基于多智能体工作流的遥感大模型智能解译方法研究

Research on Intelligent Interpretation Method of Remote Sensing Large Model Based on Multi Agent Workflow

摘要: 随着高分辨率遥感卫星、航空遥感平台以及无人机遥感系统的快速发展，全球遥感数据规模呈现指数级增长。传统遥感智能解译方法虽然在分类、检测与变化监测等任务中取得显著进展，但仍存在标注依赖强、跨场景泛化不足，以及缺乏面向复杂任务的系统化编排能力等问题。遥感基础模型与多模态大模型为智能解译提供了强大的表征与认知能力，但如何将这些模型能力组织为可复现、可扩展、可交互的解译流程，仍是制约其业务落地的关键瓶颈。为了解决这一问题，本文聚焦于多智能体工作流，研究了一种面向遥感大模型智能解译的系统化方法。首先，回顾了基础模型和多模态大模型的发展现状，分析了SatMAE、Prithvi、GeoCLIP和GeoGPT等具有代表性的模型。随后，阐明了遥感智能解译范式正由传统的目标识别向场景理解、时空推理以及智能决策演进。在此基础上，提出了一种融合基于MCP的工具服务、多智能体协同、工具调用和工作流编排的遥感大模型智能解译方法，构建了从自然语言任务输入到分析报告输出的完整处理流程。最后，讨论了遥感大模型智能解译面临的数据构建、模型幻觉以及可信人工智能等关键问题，并展望了遥感智能体和数字地球大模型的发展方向。

Abstract: With the rapid development of high-resolution remote sensing satellites, aerial platforms, and unmanned aerial vehicle systems, global remote sensing data have grown exponentially. Although traditional intelligent interpretation methods have achieved remarkable progress in classification, detection, and change monitoring, they still suffer from heavy annotation dependence, insufficient cross-scene generalization, and a lack of systematic orchestration for complex tasks. Remote sensing foundation models and multimodal large models provide strong representation and cognitive capabilities, yet how to organize these capabilities into reproducible, scalable, and interactive interpretation workflows remains a critical bottleneck for operational deployment. To address this issue, this paper focuses on multi-agent workflows and investigates a systematic method for remote sensing large model intelligent interpretation. It first reviews foundation models and multimodal large models, analyzing representative models such as SatMAE, Prithvi, GeoCLIP, and GeoGPT. It then clarifies the paradigm shift from target recognition to scene understanding, spatiotemporal reasoning, and intelligent decision-making. On this basis, a remote sensing large model intelligent interpretation method integrating MCP-based tool services, multi-agent collaboration, tool calling, and workflow orchestration is proposed, forming a complete pipeline from natural language task input to analytical report output. Finally, key issues such as data construction, model hallucination, and trustworthy artificial intelligence are discussed, and future directions of remote sensing agents and digital earth large models are outlined.