Synthetic aperture radar (SAR) images have been extensively used in earthquake monitoring, resource survey, agricultural forecasting, etc. However, it is a challenge to interpret SAR images with severe speckle noise and geometric deformation due to the nature of radar imaging. The translation of SAR-to-optical images provides new support for the interpretation of SAR images. Most of the existing translation networks, which are based on generative adversarial networks (GANs), are vulnerable to part information loss during the feature reasoning stage, making the outline of the translated images blurred and semantic information missing. Aiming to solve these problems, cross-fusion reasoning and wavelet decomposition GAN (CFRWD-GAN) is proposed to preserve structural details and enhance high-frequency band information. Specifically, the cross-fusion reasoning (CFR) structure is proposed to preserve high-resolution, detailed features and low-resolution semantic features in the whole process of feature reasoning. Moreover, the discrete wavelet decomposition (WD) method is adopted to handle the speckle noise in SAR images and achieve the translation of high-frequency components. Finally, the WD branch is integrated with the CFR branch through an adaptive parameter learning method to translate SAR images to optical ones. Extensive experiments conducted on two publicly available datasets, QXS-SAROPT and SEN1-2, demonstrate a better translation performance of the proposed CFRWD-GAN compared to five other state-of-the-art models.

Pointly Supervised Object Detection (PSOD) has attracted considerable interests due to its lower labeling cost as compared to box-level supervised object detection. However, the complex scenes, densely packed and dynamic-scale objects in Remote Sensing (RS) images hinder the development of PSOD methods in RS field. In this paper, we make the first attempt to achieve RS object detection with single point supervision, and propose a PSOD method tailored for RS images. Specifically, we design a point label upgrader (PLUG) to generate pseudo box labels from single point labels, and then use the pseudo boxes to supervise the optimization of existing detectors. Moreover, to handle the challenge of the densely packed objects in RS images, we propose a sparse feature guided semantic prediction module which can generate high-quality semantic maps by fully exploiting informative cues from sparse objects. Extensive ablation studies on the DOTA dataset have validated the effectiveness of our method. Our method can achieve significantly better performance as compared to state-of-the-art image-level and point-level supervised detection methods, and reduce the performance gap between PSOD and box-level supervised object detection. Code is available at https://github.com/heshitian/PLUG.

To enhance mid–low-resolution ship detection, existing methods generally use image super-resolution (SR) as a preprocessing step and feed the super-resolved images to the detectors. However, these methods only use high-resolution (HR) images as ground-truth labels to supervise the training of their SR module but overlook the rich HR information in the detection stage. Inspired by the recent advances in knowledge distillation, in this letter, we design a feature distillation framework to fully exploit the information in ground-truth HR images to handle mid–low-resolution ship detection. Our framework consists of a student network and a teacher network. The student network first super-resolves input images using an SR module and then feeds the super-resolved images to the detection module. The teacher network whose architecture is the same as the student detection module directly takes HR images as input to generate HR feature representation and then distills these HR features to the student network through a distillation loss. Using our feature distillation framework, HR images are not only used as ground-truth labels to train the SR module but also provide “ground-truth” features to train the detection module, which enhances the detection performance of the student network. We apply our framework to several popular detectors, including FCOS , Faster-RCNN , Mask-RCNN , and Cascase-RCNN , and conduct extensive ablation studies to validate its effectiveness and generality. Experimental results on the HRSC2016, DOTA, and NWPU VHR-10 datasets demonstrate that, when applying our framework to Faster-RCNN , our method can outperform several state-of-the-art detection methods in terms of mAP50 and mAP75.

Semantic segmentation of high-resolution remote sensing images (HRSI) is significant, yet challenging. Recently, several research works have utilized the self-attention operation to capture global dependencies. HRSI have complex scenes and rich details, and the implementation of self-attention on a whole image will introduce redundant information and interfere with semantic segmentation. The detail recovery of HRSI is another challenging aspect of semantic segmentation. Several networks use up-sampling, skip-connections, parallel structure, and enhanced edge features to obtain more precise results. However, the above methods ignore the misalignment of features with different resolutions, which affects the accuracy of the segmentation results. To resolve these problems, this paper proposes a semantic segmentation network based on sparse self-attention and feature alignment (SAANet). Specifically, the sparse position self-attention module (SPAM) divides, rearranges, and resorts the feature maps in the position dimension and performs position attention operations (PAM) in rearranged and restored sub-regions, respectively. Meanwhile, the proposed sparse channel self-attention module (SCAM) groups, rearranges, and resorts the feature maps in the channel dimension and performs channel attention operations (CAM) in the rearranged and restored sub-channels, respectively. SPAM and SCAM effectively model long-range context information and interdependencies between channels, while reducing the introduction of redundant information. Finally, the feature alignment module (FAM) utilizes convolutions to obtain a learnable offset map and aligns feature maps with different resolutions, helping to recover details and refine feature representations. Extensive experiments conducted on the ISPRS Vaihingen, Potsdam, and LoveDA datasets demonstrate that the proposed method precedes general semantic segmentation- and self-attention-based networks.

A significant challenge in detecting objects in complex remote sensing (RS) datasets is from small objects. Existing detection methods achieve much lower accuracy on small objects than medium and large ones. These methods suffer from limited feature information, susceptibility to complex background interferences, and insufficient contextual information. To address these issues, a small object detection method with the enhanced receptive field, ERF-RTMDet, is proposed to achieve a more robust detection capability on small objects in RS images. Specifically, three modules are employed to enhance the receptive field of small objects’ features. First, the Dilated Spatial Pyramid Pooling Fast Module is proposed to gather more contextual information on small objects and suppress the interference of background information. Second, the Content-Aware Reassembly of Features Module is employed for more efficient feature fusion instead of the nearest-neighbor upsampling operator. Finally, the Hybrid Dilated Attention Module is proposed to expand the receptive field of object features after the feature fusion network. Extensive experiments are conducted on the MAR20 and NWPU VHR-10 datasets. The experimental results show that our ERF-RTMDet attains higher detection precision on small objects while maintaining or slightly enhancing the detection precision on mid-scale and large-scale objects.

Medium-low resolution (M-LR) remote sensing ship detection is a challenging problem due to the small target sizes and insufficient appearance information. Although image super resolution (SR) has become a popular solution in recent years, the ability of image SR is limited since much information is lost in input images. Inspired by the powerful information embedding ability of the encoder in image rescaling, in this paper, we introduce image rescaling to guide the training of image SR. Specifically, we add an adaption module before the SR network, and use the pre-trained rescaling network to guide the optimization of the adaption module. In this way, more information is embedded in the adapted M-LR images, and the subsequent SR module can utilize more information to achieve better performance. Extensive experimental results demonstrate the effectiveness of our method on image SR. More importantly, our method can be used as a pre-processing approach to improve the detection performance.