Random Spatial Forests

TL;DR

This paper presents random spatial forests, an efficient method combining random forests with spatial correlation modeling, improving prediction accuracy for spatial data in environmental applications.

Contribution

The paper introduces a novel computationally efficient algorithm for spatially correlated tree-based regression, integrating kriging and non-parametric spatial smoothers.

Findings

Improved prediction accuracy over existing methods

Effective modeling of spatial correlation in regression trees

Validated on environmental measurement datasets

Abstract

We introduce random spatial forests, a method of bagging regression trees allowing for spatial correlation. Our main contribution is the development of a computationally efficient tree building algorithm which selects each split of the tree adjusting for spatial correlation. We evaluate two different approaches for estimation of random spatial forests, a pseudo-likelihood approach combining random forests with kriging and a non-parametric version for a general class of spatial smoothers. We show improved prediction accuracy of our method compared to existing two-step approaches combining random forests and kriging across a range of numerical simulations and demonstrate its performance on elemental carbon, organic carbon, silicon, and sulfur measurements across the continental United States from 2009-2010.