for small targets and stealth targets)		Linear feature detection of wake edge		The detection problem is abstrac ted into the linear feature detec tion problem, and the method is relatively simple		It is difficult to detect different types of weak wakes in the cluttered background
		Background difference detection in the wake region		Due to the different scattering mechanisms of the wake region and background sea surface, there are considerable differences in spectrum characteristics, statisti cal characteristics, polarimetric characteristics, and other aspects, and the effect is good		The theoretical analysis and method implementation are more complex
Target detection algorithm based on deep learning (it has achieved great success in intelligent PolSAR image interpretation without extracting target features man ually)		Detection method based on CNN		Flexible structure can automa tically extract structured features, including both low-dimensional and high-dimensional features of images		1. It needs a large number of samples to train with 2. Interpretability needs to be investigated

 | Show Table

DownLoad: CSV

Simultaneously, with the continuous development and progress of radar technology, the future development trend of PolSAR ship target detection can be predicted as follows:

(1) At present, the optimal polarimetric technology is gradually combined with the interpretable machine learning method, and the classification optimization criterion is adopted for target detection, which may be a new method to improve the detection performance of small targets, and can be used as the development direction of subsequent polarimetric optimization detection technology. New feature vectors can be constructed from different component coefficients of different polarimetric decomposition models, and the detection performance of targets can be improved by adopting the Fisher classification criterion or other feature vector space, which is also the development trend of ship target detection based on the polarimetric decomposition theory.

(2) In the dense target environment, in addition to the pollution between the target and the clutter, some problems in the ship target group, such as splitting and merging, cause difficult image interpretation. Targeted separation of dense target obfuscation and merger of split targets are also the research direction of ship target detection in the future.

(3) The integrated system of distributed fully polarimetric radar is developed by utilizing interferometric processing technology. Interference and polarization both belong to the category of multidimensional information processing. Through the extension of spatial and polarimetric information, the detection capability of the slow and small ship targets of PolSAR can be significantly improved.

(4) There is still a contradiction between the complexity of the statistical model, the accuracy of parameter estimation, and the number of samples. The statistical model of clutter and the theory of parameter estimation are further improved. With the improvement of PolSAR resolution and the complexity of sea conditions, a uniform and universal statistical modeling framework of clutter polarimetric covariance matrix needs to be established, and its CFAR detection closed form, which will help improve the statistical theory of PolSAR ship detection, needs to be deduced.

(5) In combination with AI technologies, such as deep learning, PolSAR ship detection is driven by a small amount of data, focusing not only on the detection of ship ontology but also on the extraction of ship wake features to further enhance the ship target detection capability of PolSAR image.