HyDA-Net: A Hybrid Dense Attention Network for Remote Sensing Multi-Image Super-Resolution

Remote sensing imagery super-resolution (SR) has gained focus due to its importance in enhancing satellite images to aid in the study of Earth. Recently, deep learning has advanced significantly in SR, particularly in single-image super-resolution (SISR). However, SISR still struggles with challenges, such as atmospheric occlusion and sensor noise, which degrade image quality. In this article, a novel hybrid dense attention network (HyDA-Net) is proposed that highlights the idea of multi-image SR to address SISR problems. HyDA-Net is a three-branch architecture that emphasizes the importance of information compensation and the correlation between multiple images of the same scene by fusing 3-D and 2-D features. HyDA-Net introduces a novel 3-D dense attention block (3D-DAB) designed to improve the preservation of fine details in satellite images. 3D-DAB integrates a 3-D dense block and a tailored feature attention mechanism with a 3-D convolution to skillfully capture high-frequency details of dense features from multiple low-resolution satellite images that will complement the low-frequency components. In addition, 3D-DAB has a global residual connection and multilevel local residual connections inside 3-D dense block to avoid the vanishing gradient problem during training. Extensive experiments using real-captured satellite datasets, namely PROBA-V and MuS2, show that HyDA-Net outperforms state-of-the-art models in different spectral bands. Moreover, a cross-dataset experiment is conducted to further evaluate the robustness and generalizability of the proposed HyDA-Net.