An assessment of spatial random forests for environmental mapping: the case of groundwater nitrate concentration

Machine-learning models such as random forests (RF) and their spatial variants are increasingly popular in the regionalization of environmental contaminants. To date, no empirical comparison of spatial RF variants is available in this field. The purpose of this study is to evaluate six spatial RF variants, benchmarking them against universal kriging (UK) and multiple linear regression (MLR). We empirically examine predictive performances over different prediction distances using the regionalization of nitrate concentrations in groundwater as a case study. Differences among spatial RF variants were generally small. Over prediction distances shorter than the practical range of autocorrelation, spatial variants tended to achieve higher precisions than non-spatial RF and MLR. RF-OOB-OK that uses ordinary kriging predictions based on the out-of-bag error appeared as one of the more consistently well-performing methods. Computationally tractable spatial RF variants can be considered viable alternatives to geostatistical regionalization methods in making spatial predictions of environmental contaminants. • Over short prediction distances spatial RF variants tend to outperform RF • Differences between spatial RF variants are usually small • Computational complexity can be a limiting factor • Leave-one-out Ordinary-Kriging predictions are a beneficial spatial covariate in RF