Edge SpAIce: Deep Learning Deployment Pipeline for Onboard Data Reduction on Satellite FPGAs

Earth Observation satellites are essential for monitoring global phenomena. However, their potential is often constrained by limited downlink bandwidth and ground-based processing capabilities. To address these challenges, the EU-funded Edge SpAIce project aims to advance the deployment of Deep Convolutional Neural Networks onboard satellites, enabling efficient data processing directly at the edge. This project has also the objective of accelerating the spread of AI technology deployment in space by demonstrating a challenging use case scenario for marine plastic litter detection. We introduce a pipeline to deploy optimized Neural Networks on Field Programmable Gate Arrays through hls4ml, for the purpose of segmenting satellite imagery into object classes. The Edge SpAIce project is presented, along with its current achievements and future developments.