Entropy-Driven Scanning Optimization for Near Real-Time Earth Observation

ICLR 2026 Conference Submission13315 Authors

Keywords: Earth Observation, Optimization

Abstract: Earth observation aims to collect geospatial information using remote sensing satellites. However, traditional systems often require days or even weeks to achieve full-region coverage. In this paper, we present the first entropy-based formulation of satellite scanning optimization, designed to enable near real-time Earth observation with large-scale Low Earth Orbit (LEO) constellations. Unlike conventional coverage plans that follow rigid orbital patterns, our approach directly maximizes spatial entropy over imaging point distributions, promoting diversity and fairness in spatiotemporal coverage. This principled objective prevents redundant observations, ensures balanced regional attention, and provides smooth transitions between successive scan plans. To operationalize the framework, we introduce a differentiable solver that maps optimized imaging points into physically executable camera angles, and an efficient satellite-to-task assignment module that minimizes slewing effort through a hybrid of the Hungarian algorithm and nearest-neighbor heuristics. Experimental results demonstrate that our framework achieves full-region coverage within minutes and delivers up to 10× faster scanning compared to conventional orbit-based strategies.

Primary Area: optimization

Submission Number: 13315