Learning the Feasibility of Sets of Acquisition Tasks for Earth Observation Satellites
Résumé
Planning Earth Observation satellite activities involves computing observation sequences while considering timedependent orientation maneuvers. We introduce a novel approach leveraging neural networks and image classification to learn scheduling feasibility for observation tasks. A 0/1 classifier takes an image of POIs along the satellite's orbit as input, estimating whether a schedulable sequence exists within temporal constraints. Learned offline, this classifier enables quick online selection of optimal POI subsets (among candidates) without requiring numerous costly maneuver computations. Experiments show our classifier achieves high accuracy in approximating scheduling feasibility, allowing for efficient optimization with minimal risk of misclassified solutions.
| Origine | Fichiers éditeurs autorisés sur une archive ouverte |
|---|---|
| Licence |