Semantic segmentation in very-high-resolution (VHR) aerial images is one of the most challenging tasks in remote sensing image understanding. Most of the current approaches are based on deep convolutional neural networks (DCNNs). However, standard convolution with local receptive fields fails in modeling global dependencies. Prior research works have indicated that attention-based methods can capture long-range dependencies and further reconstruct the feature maps for better representation. Nevertheless, limited by the mere perspective of spatial and channel attention and huge computation complexity of self-attention (SA) mechanism, it is unlikely to model the effective semantic interdependencies between each pixel pair of remote sensing data with complex spectra. In this work, we propose a novel attention-based framework named hybrid multiple attention network (HMANet) to adaptively capture global correlations from the perspective of space, channel, and category in a more effective and efficient manner. Concretely, a class augmented attention (CAA) module embedded with a class channel attention (CCA) module can be used to compute category-based correlation and recalibrate the class-level information. In addition, we introduce a simple yet effective region shuffle attention (RSA) module to reduce feature redundant and improve the efficiency of SA mechanism via regionwise representations. Extensive experimental results on the ISPRS Vaihingen, Potsdam benchmark, and iSAID data set demonstrate the effectiveness and efficiency of our HMANet over other state-of-the-art methods.

Semantic segmentation of aerial images refers to assigning one land cover category to each pixel. This is a challenging task due to the great differences in the appearances of ground objects. Many attempts have been made during the past decades. In recent years, convolutional neural networks (CNNs) have been introduced in the remote sensing field, and various solutions have been proposed to realize dense semantic labeling with CNNs. In this letter, we propose shuffling CNNs to realize semantic segmentation of aerial images in a periodic shuffling manner. This approach is a supplement to current methods for semantic segmentation of aerial images. We propose a naive version and a deeper version of this method, and both are adept at detecting small objects. Additionally, we propose a method called field-of-view (FoV) enhancement that can enhance the predictions. This method can be applied to various networks, and our experiments verify its effectiveness. The final results are further improved through an ensemble method that averages the score maps generated by the models at different checkpoints of the same network. We evaluate our models using the ISPRS Vaihingen and Potsdam data sets, and we acquire promising results using these two data sets.

The building extraction from synthetic aperture radar (SAR) images has always been a challenging research topic. Recently, the deep convolution neural network brings excellent improvements in SAR segmentation. The fully convolutional network and other variants are widely transferred to the SAR studies because of their high precision in optical images. They are still limited by their processing in terms of the geometric distortion of buildings, the variability of building structures, and scattering interference between adjacent targets in the SAR images. In this article, a unified framework called selective spatial pyramid dilated (SSPD) network is proposed for the fine building segmentation in SAR images. First, we propose a novel encoder-decoder structure for the fine building feature reconstruction. The enhanced encoder and the dual-stage decoder, composed of the CBM and the SSPD module, extract and recover the crucial multiscale information better. Second, we design the multilayer SSPD module based on the selective spatial attention. The multiscale building information with different attention on multiple branches is combined, optimized, and adaptively selected for adaptive filtering and extracting features of complex multiscale building targets in SAR images. Third, according to the building features and SAR imaging mechanism, a new loss function called L-shape weighting loss (LWloss) is proposed to heighten the attention on the L-shape footprint characteristics of the buildings and reduce the missing detection of line buildings. Besides, LWloss can also alleviate the class imbalance problem in the optimization stage. Finally, the experiments on a large-scene SAR image dataset demonstrate the effectiveness of the proposed method and verify its superiority over other approaches, such as the region-based Markov random field, U-net, and DeepLabv3+.

In recent years, remote sensing (RS) vision foundation models such as RingMo have emerged and achieved excellent performance in various downstream tasks. However, the high demand for computing resources limits the application of these models on edge devices. It is necessary to design a more lightweight foundation model to support on-orbit RS image interpretation. Existing methods face challenges in achieving lightweight solutions while retaining generalization in RS image interpretation. This is due to the complex high and low-frequency spectral components in RS images, which make traditional single CNN or Vision Transformer methods unsuitable for the task. Therefore, this paper proposes RingMo-lite, a RS lightweight network with a CNN-Transformer hybrid framework, which effectively exploits the frequency-domain properties of RS to optimize the interpretation process on several tasks like classification, object detection, semantic segmentation, and change detection. It is combined by the Transformer module as a low-pass filter to extract global features of RS images through a dual-branch structure, and the CNN module as a stacked high-pass filter to extract fine-grained details effectively. Furthermore, a novelty-designed frequency-domain masked image modeling (FD-MIM) is employed during the pretraining stage for self-supervised learning, which combines the high-frequency and low-frequency characteristics of each image patch. This approach effectively captures the latent feature representation in RS data. As shown in Fig. 1, compared with RingMo, the proposed RingMo-lite reduces the parameters over 60% in various RS image interpretation tasks, the average accuracy drops by less than 2% in most of the scenes and achieves SOTA performance compared to models of the similar size. In addition, our work will be integrated into the MindSpore computing platform in the near future.