Enhancing the Spatial Resolution of Aerial Images using a Deep Generative Approach with Multi-Stage Residual Refinement

Aerial images are photos taken from above, usually from planes or drones, displaying a landscape as seen from above. High-quality aerial images are crucial in various tasks, like precision agriculture, post-disaster monitoring, and land use classification. However, acquiring these images can be challenging due to the constraints of underlying image sensors and several other factors. This paper introduces a deep generative approach to enhance the spatial resolution of aerial imagery. The proposed method uses an Enhanced Deep Super-Resolution (EDSR) model architecture with a multi-stage residual refinement strategy. This particular model is comprised of residual blocks with two convolutional layers and residual connections. Through a succession of these residual blocks, the model processes low-resolution, 3-channel aerial images and finally upscales the images with a convolutional layer and a depth-to-space transformation. By increasing the spatial resolution of aerial imagery, this deep generative method seeks to overcome the problems that low-quality images present in real-world applications. Findings from the experiment on the VEDAI dataset demonstrate the effectiveness of the proposed method in producing high-quality aerial images. The improvements of PSNR and SSIM metrics by 1.05% and 0.34% respectively and the reduction in MAE and NRMSE by 0.0045 and 1.23% highlight its potential impact on various fields that depend on accurate and comprehensive aerial information. The outcomes of this work will contribute to the future development of picture super-resolution techniques for use in remote sensing and geospatial analysis.