Improving Satellite Imagery using Artificial Intelligence

High-Resolution images are required for satellite sensing, and Super Resolution (SR) provides a feasible solution. Super Resolution Image Reconstruction involves generating a High-Resolution (HR) output image from a Low-Resolution (LR) input image. The recent developments in Deep Learning techniques have boosted the capability of image enhancement methods such as Super Resolution, achieving better performance compared to traditional approaches.In this paper, we explore the performance of various training methods used to achieve Super Resolution, while explaining the overall process in simple terms. Additionally, we validate the effectiveness of the generated SR images using a practical implementation of Road Detection through Semantic Segmentation. This provides an added perspective over conventional evaluation metrics such as PSNR, SSIM, and the Human Visual System (HVS).Our assessment shows that the down-sampling method used to create LR images during training significantly impacts reconstruction performance. Although traditional metrics correlate well with human perception, the use of a semantic segmentation model offers a more application-driven evaluation method. The experimental results indicate that SR outputs with higher visual clarity contribute to better segmentation accuracy, making this combined framework suitable for real-world satellite image enhancement.