U-Net Network for Remote Sensing Image Super-Resolution

Abstract

Remote sensing image super-resolution is an important technique for enhancing low-resolution satellite and aerial images, especially when high-resolution imagery is expensive, unavailable, or difficult to process in real time. This thesis presents a U-Net-based remote sensing image super-resolution framework that combines AERU-Net fine-tuning, OpenSR-Test-inspired loss analysis, and Edge AI deployment. The study aims to improve reconstruction quality while considering both conventional image-quality metrics and remote-sensing-specific evaluation criteria. The proposed framework was evaluated on the UC Merced Land Use dataset under ×2 and ×4 upscaling settings. Low-resolution inputs were generated from high-resolution reference images using bicubic degradation. Several loss configurations were first investigated using HAU-Net, including L1 loss and OpenSR-inspired auxiliary terms such as SAD and MTF. The selected loss behavior was then further examined using AERU-Net. The results showed that L1 loss remained important for stable training, while properly weighted auxiliary terms could improve reconstruction quality. For AERU-Net fine-tuning, the SAD-based loss configuration achieved the best overall performance. At ×2, scaling, the proposed model achieved 34.66 dB PSNR, 0.9353 SSIM, 0.0476 SAM, and 0.6514 SCC. At ×4 scaling, it achieved 28.02 dB PSNR, 0.7703 SSIM, 0.1002 SAM, and 0.2904 SCC, improving over the original AERU-Net baseline. OpenSR-Test metrics were also used to assess reflectance, spectral, spatial, synthesis, hallucination, omission, and improvement behavior. The fine-tuned model improved synthesis, omission, and improvement scores, indicating better recovery of meaningful high-frequency details. Finally, the optimized model was quantized to INT8 and deployed on the Xilinx Kria KV260 using Vitis AI. The results demonstrate the feasibility of FPGA-based Edge AI deployment for remote sensing super-resolution, although with lower accuracy than the full-precision model.