Accelerating RavÆn for Real-Time Satellite Image Change Detection

Data processing on satellites supports real-time identification of critical information and preferential downlink. RaVÆn, a lightweight, unsupervised learning approach based on Variational Auto-Encoders, has been proposed to detect changes in satellite images. However, there is a significant gap between the required processing time and the practical image update frequency generated by Earth observation satellites. In this paper, we explore the use of Field Programmable Gate Arrays (FPGAs) to deploy RaVÆn to achieve high computing speed within strict power constraints. We used high-level synthesis to structure the hardware task into six processing stages and optimized each stage to balance resource utilization and performances. Our system, based on a Kria KV260 board, is capable of processing non-overlapping images as captured on the Sentinel2 satellite in real-time, with a 23 % time margin. Compared to a CPU-based implementation, our accelerated implementation achieves a 109-fold throughput improvement and 69-fold energy saving while maintaining high accuracy. The outcome demonstrates the exceptional performance of an FPGA-based design for executing the RavÆn algorithm on satellite platforms.