Compared with optical images, polarimetric synthetic aperture radar (PolSAR) images usually maintain lower resolution. Ship targets in PolSAR images have fewer pixels than those in optical images. Therefore, architectures such as Faster R-CNN and its variations that focus more on target pixels may fail for PolSAR images. Due to limited data and the simple geometric structures of small ship targets, the R-CNN framework with a large neural network as its backbone easily overfits the training set. In this article, we propose a novel lightweight patch-to-pixel convolutional neural network (P2P-CNN) for ship detection via PolSAR images. P2P-CNN focuses on both the target and its surroundings. A patch of proper size that contains both a target and its surroundings is used as the input of the neural network to determine whether the pixel in the center of the patch belongs to the target. To utilize contextual semantic information at all scales, all feature maps from the top down are combined to improve the final result. Instead of conventional convolution, dilated convolutions are used in the proposed neural network to exponentially expand receptive fields without adding any model parameters. The proposed approach and comparative methods are tested and compared on PolSAR images in various environments under varying image resolution, target size, sea condition, sensor type, etc. The experimental results demonstrate that the proposed approach outperforms all the compared methods.

Ocean surveillance is one of the important applications in synthetic aperture radar (SAR) imagery. Polarimetric SAR provides multichannel information and shows great potential for ocean target observation. Oil-spills and ships possess different polarimetric features from the ocean surface, whose physical backscattering property is generally admitted as being dominated by the Bragg resonant scattering. In this study, we focus on the oil-spill and ship observation based on the polarimetric features. Oil-spills present a non-Bragg scattering property. While backscatter from ships is predominated by the double-bounce scattering and the multiple reflections between the ship and sea surface. Ships exhibit very different scattering characteristics from the ocean Bragg scattering. Based on the extended Bragg (X-Bragg) model, a new method is proposed for observing oil-spills and ships. This method allows distinguishing oil-spills from two kinds of important look-alikes, i.e., biogenic slicks and the low-wind region (LWR), and also shows good performance for ship enhancement. Experiments are performed on the C-band fully polarimetric SAR data acquired by both SIR-C/X-SAR and RADARSAT-2. The other issue concerned is the application potential of the hybrid dual-polarimetric (i.e., compact polarimetry (CP)) SAR mode, which has the advantage of providing larger imaging coverage compared to the full polarimetry. Based on the X-Bragg model, where the backscatter reflection symmetry is assumed, an equivalent method is proposed with the circularly polarized transmission CP mode. Experimental results show that similar results can be obtained between the CP and full polarimetry for ocean target observation.

SAR image targets detection and recognition (SAR-TDR) have become research hotpots in the remote sensing application. These targets include ships, vehicles, aircrafts, oil tanks, bridges, and so on. However, with the rapid development of SAR technology and increasingly complex electromagnetic environment, complex characteristics of SAR images bring severe challenges to the accurate SAR-TDR via traditional physical models or manually features extraction based machine learning methods. In recent years, deep learning (DL), as a powerful automatic features extraction algorithm, has been widely used in the computer vision domain. More specifically, DL has also been introduced into the SAR-TDR tasks and effectively achieved good performances in terms of accuracy, real-time, etc. With the rapid development of DL, SAR images processing, and practical requirements of SAR-TDR in civilian and military domains, it is very necessary to make a systematic survey on the SAR-TDR in past few years. In this survey paper, we mainly conduct a systematic overview of DL-based SAR-TDR literatures on two tasks, i.e., targets recognition (e.g., ground vehicles, ships, and aircrafts) and targets detection (e.g., ships, aircrafts,, change detection, seasurface oil spill, and oil tanks), respectively. More specifically, our related works about these topics are also presented to verify the effectiveness of DL-based methods. Firstly, several DL methods (e.g., CNN, RNN, AEs, and GNN), commonly used to SAR-TDR, are briefly introduced. The systematic review of DL-based SAR-TDR (including our related works) is then presented. Finally, the current challenges and future possible research directions are deeply analyzed and discussed.

Deep learning methods have been widely studied for Polarimetric synthetic aperture radar (PolSAR) land cover classification. The scarcity of PolSAR labeled samples and the small receptive field of the model limit the performance of deep learning methods for land cover classification. In this paper, a vision Transformer (ViT)-based classification method is proposed. The ViT structure can extract features from the global range of images based on a self-attention block. The powerful feature representation capability of the model is equivalent to a flexible receptive field, which is suitable for PolSAR image classification at different resolutions. In addition, because of the lack of labeled data, the Mask Autoencoder method is used to pre-train the proposed model with unlabeled data. Experiments are carried out on the Flevoland dataset acquired by NASA/JPL AIRSAR and the Hainan dataset acquired by the Aerial Remote Sensing System of the Chinese Academy of Sciences. The experimental results on both datasets demonstrate the superiority of the proposed method.

This letter proposes a novel convolutional neural network (CNN) method for dual-polarized synthetic aperture radar (SAR) ship grained classification. The network employs hybrid channel feature loss that jointly utilizes the information contained in the polarized channels (VV and VH). It is demonstrated that, by adopting the proposed CNN framework and the novel loss function, the classification performance can be efficiently improved. First, instead of the prevalently used threefold or fourfold division (container ship, oil tanker, bulk carrier, and so on), the proposed method can further divide vessels into eight accurate categories. Second, this method can not only effectively classify targets into eight categories but also its accuracy in terms of fewer category classifications surpasses existing methods. Third, the method can achieve good performance on a small training data set. Experiments conducted on the OpenSARShip data sets indicate that the proposed classification method achieves state-of-the-art results.

Synthetic aperture radar (SAR) plays an essential role in earth observation and projection due to its capability to penetrate clouds, which makes it possible to monitor terrestrial surfaces under all weather conditions. Multiplicative noise often occurs in the SAR signal, hampering the retrieval of information from SAR imagery. Convolutional neural networks (CNNs) have been used in many computer vision tasks and are helpful in image denoising. However, current CNN-based denoising approaches inevitably lead to a “washed out” effect that loses spatial details. Another limitation is that most typical CNN-based denoising models require a noise-free image for training. To address these issues, we propose a novel end-to-end self-supervised SAR denoising model: Enhanced Noise2Noise (EN2N), which can be trained without a noise-free image. To enhance the quality of the result images, the perceptual features from a pre-learned CNN are introduced to restore the spatial details by a hybrid loss function. Experiments show that our proposed method outperforms the typical denoising methods in terms of noise reduction and feature preservation based on image quality metrics. Also, the new hybrid loss could enhance the spatial details significantly. The good performance maintains the robustness throughout time, which reduces the uncertainty in time-series SAR caused by random noise. Benefiting from optimization of graphics processing unit (GPU) and multi-threading, the proposed method has higher computation efficiency than traditional methods. This study demonstrates the great potential of using our self-supervised deep learning approaches for SAR image denoising in the future.