On-Orbit AI: Cloud Detection Technique for Resource-Limited Nanosatellite

This paper presents a novel cloud detection technique for resource-limited nanosatellites, addressing challenges of limited onboard computational power and communication bandwidth. We developed a multi-phase neural network-based algorithm, combining a lightweight TriCloudNet classifier with a pruned U-Net architecture, optimized for space-proven onboard computers. Validated on a Xilinx Zynq-7000 SoC-based OBC operational in SNIPE mission satellites, our approach achieved a 6.21 × speedup in inference time through custom FPGA-based hardware acceleration. The system reduces downlink data volume by 40–50% via onboard cloud detection and image prioritization. Experimental results confirm real-time processing feasibility within 6U CubeSat constraints, with minimal power consumption increase. This research significantly enhances efficiency of nanosatellite Earth-observation missions, demonstrating effective implementation of complex neural networks on space-constrained hardware. Our work paves the way for more intelligent, autonomous space-based imaging systems, maximizing scientific and commercial value of nanosatellite missions.