RS-JointNet: A Lightweight Framework for Procedural Knowledge Extraction from Remote Sensing Documents

Abstract

Valuable procedural knowledge in the remote sensing (RS) domain-such as algorithm selection and data processing workflows-is frequently trapped within unstructured academic PDFs and text documents. Extracting this information using traditional Natural Language Processing (NLP) techniques is notoriously difficult due to dense terminology, nested entities, and overlapping semantic relations. Furthermore, deploying Large Language Models (LLMs) for direct online extraction introduces unacceptable computational overhead and hallucination risks. To address these issues, this paper presents a novel, end-to-end NLP framework explicitly designed to mine and structure process-oriented knowledge from raw PDFs and TXT files. The proposed methodology initiates with a dynamic sliding-window chunking strategy to effectively parse lengthy texts. Subsequently, an LLM-guided distillation module leverages multi-level consistency verification to automatically generate a high-fidelity training corpus. This corpus supervises RS-JointNet, a customized lightweight extraction network that combines a domain-adapted SciBERT encoder with a two-dimensional grid tagging decoder. This architecture successfully transforms sequence labeling into matrix classification, adeptly resolving nested and overlapping structures. Ultimately, the extracted triples are instantiated into a Neo4j graph database. Experimental results indicate that our framework achieves an outstanding F1-score of 88.7%. Compared to baseline LLMs, it accelerates inference speed by a factor of 37 while drastically reducing memory consumption, offering a highly accurate and cost-effective NLP pipeline for constructing computable RS knowledge graphs.