A MULTI-AGENT SYSTEM TO ORCHESTRATE INTERACTIONS WITH DIGITAL TWINS OF EARTH
We present a new-generation, AI-agent-powered digital assis tant featuring four specialized engines for satellite imagery: search by image, search by caption, visual question answering, and knowledge graph question answering. At the core of the system is a Task Interpreter, designed as a multi-agent system, which coordinates these engines to address complex user requests for Earth observation data. The Task Interpreter comprises four agents: an Engine Routing Agent that selects the appropriate engine or rejects unmanageable requests; a Conversational Agent that handles general or out-of-scope queries; an Argument Extraction Agent that identifies image type parameters for retrieval tasks; and a Tool Feasibility Agent that assesses the applicability of tools for domain specific queries. This multi-agent system enables seamless interaction with Digital Twins of Earth, with an emphasis on modularity and extensibility to adapt to the rapid evolution of remote sensing technologies.
