The increasing demand for artificial intelligence (AI)-based on-board data processing in space missions is reshaping the design of satellite payloads. Field-Programmable Gate Arrays (FPGAs) are considered one of the most promising solutions for accelerating deep neural networks (DNNs) in orbit, thanks to their flexibility, power efficiency, and availability in radiation-tolerant and radiation-hardened versions. However, conventional design methodologies for FPGA-based accelerators suffer from high development effort and long time-to-market. To address this issue, this paper evaluates FPG-AI, a vendor-agnostic automated toolflow for FPGA-based AI acceleration, in a representative space use case. Specifically, we consider the semantic segmentation of maritime and terrestrial areas from Sentinel-2 imagery using a U-Net–like architecture trained on the Sentinel-2 Water Edges Dataset (SWED). The network achieves 90% accuracy and a mean IoU of 0.81, demonstrating robustness for coastal monitoring applications. The hardware acceleration was applied to the encoder through FPG-AI, and the designs were implemented on both commercial (Xilinx XCZU7EV) and space-grade devices (Xilinx RT Kintex UltraScale XQRKU060 and NanoXplore NG-ULTRA). Results show that FPG-AI can generate hardware accelerators capable of near real-time inference (49 ms, 0.98 W on XCZU7EV; 73 ms, 0.93 W on XQRKU060), with negligible accuracy degradation (<0.15%) after quantization. These findings demonstrate the readiness of FPG-AI for practical in-orbit deployment, significantly raising its Technology Readiness Level (TRL).
Bocchi∗ et al. (Wed,) studied this question.