Edge AI for Earth Observation

Earth observation (EO), edge computing, and artificial intelligence (AI) are rapidly advancing technologies with diverse applications and benefits. Integrating edge computing and AI with EO enables the preprocessing and analysis of EO data near its source, supporting efficient decision-making and in-orbit information interpretation. In this context, this article provides a review of the current state of edge AI in EO applications, summarizes the key challenges, including data sample limitations, computing resource constraints, catastrophic forgetting, and difficulties with satellite-ground coordination. Also, we explore possible solutions and techniques such as including generalization under small sample conditions, lightweight model design and training (e.g., pruning, quantization, distillation), continuous learning for multiple tasks, and satellite-ground continuum systems (e.g., federated learning and resource-constrained inference). Finally, we outline possible future research directions to address the challenges using edge AI for EO scenarios.