Accounting for Spatial Dependence in Machine Learning: From Improved Prediction to Reliable Interpretation

Abstract

Spatial data are common across environmental, ecological, and scientific applications, yet most machine learning methods assume independence among observations, an assumption that is routinely violated in practice. When spatial dependence is ignored, it can both degrade predictive performance and distort model interpretation, with spatially structured predictors appearing artificially important because they reflect shared spatial patterns rather than meaningful relationships with the response.   In this talk, we present a unified framework for addressing spatial dependence through a scalable spatial decorrelation (whitening) transformation that removes local spatial correlation from both predictors and responses using Vecchia-based approximations, allowing standard machine learning and deep learning models to be applied without modification and predictions to be mapped back to the original spatial scale.   Through simulations and real-world ecological data, we show that this approach improves predictive accuracy while, more importantly, mitigating “illusions of importance,” yielding variable importance measures and partial dependence plots that reflect true functional relationships rather than spatial artifacts, thereby enabling more trustworthy scientific inference.