With the development of supervised deep neural networks, classification performance on existing remote sensing scene datasets has been markedly improved. However, supervised learning methods rely heavily on large-scale tagged examples to obtain a better prediction performance. The lack of large-scale tagged remote sensing scene images has become the primary bottleneck in scene classification. To deal with this issue, a novel scene classification method using self-supervised gated self-attention generative adversarial networks (GANs) with similarity loss is proposed. Specifically, the gated self-attention module is first introduced into GANs to focus on key scene areas and filter useless information for strengthening feature representations. Then, the pyramidal convolution block is introduced into the residual block of the discriminator to capture different levels of details in the image using different types of filters with varying sizes and depths for enhancing the feature representations of the discriminator. Additionally, a novel similarity loss item is integrated into the discriminator to leverage self-supervised learning. Besides, spectral normalization is introduced into both the generative network and discriminative network to stabilize training and enhance feature representations. The architecture of multilevel feature fusion is integrated into the discriminative network to achieve more discriminant representation. Experimental results on the AID and NWPU-RESISC45 datasets show that the proposed method achieves the best performance compared to the existing unsupervised classification methods.

Scene classification of high-resolution Remote Sensing Images (RSI) is one of basic challenges in RSI interpretation. Existing scene classification methods based on deep learning have achieved impressive performances. However, since RSI commonly contain various types of ground objects and complex backgrounds, most methods cannot focus on saliency features of scene, which limits the classification performances. To address this issue, we propose a novel Saliency Dual Attention Residual Network (SDAResNet) to extract both cross-channel and spatial saliency information for scene classification of RSI. More specifically, the proposed SDAResNet consists of spatial attention and channel attention, in which spatial attention is embedded into low-level feature to emphasize saliency location information and suppress background information, while channel attention is integrated into high-level features to extract saliency meaningful information. Additionally, several image classification tricks are used to further improve classification accuracy. Finally, extensive experiments on two challenging benchmark RSI datasets are presented to demonstrate that our methods outperform most state-of-the-art approaches significantly.

In the field of object detection, feature pyramid network (FPN) can effectively extract multi-scale information. However, the majority of FPN-based methods suffer from a semantic gap between features of various sizes before feature fusion, which can lead to feature maps with significant aliasing. In this paper, we present a novel multi-scale semantic enhancement feature pyramid network (MSE-FPN) which consists of three effective modules: semantic enhancement module, semantic injection module, and gated channel guidance module to alleviate these problems. Specifically, inspired by the strong ability of the self-attention mechanism to model context, we propose a semantic enhancement module to model global context to obtain the global semantic information before feature fusion. Then we propose the semantic injection module to divide and merge global semantic information into feature maps at various scales to narrow the semantic gap between features at different scales and efficiently utilize the semantic information of high-level features. Finally, to mitigate feature aliasing caused by feature fusion, the gated channel guidance module selectively outputs crucial features via a gating unit. By replacing FPN with MSE-FPN in Faster R-CNN, our models achieve 39.4 and 41.2 Average precision (AP) using ResNet50 and ResNet101 as the backbone network respectively. When using ResNet-101-64x4d as the backbone, MSE-FPN achieved up to 43.4 AP. Our results demonstrate that replacing FPN with MSE-FPN significantly enhances the detection performance of state-of-the-art FPN-based detectors.

With the advent of a large number of remote sensing images (RSIs), scene classification of RSI is widely applied to many fields such as urban planning, natural disaster detection, and environmental monitoring. Compared with the natural image field, the lack of labeled RSI is a bottleneck of supervised scene classification methods based on deep learning. Meanwhile, unsupervised scene classification is difficult to meet actual needs. To this end, we propose a novel semisupervised scene classification method for RSI using generative adversarial nets (GANs), in which a gating unit, a self-attention gating (SAG) module, and a pretrained Inception V3 branch are introduced into discriminative network to enhance the feature representation capability for facilitating semisupervised classification. To be specific, the gating unit aims to learn the weights of each feature map and capture the dependence relationship between features. The SAG module aims to capture a long-range dependence for adaptively focusing on important scene regions. The Inception V3 branch aims to extract the high-level semantic representation of input images and further enhance the discriminant ability by gating unit and SAG module. Furthermore, a new optimization term is incorporated into the generator loss to indirectly drive discriminator to correctly classify scene images. To verify the effectiveness of the proposed method, extensive experimental results on UC Merced and EuroSAT data sets demonstrate that the method surpasses most of the state-of-the-art semisupervised image classification methods significantly, especially when only few samples are tagged.

Remote sensing image (RSI) target detection methods based on traditional multiscale feature fusion (MSFF) have achieved great success. However, the traditional MSFF method significantly increases the computational cost during model training and inference, and the simple fusion operation may lead to the semantic confusion of the feature map, which cannot realize the refined extraction of features by the model. In order to reduce the computational effort associated with the MSFF operation and to enable the features in the feature map to present an accurate, fine-grained distribution, we propose a single-stage detection model (RS-YOLO). Our main additions to RS-YOLO are a computationally smaller and faster Quick and Small E-ELEN (QS-E-ELEN) module and a feature refinement extraction (FRE) module. In the QS-E-ELEN module, We utilize QSBlock, jump-join, and convolution operations to fuse features on different scales and reduce the computational effort of the model by exploiting the similarity of the RSI feature map channels. In order for the model to better utilize the enhanced features, we designed the FRE module to make the location of the enhanced features more accurate and fine. By conducting experiments on the popular NWPU-VHR- 10 and SSDD datasets, we derive results showing that RS-YOLO outperforms most mainstream models in terms of the tradeoff between accuracy and speed. Specifically, in terms of accuracy, it improves 1.6 $\%$ and 1.7 $\%$ compared to the current state-of-the-art models, respectively. At the same time, RS-YOLO reduces the number of parameters and computational effort.