From Commands to Cognition: An LLM-Driven Satellite Agent for Autonomous Spectrum Sensing

A low Earth orbit (LEO) satellite traverses multiple distinct electromagnetic environments within every 94-min orbit, from signal-dense urban corridors that demand high-precision detection to near-silent oceans where lighter sensing conserves scarce battery. Existing on-board schedulers can only switch among preconfigured sensing modules, and cannot reason about task feasibility, resource implications, or ambiguous mission intents issued from the ground. This limitation is increasingly critical for cognitive radio IoT (CR-IoT), where timely and wide-area spectrum awareness underpins dynamic spectrum access and interference management. In this article, we propose LEOScA, an intent-aware large language model (LLM)-driven orchestration framework for LEO satellite spectrum monitoring. The framework employs an LLM as a high-level reasoning engine to transform natural language monitoring intents into structured execution strategies. To further mitigate the inherent limitations of LLM in structured decision-making scenarios, we design a prompt-chain-based hierarchical strategy scheduling (PC-HSS) algorithm that decomposes on-board decision-making into three coordinated modules and invokes the LLM only upon the arrival of a new task. Experimental results show that LEOScA achieves 100% feasible-task completion rate (TCR), compared with 88% for the best static baseline, while reducing cumulative energy consumption by 9.7% under mixed-instruction scenarios. These results demonstrate the feasibility of autonomous LLM-driven spectrum sensing on resource-constrained satellite platforms, thereby enabling self-managing CR-IoT infrastructure in space.