Synthetic aperture radar (SAR) ship detection especially for small ships has issues, such as dense distribution of ships, interference from land and small islands. To address these issues, many deep learning methods, including anchor-based and anchor-free methods, have been successfully migrated from optical scenes to SAR images. However, when the preset scale of anchors does not match well with the ships, it will seriously reduce the detection precision. Due to the lack of anchor-based refinement process, anchor-free methods may generate missing or false alarms in complex scenarios. In this article, a two-stage ship detection network which can generate anchors is proposed. First, our method generates high-quality anchors by network, which is more beneficial for the network to capture small ships. In addition, the generated anchors are centrally set in the region of ships, which reduces the number of anchors unrelated to ships. Second, the receptive field enhancement module is inserted into the feature pyramid network. It sets different dilation ratios of atrous convolution according to the scale of the feature map, which further enriches the semantic information of the elements in the feature map. Therefore, the network can use the information of a wider region effectively to detect ships. Finally, to verify the effectiveness of our method, extensive experiments are carried out on SAR ship detection dataset and high-resolution SAR images dataset. The results show that our method has more strong ability of detecting small ships, and achieves better detection performance than some state-of-the-art methods.

The advancement of convolutional neural networks (CNNs) has greatly promoted the development of ship detection in SAR images. The oriented ship detection is practical and imperative, but the development of its fully-supervised methods is limited by the insufficiency of data annotated in oriented bounding box (OBB). Therefore, we proposed a weakly supervised learning (WSL) paradigm to train network by only data annotated in horizontal bounding box (HBB) for oriented ship detection. The paradigm follows a coarse-to-fine prediction route, and the ships coarse orientations and high-quality pseudo-labels both mined from images are used to train the network combined with the HBB annotations. The designed orientation initialization decoder (OID) utilizes the output from the prediction branch of ship coarse orientation (PBSCO) and the regression branch to decode coarse OBBs, based on the horizontal proposals, and then these OBBs are refined by the fine prediction head, which is jointly trained by high-quality pseudo-labels and HBB annotation. Extensive experiments were conducted on the SAR ship detection dataset (SSDD) and the high resolution SAR images dataset (HRSID) with both OBB and HBB annotations, and the results show that our proposed WSL paradigm can train network to attain the AP ₅₀ and R in the same level of mainstream fully-supervised methods in remote sensing.

Due to a lack of labeled samples, deep learning methods generally tend to have poor classification performance in practical applications. Few-shot learning (FSL), as an emerging learning paradigm, has been widely utilized in hyperspectral image (HSI) classification with limited labeled samples. However, the existing FSL methods generally ignore the domain shift problem in cross-domain scenes and rarely explore the associations between samples in the source and target domain. To tackle the above issues, a graph-based domain adaptation FSL (GDAFSL) method is proposed for HSI classification with limited training samples, which utilizes the graph method to guide the domain adaptation learning process in a uniformed framework. First, a novel deep residual hybrid attention network (DRHAN) is designed to extract discriminative embedded features efficiently for few-shot HSI classification. Then, a graph-based domain adaptation network (GDAN), which combines graph construction with domain adversarial strategy, is proposed to fully explore the domain correlation between source and target embedded features. By utilizing the fully explored domain correlations to guide the domain adaptation process, a domain invariant feature metric space is learned for few-shot HSI classification. Comprehensive experimental results conducted on three public HSI datasets demonstrate that GDAFSL is superior to the state-of-the-art with a small sample size.

The synthetic aperture radar (SAR) oriented ship detection is a more complex and finer task than the horizontal, and it provides ship orientation for military and civilian applications, such as trajectory prediction. However, many oriented target detection methods discarded high-resolution (low-level) feature map, which provides morphological features for predicting accurate bounding box for small ships, due to computational cost and lack of semantic information. Additionally, most methods follow the samples assignment in horizontal detection after padding oriented bounding box. The inaccurate geometric relationship causes that some low-quality positive samples without contributing accuracy improvement are introduced. To improve the precision of small oriented ship detection, we proposed the feature pyramid network contained high-resolution feature (FPNHRF) to fuse high-level semantics and highlight contours and textures of ships in the low-level feature map. To select high-quality positive rotated samples, dynamic rotated positive samples mining (DRPSM) is designed to mine hard positive samples for oriented ship detection specially. In one word, a two-stage and anchor-free detector trained by DRPSM and utilized high-resolution feature map is proposed for detecting oriented ships in SAR imagery. Subsequently, extensive experiments have been conducted on two oriented SAR datasets: the SSDD and the HRSID. As a result, our method attains Recall of 96.2% and 90.1%, and AP50 of 90.7% and 86.5% on SSDD and HRSID respectively. It shows that our detector achieves state-of-the-art results applied in capturing ships, when compared with other popular remote sensing oriented detectors.

With the rapid development of deep learning, the convolutional neural networks (CNNs) have been widely used in hyperspectral image classification (HSIC) and achieved excellent performance. However, CNNs reuse the same kernel weights over different locations, thus resulting in the insufficient capability of capturing diversity spatial interactions. Moreover, CNNs usually require a large amount of training samples to optimize the learnable parameters. When training samples are limited, the classification performance of CNN tends to drop off a cliff. To tackle the aforementioned issues, a novel residual spatial attention kernel generation network (RSAKGN) is proposed for HSIC. First, a spatial attention kernel generation module (SAKGM) is built to extract discriminative semantic features, which can dynamically calculate the attention weights to generate specific spatial attention kernels over different locations. Then, we combine the SAKGM with residual learning framework by embedding the SAKGM into a bottleneck residual block to obtain the residual spatial attention block (RSAB). The RSAKGN is constructed by stacking several RSABs. Experimental results on three public HSI datasets demonstrate that the proposed RSAKGN method outperforms several state-of-the-arts with small sample size.

In hyperspectral (HS) and light detection and ranging (LiDAR) collaborative classification, HS provides rich spectral information, while LiDAR offers unique elevation data. However, existing methods often focus on feature extraction within individual modalities before fusion, which may bring about insufficient fusion due to a lack of intermodal complementarity and interaction. To address this, we propose a framework for HS and LiDAR fusion classification based on the VMamba model, called SSFN, which includes a dual supplement network (DSN) and a VMamba-based integration network (VMIN), modeling long-range dependencies and fully leveraging the correlation and complementarity of heterogeneous information. The DSN, comprising a spatial supplement network (Spa-SN) and a spectral supplement network (Spe-SN), is devised to supplement missing features for each modality. The Spa-SN complements the spatial features of HS by capturing spatial correlations between LiDAR and HS, and the Spe-SN employs spectral information from HS to compensate for the spectral features missing in LiDAR. Thus, both HS and LiDAR have a comprehensive spatial–spectral description. The VMIN is then utilized for augmentation and interaction of supplemented features, and discriminative features are adaptively selected for classification. Extensive experiments on three benchmark datasets demonstrate that our method outperforms multiple state-of-the-art methods and needs the fewest parameters.