Towards operational application of Deep Reinforcement Learning to Earth Observation satellite scheduling

Adrien Hadj-Salah (1) , Jonathan Guerra (1, 2) , Mathieu Picard (1, 2) , Mikaël Capelle (1)

1 IRT Saint Exupéry - Institut de Recherche Technologique
2 Airbus Defence and Space [Les Mureaux]

Scheduling an agile optical Earth Observation satellite (AEOS) requires to choose a limited number of images to take among a large set of possibilities. It is a NP-hard problem which is made even more difficult by the presence of uncertainties: an image is useless if it is too cloudy and weather uncertainty cannot be removed, no matter the accuracy of the forecasts. Moreover, among the various types of customer requests of commercial satellites, we focus here on large area acquisitions that cover a country or a continent. Such requests require several months or years to complete even with a constellation of satellites. Considering such long time frames, the completion time highly depends on weather uncertainties and there are currently no trustworthy forecasts. Therefore the selection of the requests is crucial to speed up the completion using a long-term strategy. Reinforcement Learning is an interesting solution to explore when it comes to uncertain environments. We propose to use the well-known Actor Critic (A2C) algorithm combined with Transfer Learning, Domain Knowledge and Domain Randomization (TDDR). We demonstrate how transfer learning is a way to address real-world problem. We find that TDDR method challenge state-of-the-art heuristics for satellite scheduling on various real weather conditions.

Domaines: Intelligence artificielle [cs.AI]

Adrien Hadj-Salah, Jonathan Guerra, Mathieu Picard, Mikaël Capelle. Towards operational application of Deep Reinforcement Learning to Earth Observation satellite scheduling. 2020. ⟨hal-02925740⟩