LLM-Based Map Conflation: Performance Assessment on Matching Embedded Road Lines

Abstract

Map conflation is essential for integrating heterogeneous road datasets, but it often requires region- and data-specific algorithm design to automate the complex identification of feature-to-feature correspondences. This effort is increased when only cartographic products are available instead of GIS-ready vectors since both digitization or matching corresponding features manually are labor-intensive. In this study, we assess the performance of a multimodal LLM, GPT-5 thinking mode for map conflation directly on a PDF map where road networks from TomTom and OpenStreetMap are embedded as colored polylines. We instruct the LLM to interpret the PDF, extract road geometries and their identifiers, and generate both strict 1:1 and flexible M:N matches. In any hybrid-patterned network cases located around Bosphorus, Istanbul, while M:N matching process increased the number of matches, it also increased false positives and lowered overall F1 scores. In contrast, 1:1 matching produced more balanced correctness-completeness results. The model achieves its highest performance in the cellular-patterned networks. The results show that LLM-based matching can detect a substantial share of true correspondences in such a challenging hybrid setting, but performance clearly depends on the matching strategy: strict or flexible. It highlights both the potential promise and the current limitations of matching embedded road lines.