Enhancing Dual Aggregation Approach for Remote Sensing Image Super-Resolution using Transformer

Convolutional neural networks (CNNs) have significantly progressed in various remote sensing tasks. However, they often rely on basic or residual blocks for low-level feature extraction, which overlook important features during preprocessing and fail to capture spatial details effectively. Additionally, these models struggle to incorporate global information necessary for reconstructing high-resolution images. To overcome these problems, we propose a novel feature extraction block called the Enhanced Dual Aggregation Approach (EDAA). This module consists of Enhanced Global Contextual Attention (EGCA) and Enhanced Deep Spatial Attention (EDSA). These are followed by the Activated Sparsely Sub-Pixel Transformer (ASSPT), which exploits features at different stages. EGCA is designed to capture long-range dependencies and contextual information across the image. On the other hand, EDSA focuses on refining spatial features at a more local or fine-grained level. It ensures that small, detailed structures in the image are accurately captured and enhanced during the feature extraction stage. ASSPT is used to leverage remote sensing information fully. It incorporates EDAA at various levels, maximizing the efficient use of global information. The reconstruction block adds a local detailed refinement process before the final output. The experimental results show that our proposed model achieves superior results. It also demonstrates better image reconstruction quality compared to state-of-the-art models.