Measuring the spatial explicitness of GeoAI and geospatial foundation models: A systematic review and future roadmap

GeoAI and geospatial foundation models (GeoFMs) are fundamentally reshaping how geospatial scientists analyze data, model spatial processes, and support decisions. However, the rapid expansion of this field renders static literature reviews quickly obsolete. A timely and dynamic review is therefore crucial for accurately capturing the current research landscape and projecting future trajectories in GeoAI and GeoFM research. This systematic review synthesizes 1,122 peer-reviewed manuscripts to address a central question: are current methodologies becoming truly spatially explicit? Adhering to a PRISMA-guided workflow, we retrieved records from the Web of Science and DBLP, conducted structured human screening, and extracted standardized study metadata, including methodological labels, from full texts. This extraction utilized a Retrieval-Augmented Generation (RAG) workflow with rigorous manual validation, achieving over 90% agreement in a 5% random audit. The results reveal a research landscape dominated by remote sensing and raster data, with a technical mainstream centered on Convolutional Neural Networks (CNNs) and transformers targeting classification, segmentation, and representation learning tasks. Nevertheless, most methodological novelty remains framed as application-driven or as non-theory-driven structural modifications. Designs explicitly grounded in geographic knowledge, geospatial pretraining, and genuine GeoFM innovations remain comparatively limited. We conclude by presenting a practical roadmap that offers insights into data demands, GeoAI/GeoFM development, educational paradigms, and industry evolution, all of which are critical for advancing the next generation of spatially explicit GeoAI and GeoFMs.