COSMOS: A Distributed Multi-Granularity Recognition Framework for Earth Observation Applications in Satellite Edge Computing Networks

Advances in Earth observation have led to a deluge of Ultra-High Resolution (UHR) satellite imagery, overwhelming limited downlink bandwidth, while on-orbit processing is a critical solution. However, UHR object detection on a single resource-constrained satellite faces prohibitive computational costs and poor accuracy. To address this, we propose a distributed multi-granularity recognition framework, which we term COSMOS (COllaborative Satellite Multi-granularity Object Sensing). It first employs lightweight, on-orbit coarse-grained screening to efficiently find Regions of Interest (ROIs). Subsequently, high-resolution ROI data is offloaded to more resourceful edge computing satellites for precise, fine-grained analysis. The complex offloading decisions are solved using a QMIX-based multi-agent reinforcement learning algorithm. Experiments on a hardware-in-the-loop testbed emulating a heterogeneous satellite constellation demonstrate that our COSMOS framework outperforms a slice-and-detect baseline (SAHI) with a centralized DQN scheduler by over 14% in average completion time. Furthermore, this holistic design significantly reduces data transmission volume and maintains high fairness and reliability, proving its practical viability for future on-orbit processing systems.