OPTIMIZED U-NET ARCHITECTURE FOR CLOUD DETECTION FOR THE UM5-EOSAT NANOSATELLITE

Speaker: Imane Khalil, Mohammadia School of Engineers, Rabat, MoroccoThe implementation of deep learning for cloud detection onboard Earth observation nanosatellites has shown great potential in improving mission performance. However, challenges related to onboard processing efficiency and energy consumption persist. Traditional methods, including image capture, onboard storage, and transmission of cloud-obstructed images, face significant challenges due to limited bandwidth and storage capacity. These constraints highlight the need for efficient onboard artificial intelligence to process data in real-time and ensure effective cloud detection, especially within the edge constraints of nanosatellite systems. In this work, we present a dual-optimization strategy for cloud detection, where we customize and optimize the U-Net architecture to meet strict energy constraints.