Scalable conflation for maps data replay between heterogeneous geospatial data sources

Last-mile logistics efficiency largely depends on map data being accurate, fresh, and having high geospatial coverage. Operators must preserve these qualities when migrating between heterogeneous map data providers. Existing work in the literature addresses road-geometry conflation; however, the conflation of road network attributes and their transition relations remains under-studied, especially at a production-level scale. This leaves significant gaps in the map data when operators switch map providers. In this work, we present a scalable map conflation system to close those gaps and replay accumulated map edits onto a refreshed base map. The proposed system has four components. Change-driven partitioning groups in-scope edits into independently conflatable clusters; connected components of a shared-node graph subdivided by size into a hierarchy of uniform spatial grid cells (S2 cells), so the same pipeline drives both continental-level bootstrapping and small-scale incremental refreshing use cases. A logistic-regression map-matching model (with road-segment normalization) conflates both road segment and intersection-level attributes. A random-forest model formulates turn-restriction conflation as intersection matching followed by road matching and scoring candidate pairs with geometric features. An orchestration layer ties these core components into a scalable, unified production pipeline. On the OpenStreetMap and Overture Maps benchmark, the proposed system successfully conflated ∼94.9% of intersection attributes, ∼94.5% of road-segment attributes, and ∼60% of turn restrictions at 99% precision; the intersection-matching classifier reached ∼92% accuracy.