LLM-Based Satellite Multimodal Knowledge Graph Construction: A Context-Aware Self-Interpretable Paradigm

This paper addresses three main bottlenecks in commercial aerospace multimodal data: semantic alignment difficulties, rigid ontologies, and context loss. We propose a method to construct satellite multimodal knowledge graphs (KGs) using Large Language Model (LLM).The process begins with layout analysis. This step preprocesses interleaved text-image satellite documents to align cross-modal semantics. To resolve ambiguity in technical terms, we developed a triple-positioning entity extraction mechanism. This mechanism directly identifies the original text, summaries, and entity types. For event nodes, we use an event-centric modeling paradigm. This paradigm captures spatio-temporal elements and other meta informations to ensure the nodes are self-interpretable. Entity normalization and automatic schema recognition rely on deep vector embedding. This embedding can semantic clustering the entities for the KG, eliminates the need for prior ontologies. We tested the method on real satellite datasets. The experimental results show high logical integrity and confirm its practical value for aerospace failure analysis.