Home // GEOProcessing 2025, The Seventeenth International Conference on Advanced Geographic Information Systems, Applications, and Services // View article

Coordinates Are Just Features: Rethinking Spatial Dependence in Geospatial Modeling

Authors:
Yameng Guo
Seppe vanden Broucke

Keywords: geospatial regression; ensembles modeling; spatial statistics; comparative performance.

Abstract:
Geospatial inference is crucial for various spatial prediction tasks, where the choice of modeling approach significantly impacts both inference performance and computational efficiency. Traditional geospatial statistical models, such as Geographically Weighted Regression (GWR) and Kriging, explicitly account for spatial dependence, but often come with high computational costs. In this study, we argue that treating coordinates as standard input features can yield competitive inference performance while significantly reducing computational costs when selecting a predictive model with an appropriate level of complexity. To support this, we compare geospatial statistical models with various machine learning approaches, including linear methods, tree ensemble methods, hybrid kernel-based methods that incorporate explicit geospatial learning, and a recent state-of-the-art tabular deep learning model—TabPFN—to assess their effectiveness in spatial prediction tasks (to the best of our knowledge, this is the first study to investigate the performance of TabPFN in the geospatial domain using explicit coordinate inputs). Our results demonstrate that when coordinates are sufficiently informative, tree-based ensemble models and tabular deep learning can implicitly capture spatial dependence without requiring explicit geospatial modeling, achieving superior performance whilst maintaining a reasonable computational cost.

Pages: 48 to 55

Copyright: Copyright (c) IARIA, 2025

Publication date: May 18, 2025

Published in: conference

ISSN: 2308-393X

ISBN: 978-1-68558-269-2

Location: Nice, France

Dates: from May 18, 2025 to May 22, 2025