Convolutional neural network (CNN)-based research has been successfully applied in remote sensing image classification due to its powerful feature representation ability. However, these high-capacity networks bring heavy inference costs and are easily overparameterized, especially for the deep CNNs pretrained on natural datasets. Network pruning is regarded as a prevalent approach compressing networks, but most existing ignores model interpretability while formulating criterion. To address...

Ship detection in synthetic aperture radar (SAR) images has important application value. Sea clutter, complex scenes, a large size change ships, and the arbitrary directionality of ships make ship challenging. With development deep learning, many learning algorithms have been applied to SAR images. These need lot labeled data for training. It is time-consuming label data, unlabeled are easy obtain. necessary use effectively improve performance algorithm. In this study, semisupervised...

Semantic segmentation, a fundamental research direction in synthetic aperture radar (SAR) image interpretation, has significant application value for multiple sectors. However, noise, multi-style terrains, geometric distortion, and shadows make SAR segmentation challenging. Although existing deep learning algorithms tend to mine the semantic relationship between pixels within individual images, they disregard global context of training data different regions from images. Moreover, noise...

In the annotation of remote sensing images (RSIs), effectiveness common object detection methods trained on only a few samples decreases instantly, which has prompted increasing research few-shot problem in sensing. RSIs often exhibit suboptimal performance scenarios due to intricate nature scene information interference and high degree cosine similarity, both present significant challenges their effectiveness. this paper, two-stage framework based fine-tuning is selected deal with problems...

Localization regression in oriented object detection tasks has long faced boundary discontinuity and angular problems induced by periodic angles. These were successfully resolved using a 2d Gaussian distribution to modelling the bounding box (OBB). However, information of square-like objects will be lost when they are converted distribution, forming systematic problem. Its fundamental reason is that aspect ratio tends 1, equiprobability curve degenerates from an ellipse circle, thus losing...

The overwhelming majority of models for remote sensing image (RSI) scene classification generally require the weights pre-trained on natural images initialization before formal training. However, differences in imaging mechanisms lead to huge discrepancies between and RSIs, strong visual representation learned from massive limits performance when inferencing RSIs. To address this issue, well-established self-supervised contrastive learning paradigm field is introduced RSI field. We propose a...

Labeling data in the field of remote sensing is time-consuming and labor-intensive, making domain adaptation between different domains an urgently needed solution. To address gap diverse datasets domain, numerous methods tailored for high-resolution imagery have emerged. Some existing focus on reducing at either feature level or pixel level, often overlooking their underlying connection. tackle this issue, we introduce a prototype-wise contrastive alignment paradigm (PCFA) aimed bridging...

For the past few years, barrier of explainability accompanying by deep neural networks (DNNs) has been increasingly studied. The methods based on class activation map (CAM) which interpret model decision mapping output back to input space, have achieved a notable momentum among research. However, CAM-based cannot stably produce effective explanation results remote sensing images (RSIs), owing coarse location generated high-level features, whereas, RSIs contain abundant detailed spatial...

During the past decades, convolutional neural network (CNN)-based models have achieved notable success in remote sensing image classification due to their powerful feature representation ability. However, lack of explainability during decision-making process is a common criticism these high-capacity networks. Local explanation methods that provide visual saliency maps attracted increasing attention as means surmount barrier explainability. vast majority research conducted on last layer,...

This paper proposes a new semi-supervised PolSAR image classification method using deep belief network (DBN) and tensor dimensionality reduction, which uses multilinear principle component analysis (MPCA) to reduce the dimension of form data, regards multiple features data as input DBN. In order take full advantage neighborhood information each pixel we its form. For simple feature has been proven not be able effectively classify complex terrains. Therefore, combine obtain more abundant...

The gap between self-supervised visual representation learning and supervised is gradually closing. Self-supervised does not rely on a large amount of labeled data reduces the loss human information. Compared with natural images, remote sensing images require rich samples annotation by experts. Moreover, many algorithms have poor interpretability unconvincing results. Therefore, this paper proposes method based prototype assignment designing pretext task so that network maps features to...

The unique visual properties and the huge size of synthetic aperture radar (SAR) images pose challenges in labeling data. scarcity labeled data limits training SAR image segmentation networks. To address this issue, semi-supervised methods are used to train network. However, traditional algorithms like Mean Teacher often generate erroneous pseudo-labels that carry over subsequent epochs, leading overfitting affecting network performance. This stems from an overreliance on unreliable pixels...

To overcome the inherent domain gap between remote sensing (RS) images and natural images, some self-supervised representation learning methods have made promising progress. However, they overlooked diverse angles present in RS objects. This paper proposes Masked Angle-Aware Autoencoder (MA3E) to perceive learn during pre-training. We design a \textit{scaling center crop} operation create rotated crop with random orientation on each original image, introducing explicit angle variation. MA3E...

Target decomposition features are the cornerstone of subsequent analyses for PolSAR images. Generally, adopting single or several algorithms limits representation ability original terrain characteristics. Using all existing features, however, will definitely increase computational complexity. Besides, some even have a negative effect on following tasks. To address these problems, sparse variational autoencoder feature selection framework (SVAE-FS) is proposed in this article. In detail,...