Title: Beyond Pixel-Based Spatial Representation: Integrating Raster Imagery with Vector Semantics for Human-Centric Geospatial Foundation Models

Abstract:

Earth Observation (EO) has revolutionized our ability to monitor environmental dynamics and human activities on a global scale. The emergence of Earth Observation Foundation Models (EOFMs), driven by recent breakthroughs in self-supervised learning, now enables the extraction of transferable representations from petabytes of unlabeled EO data for various downstream geospatial applications. Nevertheless, current EOFMs are predominantly restricted to raster modalities, thereby neglecting the rich, structured insights contained within openly accessible vector datasets like OpenStreetMap and Overture.

Vector data delivers explicit, compact depictions of geographic entities, encompassing geometry, topology, and semantic linkages. These elements provide vital contextual cues that are frequently ambiguous or entirely missing in imagery alone. Raster and vector data serve as complementary perspectives of geographic space: while raster captures continuous physical and spectral patterns, vector data encodes discrete objects along with their relational structures. Notably, vector data often reflects human-centric systems—such as social or demographic information—rather than purely physical phenomena.

However, prevailing geospatial representation learning frameworks address these modalities in isolation, depending on imperfect and frequently lossy transformations to connect them. This perspective paper advocates for a paradigm shift toward joint Spatial Representation Learning (SRL) within a unified embedding space that merges raster-based perception with vector-driven reasoning. Drawing on nascent developments in multimodal geospatial learning, we outline the conceptual underpinnings, technical hurdles, and promising avenues for aligning diverse spatial data sources. We argue that such integration is indispensable for crafting next-generation geospatial AI systems that offer a more accurate, interpretable, and semantically grounded comprehension of our planet.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC