ICMTEST-2026 - Geotoolrag: an LLM-Based Workflow Planning Framework for Geospatial Decision-Making

The rapid expansion of high-resolution geospatial data from sources such as satellite imagery, remote sensing platforms, and open geospatial databases has significantly increased the demand for intelligent systems capable of automating complex spatial workflows. Traditional Geographic Information System (GIS) pipelines rely heavily on manual scripting, tool selection, and expert-driven decision-making, which limits scalability, reproducibility, and accessibility for non-expert users (Goodchild, 2007; Longley et al., 2015). As geospatial applications continue to grow in areas such as disaster management, urban planning, and environmental monitoring, there is a critical need for automated systems that can translate high-level user intent into executable spatial analysis workflows. This paper proposes GeoToolRAG, a novel Large Language Model (LLM)-based workflow planning framework that integrates dual-channel Retrieval-Augmented Generation (RAG) to bridge the gap between natural language understanding and geospatial computation. The proposed system distinguishes between tool-level knowledge (e.g., APIs, parameters, and data formats) and domain-level knowledge (e.g., hydrological modeling, terrain analysis), enabling more accurate and context-aware workflow generation. Using Chain-of-Thought reasoning (Wei et al., 2022), the system generates structured, interpretable workflows represented as Directed Acyclic Graphs (DAGs), which are subsequently compiled and executed using Python-based GIS libraries such as GeoPandas and Rasterio. The effectiveness of the proposed framework is evaluated on representative geospatial tasks, including flood-risk mapping and site suitability analysis. Experimental results demonstrate that GeoToolRAG achieves improved workflow validity, higher spatial prediction accuracy, and significantly reduced execution errors compared to baseline approaches, including rule-based GIS pipelines and code-only LLM agents. The framework also enhances interpretability by producing explicit workflows that align closely with expert-designed processes. Overall, this study contributes (i) a unified architecture for LLM-driven geospatial workflow automation, (ii) a dual-channel retrieval strategy for integrating heterogeneous knowledge sources, and (iii) a comprehensive evaluation framework that jointly assesses workflow planning quality and task performance. The proposed approach highlights the potential of combining large language models with geospatial systems to enable scalable, transparent, and intelligent spatial decision-making.