2017 Vol. 6, No. 2

Reviews
Automatic Target Recognition (ATR) is one of the most difficult problems in Synthetic Aperture Radar (SAR) data interpretation. In recent years, the model-based SAR target recognition method has attracted much attention because of its good performance in the extended operation condition. Based on the research of a few domestic research institutes, this paper briefly introduces the preliminary research results and gives some thoughts about SAR ATR problem. First of all, the development of parametric scattering model are discussed from three aspects. Next, two ways to model the parametric electromagnetic scattering for complex target are put forward. Finally, we propose a new framework for a Three-Dimensional (3D) parametric scattering model based SAR ATR. In the end, the future research direction of model-based SAR target recognition is prospected. Automatic Target Recognition (ATR) is one of the most difficult problems in Synthetic Aperture Radar (SAR) data interpretation. In recent years, the model-based SAR target recognition method has attracted much attention because of its good performance in the extended operation condition. Based on the research of a few domestic research institutes, this paper briefly introduces the preliminary research results and gives some thoughts about SAR ATR problem. First of all, the development of parametric scattering model are discussed from three aspects. Next, two ways to model the parametric electromagnetic scattering for complex target are put forward. Finally, we propose a new framework for a Three-Dimensional (3D) parametric scattering model based SAR ATR. In the end, the future research direction of model-based SAR target recognition is prospected.
Papers
Deep learning such as deep neural networks has revolutionized the computer vision area. Deep learning-based algorithms have surpassed conventional algorithms in terms of performance by a significant margin. This paper reviews our works in the application of deep convolutional neural networks to target recognition and terrain classification using the SAR image. A convolutional neural network is employed to automatically extract a hierarchic feature representation from the data, based on which the target recognition and terrain classification can be conducted. Experimental results on the MSTAR benchmark dataset reveal that deep convolutional network could achieve a state-of-the-art classification accuracy of 99% for the 10-class task. For a polarimetric SAR image classification, we propose complex-valued convolutional neural networks for complex SAR images. This algorithm achieved a state-of-the-art accuracy of 95% for the 15-class task on the Flevoland benchmark dataset. Deep learning such as deep neural networks has revolutionized the computer vision area. Deep learning-based algorithms have surpassed conventional algorithms in terms of performance by a significant margin. This paper reviews our works in the application of deep convolutional neural networks to target recognition and terrain classification using the SAR image. A convolutional neural network is employed to automatically extract a hierarchic feature representation from the data, based on which the target recognition and terrain classification can be conducted. Experimental results on the MSTAR benchmark dataset reveal that deep convolutional network could achieve a state-of-the-art classification accuracy of 99% for the 10-class task. For a polarimetric SAR image classification, we propose complex-valued convolutional neural networks for complex SAR images. This algorithm achieved a state-of-the-art accuracy of 95% for the 15-class task on the Flevoland benchmark dataset.
Feature extraction is a key step in radar target recognition. The quality of the extracted features determines the performance of target recognition. However, obtaining the deep nature of the data is difficult using the traditional method. The autoencoder can learn features by making use of data and can obtain feature expressions at different levels of data. To eliminate the influence of noise, the method of radar target recognition based on stacked denoising sparse autoencoder is proposed in this paper. This method can extract features directly and efficiently by setting different hidden layers and numbers of iterations. Experimental results show that the proposed method is superior to the K-nearest neighbor method and the traditional stacked autoencoder. Feature extraction is a key step in radar target recognition. The quality of the extracted features determines the performance of target recognition. However, obtaining the deep nature of the data is difficult using the traditional method. The autoencoder can learn features by making use of data and can obtain feature expressions at different levels of data. To eliminate the influence of noise, the method of radar target recognition based on stacked denoising sparse autoencoder is proposed in this paper. This method can extract features directly and efficiently by setting different hidden layers and numbers of iterations. Experimental results show that the proposed method is superior to the K-nearest neighbor method and the traditional stacked autoencoder.
Attributed scattering center is one of important features of Synthetic Aperture Radar (SAR) image. In this paper, a method for the matching of attributed scattering centers and its application to SAR target recognition is proposed. First, the attributed scattering centers of the test SAR image and template SAR images are extracted on the basis of the attributed scattering model. Second, the Hungarian algorithm is employed to match the two scattering center sets. Based on the one to one correspondence, we design a new similarity measure to evaluate the similarity between the two scattering center sets that will decide the target type of the test image. The similarity measure considers the effects of each individual scattering center, single matching pair, and missing alarms and false alarms; thus, it is more comprehensive. The experiment based on moving and stationary target acquisition and recognition database demonstrates the validity of the proposed method. Attributed scattering center is one of important features of Synthetic Aperture Radar (SAR) image. In this paper, a method for the matching of attributed scattering centers and its application to SAR target recognition is proposed. First, the attributed scattering centers of the test SAR image and template SAR images are extracted on the basis of the attributed scattering model. Second, the Hungarian algorithm is employed to match the two scattering center sets. Based on the one to one correspondence, we design a new similarity measure to evaluate the similarity between the two scattering center sets that will decide the target type of the test image. The similarity measure considers the effects of each individual scattering center, single matching pair, and missing alarms and false alarms; thus, it is more comprehensive. The experiment based on moving and stationary target acquisition and recognition database demonstrates the validity of the proposed method.
A feature fusion algorithm based on a Stacked AutoEncoder (SAE) for Synthetic Aperture Rader (SAR) imagery is proposed in this paper. Firstly, 25 baseline features and Three-Patch Local Binary Patterns (TPLBP) features are extracted. Then, the features are combined in series and fed into the SAE network, which is trained by a greedy layer-wise method. Finally, the softmax classifier is employed to fine tune the SAE network for better fusion performance. Additionally, the Gabor texture features of SAR images are extracted, and the fusion experiments between different features are carried out. The results show that the baseline features and TPLBP features have low redundancy and high complementarity, which makes the fused feature more discriminative. Compared with the SAR target recognition algorithm based on SAE or CNN (Convolutional Neural Network), the proposed method simplifies the network structure and increases the recognition accuracy and efficiency. 10-classes SAR targets based on an MSTAR dataset got a classification accuracy up to 95.88%, which verifies the effectiveness of the presented algorithm. A feature fusion algorithm based on a Stacked AutoEncoder (SAE) for Synthetic Aperture Rader (SAR) imagery is proposed in this paper. Firstly, 25 baseline features and Three-Patch Local Binary Patterns (TPLBP) features are extracted. Then, the features are combined in series and fed into the SAE network, which is trained by a greedy layer-wise method. Finally, the softmax classifier is employed to fine tune the SAE network for better fusion performance. Additionally, the Gabor texture features of SAR images are extracted, and the fusion experiments between different features are carried out. The results show that the baseline features and TPLBP features have low redundancy and high complementarity, which makes the fused feature more discriminative. Compared with the SAR target recognition algorithm based on SAE or CNN (Convolutional Neural Network), the proposed method simplifies the network structure and increases the recognition accuracy and efficiency. 10-classes SAR targets based on an MSTAR dataset got a classification accuracy up to 95.88%, which verifies the effectiveness of the presented algorithm.
A detection method for Synthetic Aperture Radar (SAR) targets based on single sample feature extraction is proposed. Similar targets in a SAR image are detected according to the effective features of the selected single target sample. First, the potential targets of interest in a SAR image are detected, and the area features and texture features are extracted from the target sample and potential targets, respectively. Then, the false targets are eliminated from the potential targets via different matching methods. The proposed method for texture description in this paper can be adopted for targets with different attitudes by extracting the rotationinvariance features of the local region; these features can deal with speckle noise and deformation. The experimental results show the feasibility and validity of the proposed method. A detection method for Synthetic Aperture Radar (SAR) targets based on single sample feature extraction is proposed. Similar targets in a SAR image are detected according to the effective features of the selected single target sample. First, the potential targets of interest in a SAR image are detected, and the area features and texture features are extracted from the target sample and potential targets, respectively. Then, the false targets are eliminated from the potential targets via different matching methods. The proposed method for texture description in this paper can be adopted for targets with different attitudes by extracting the rotationinvariance features of the local region; these features can deal with speckle noise and deformation. The experimental results show the feasibility and validity of the proposed method.
The Coherent Change Detection (CCD) measures the phase difference in repeat passes in SAR images and is a powerful technique for detecting minute changes between two synthetic aperture radar images taken at different times. Nevertheless, the CCD has two problems. These are the high false-alarm rates and threshold selection. To deal with these problems using the likelihood change, this study makes two improvements. First, the model parameters are optimized by the maximum likelihood method and more accurate and robust parameters are obtained by using the sliding window in the neighborhood operations. Second, the automatic change in the threshold method is proposed based on the histogram characteristics of different data. The processing of real data suggests that the proposed method is effective in detecting minute changes. The Coherent Change Detection (CCD) measures the phase difference in repeat passes in SAR images and is a powerful technique for detecting minute changes between two synthetic aperture radar images taken at different times. Nevertheless, the CCD has two problems. These are the high false-alarm rates and threshold selection. To deal with these problems using the likelihood change, this study makes two improvements. First, the model parameters are optimized by the maximum likelihood method and more accurate and robust parameters are obtained by using the sliding window in the neighborhood operations. Second, the automatic change in the threshold method is proposed based on the histogram characteristics of different data. The processing of real data suggests that the proposed method is effective in detecting minute changes.
In the field of image processing using Synthetic Aperture Radar (SAR), aircraft detection is a challenging task. Conventional approaches always extract targets from the background of an image using image segmentation methods. Nevertheless, these methods mainly focus on pixel contrast and neglect the integrity of the target, which leads to locating the object inaccurately. In this study, we build a novel SAR aircraft detection framework. Compared to traditional methods, an improved saliency-based method is proposed to locate candidates coarsely and quickly in large scenes. This proposed method is verified to be more efficient compared with the sliding window method. Next, we design a convolutional neural network fitting in SAR images to accurately identify the candidates and obtain the final detection result. Moreover, to overcome the problem of limited available SAR data, we propose four data augmentation methods comprising translation, speckle noising, contrast enhancement, and small-angle rotation. Experimental results show that our framework achieves excellent performance on the high-resolution TerraSAR-X dataset. In the field of image processing using Synthetic Aperture Radar (SAR), aircraft detection is a challenging task. Conventional approaches always extract targets from the background of an image using image segmentation methods. Nevertheless, these methods mainly focus on pixel contrast and neglect the integrity of the target, which leads to locating the object inaccurately. In this study, we build a novel SAR aircraft detection framework. Compared to traditional methods, an improved saliency-based method is proposed to locate candidates coarsely and quickly in large scenes. This proposed method is verified to be more efficient compared with the sliding window method. Next, we design a convolutional neural network fitting in SAR images to accurately identify the candidates and obtain the final detection result. Moreover, to overcome the problem of limited available SAR data, we propose four data augmentation methods comprising translation, speckle noising, contrast enhancement, and small-angle rotation. Experimental results show that our framework achieves excellent performance on the high-resolution TerraSAR-X dataset.
An improved Hybrid Change Detection (HCD) method is proposed for multi-temporal Synthetic Aperture Radar (SAR) images. Firstly, a Pixel-Based Change Detection (PBCD) method is used to extract the initial change area, and the initial cluster center is estimated based on its results. Then, Fuzzy Clustering Method (FCM) is used to get three clusters, which including water, background, and the intermediate area. The Nearest Neighbor Clustering (NNC) is adopted as the second-level clustering to divide the pixels belonging to the intermediate area into water and background respectively, afterwards merge all pixels belonging to water. Finally, the difference map of flood region in the time series images is calculated to get the final change detection result. The algorithm is validated by the Sentinel-1A data obtained from Huaihe River and Poyang Lake. The results show that our proposed method can achieve better correctness and has lower total error compared to other methods. An improved Hybrid Change Detection (HCD) method is proposed for multi-temporal Synthetic Aperture Radar (SAR) images. Firstly, a Pixel-Based Change Detection (PBCD) method is used to extract the initial change area, and the initial cluster center is estimated based on its results. Then, Fuzzy Clustering Method (FCM) is used to get three clusters, which including water, background, and the intermediate area. The Nearest Neighbor Clustering (NNC) is adopted as the second-level clustering to divide the pixels belonging to the intermediate area into water and background respectively, afterwards merge all pixels belonging to water. Finally, the difference map of flood region in the time series images is calculated to get the final change detection result. The algorithm is validated by the Sentinel-1A data obtained from Huaihe River and Poyang Lake. The results show that our proposed method can achieve better correctness and has lower total error compared to other methods.
Complex motion can cause serious defocusing phenomena in high resolution spaceborne SAR cases, which then lead to decreased image resolution. In this study, we built a simulation model to quantitatively analyze the signature and effect on maritime fluctuating targets in high resolution cases. To deal with formed Single-Look Complex (SLC) SAR images containing fluctuating targets, we implement a motion-compensation and fine-focusing method to obtain refocused images and the fluctuation parameters. We demonstrate the effectiveness and correctness of the proposed approach in focusing and estimating the parameters of fluctuating targets by processing the simulation results and archived images acquired by Terra-SAR in hybrid spotlight mode. Complex motion can cause serious defocusing phenomena in high resolution spaceborne SAR cases, which then lead to decreased image resolution. In this study, we built a simulation model to quantitatively analyze the signature and effect on maritime fluctuating targets in high resolution cases. To deal with formed Single-Look Complex (SLC) SAR images containing fluctuating targets, we implement a motion-compensation and fine-focusing method to obtain refocused images and the fluctuation parameters. We demonstrate the effectiveness and correctness of the proposed approach in focusing and estimating the parameters of fluctuating targets by processing the simulation results and archived images acquired by Terra-SAR in hybrid spotlight mode.
Crosstalk is not only one of the main error sources in the polarimetric SAR system, but is also an indicator for evaluating calibration performance. In this paper, to determine the impact of crosstalk on land cover classification, we first retrieve the mathematical relation expressions between crosstalk and the Cloude-decomposition-based scattering characteristic. Then, we verify our theoretical conclusions in a semi-physical simulation based on Radarsat-2 polarimetric data for different land covers. Finally, we perform H/a/Wishart classification on the experimental data. From the ratio curve of pixels labeled differently under changing crosstalk, we can determine the crosstalk requirement that will meet the needs of specific applications. Crosstalk is not only one of the main error sources in the polarimetric SAR system, but is also an indicator for evaluating calibration performance. In this paper, to determine the impact of crosstalk on land cover classification, we first retrieve the mathematical relation expressions between crosstalk and the Cloude-decomposition-based scattering characteristic. Then, we verify our theoretical conclusions in a semi-physical simulation based on Radarsat-2 polarimetric data for different land covers. Finally, we perform H/a/Wishart classification on the experimental data. From the ratio curve of pixels labeled differently under changing crosstalk, we can determine the crosstalk requirement that will meet the needs of specific applications.