CivicEmbed: Feature-specific embeddings for efficient geographic reasoning and retrieval

Josephine Wang, Julien Coquet, Jeffrey Huang

Keywords: geospatial retrieval, embedding space, contrastive learning, self-supervised representation learning, S2 spatial partitioning, urban planning, similarity search

TL;DR: CivicEmbed proposes compressed feature-specific embeddings that enable urban planners to search for geographically similar locations based on characteristics like topography, water proximity, vegetation, and road networks.

Abstract: Diverse geography, from natural landscapes to urban areas, poses challenges for terrain-sensitive development. Existing geospatial embeddings are largely monolithic, limiting feature-specific comparison. We present CivicEmbed, a lightweight approach that learns separate embedding spaces for topography, water proximity, vegetation cover, and road density using self-supervised contrastive learning on thematic raster layers. These modular encoders support efficient analogical reasoning via retrieval: users can emphasize individual geographic features or weighted combinations of features to retrieve spatial analogs that match the selected constraints. Feature-specific encoders achieve $32\times$ compression ($128$-D vs. $4,096$-D raw patches) while improving retrieval metrics on certain features. We implemented a FAISS-backed retrieval system at the scale of Switzerland, providing a foundation for data-driven decisions in architecture, transit design, and land-use planning.

Submission Number: 21