Development of a Super-Resolution Satellite Image Processing System on an Edge AI Platform

High-resolution satellite imagery is essential for many remote sensing applications, such as land-use mapping, infrastructure monitoring, and environmental assessment. Unfortunately, acquiring high-resolution data is expensive, constrained by limited bandwidth, and limited by infrequent satellite passes. Super-resolution (SR) techniques offer a software-based approach to enhance image resolution by reconstructing high-resolution images from low-resolution inputs, utilizing deep learning. In this work, we present a fully implemented, lightweight attention-enhanced U-Net model for real-time satellite image super-resolution on an edge AI platform. We optimize the model using the Xilinx Vitis AI toolchain and deploy it on the Xilinx Kria KV260 FPGA board. The development pipeline includes quantization-aware training, pruning, model compilation, and hardware deployment. The network architecture is carefully designed to preserve spatial detail while being compatible with quantization and FPGA constraints. We evaluated the system on the UCMerced Land Use dataset with ×2 and ×4 upscaling factors. Results show that the INT8-quantized model produces high-quality super-resolved images with low inference latency and efficient resource usage. These outcomes demonstrate the practicality of the proposed system for real-time, embedded remote sensing applications.