Machine learning for second-order nonstationary geostatistical data

Music City Center 

Machine learning methods have enjoyed a wide popularity in the last few years. While in most cases machine learning methods have been praised for their ability to capture non-linear effects of different covariates on the mean structure of the data, in recent years interest has been placed in extending these approaches to also account for spatial dependence. In this talk we investigate whether the predictions generated by these methods are still accurate when the data are a realization of a non-stationary spatial process with non-stationarity not only in the mean function but also in the second-order structure. Through simulation experiments we highlight situations in which these methods seem to perform less well, probably due to over-smoothing. We propose a solution that addresses the spatial over-smoothing and we illustrate our approach with an application in Earth System sciences.