A Review of Deep Reinforcement Learning for Autonomous Earth Observation Satellite Scheduling

The proliferation of large-scale Earth observation satellite constellations has substantially increased mission scheduling complexity, where dynamic operational constraints, heterogeneous task requirements, and stringent real-time demands must be handled concurrently. Deep reinforcement learning has emerged as a promising paradigm for autonomous scheduling by enabling agents to learn adaptive policies through interaction with simulated mission environments. This article presents a systematic review of deep reinforcement learning approaches to Earth observation satellite scheduling, synthesizing 54 eligible studies identified through a PRISMA-style screening workflow spanning IEEE Xplore, Web of Science, Scopus, and a supplementary Google Scholar search.We introduce a reference constrained Markov decision process formulation that consolidates notation and modeling choices across the reviewed literature, serving as a structured analytical lens for cross-study comparison rather than a novel algorithmic contribution. Building on this formulation, we categorize representative methods into value-based, policy-based, and model-based families, and provide critical assessments of convergence behavior, scalability, constraint satisfaction, and deployment readiness for each paradigm. A structured cross-study synthesis maps 30 representative studies against seven modeling dimensions, revealing field-wide patterns and evidence gaps. The review concludes with a standardized minimum reporting checklist to improve reproducibility and with concrete research directions toward flight-qualified autonomous scheduling.