Logo image
Quantifying Uncertainty in Spatial Prediction for Nonstationary Spatial Processes
Journal article   Peer reviewed

Quantifying Uncertainty in Spatial Prediction for Nonstationary Spatial Processes

Peng Luo
Annals of the American Association of Geographers, pp.1-25
04/16/2026
DOI: 10.1080/24694452.2026.2639721
url
https://doi.org/10.1080/24694452.2026.2639721View
Published (Version of record) Open Access

Abstract

Uncertainty quantification for geospatial prediction models plays a crucial role in evaluating model confidence and supporting informed decision-making. Conformal prediction (CP), as a model-agnostic framework, has recently been introduced into geospatial settings, leading to the development of geospatial conformal prediction (GeoCP). GeoCP considers spatial dependence and incorporates a geographic weighting mechanism into CP to capture spatially varying uncertainty, satisfying the localized exchangeability assumption. In many real-world scenarios, however, spatial processes could exhibit abrupt transitions, such as neighboring regions with distinct land-use types, resulting in significant differences despite close geographic proximity. To address this limitation, we propose geospatial similarity conformal prediction (GeoSIMCP), an extension of GeoCP that jointly considers both geographic distance and feature-space similarity when estimating local uncertainty. We further develop a parameter optimization framework to ensure robust model tuning. Through comprehensive simulation studies under varying spatial structures, we demonstrate the advantages of GeoSIMCP over GeoCP. Additionally, we validate the effectiveness of GeoSIMCP on two real-world prediction tasks-housing prices and PM2.5 concentration-characterized by different spatial processes. Our results highlight the potential of integrating geographic and feature similarity to enhance uncertainty quantification in spatial prediction, offering a more adaptive and reliable framework for geospatial decision-making.
Geography Social Sciences

Details

Metrics

1 Record Views
Logo image