Answers are grounded in this paper and cite their sources.
Large-scale modeling of environmental variables is an increasingly complex but necessary task. In this paper, we review the literature on using machine learning to cope with challenges associated with spatial autocorrelation. Our focus was on studies in which researchers predicted environmental variables using a supervised machine learning regression algorithm that accounted for spatial autocorrelation in any part of the modeling pipeline from data exploration to model validation. Methods included explicit spatial covariates, data splitting for training–testing, model calculations, and independent exploratory analysis. Authors accounted for spatial autocorrelation most often in the independent exploratory analysis which had no impact on the predicted values. We concluded that there seems to be no overall systematic approach on how to account for spatial autocorrelation in machine learning models. In the selected studies, the most appropriate spatial method depended on the specific characteristics of the study. Using explicit spatial covariates and data splitting for training-testing provided more insights into a method's applicability. We summarize these insights and provide considerations for selecting the most appropriate method.
No takes yet. Share an insight, caveat, or question.
Jemeļjanova et al. (2024) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: