Optimizing cloud detection: small-sized deep learning model for on-board implementation using RGB images

Cloud detection is a crucial aspect of remote sensing with numerous research papers focusing on improving precision. Recent advancements have enhanced precision using extensive networks, albeit at the expense of longer processing times. Implementation of these networks often demands powerful graphics processing units. The University of the Bundeswehr Munich has been developing the ATHENE-1 Lower Earth Orbit (LEO) satellite for the Seamless Radio Access Network for Internet of Space (SeRANIS) mission with constrained onboard data storage and processing capacity. Therefore, implementing large-scale models with millions of parameters on ATHENE-1 is impractical and risks operational failure. We utilize open-source labeled data for pre-launch training, considering cross-platform performance challenges and the potential for optimized parameter configuration after launch. To address the two aspects of the satellite mission, we focus on implementing small-sized networks with fewer parameter configurations. Cloud detection is a fundamental step in the ATHENE-1 onboard processing pipeline of the remote sensing imagery. It determines the cloud percentage in an image and assists in decision-making regarding captured imagery. Images heavily obscured by clouds can be promptly discarded and remaining onboard images will be used to generate cloud masks. This processing step results in the optimal usage of onboard data storage and downlink passes required for the ATHENE-1 satellite. The research uses UNet architecture for cloud detection, analyzing how reducing model size and parameters affects performance.