Gaussian process machine learning and Kriging for groundwater salinity interpolation

Abstract

Gaussian processes (GPs) provide statistically optimal predictions in the sense of unbiasedness and maximal precision. Although the modern implementation of GPs as a machine learning technique is more capable and flexible than Kriging, their employment in environmental science is less routine. Their flexibility and capability as a spatial data interpolation technique are demonstrated by applying them to groundwater salinity prediction in a data-sparse region in Australia. By learning from multiple data sources, including AEM and DEM data, GPs have generated groundwater salinity maps with rich local details and quantified uncertainty to support risk-based decision making. The results demonstrate the great worth of nonpoint data with regional spatial coverage to provide more realistic heterogeneity in aquifer properties that are critical for many studies such as contaminant transport. GPs should be further encouraged in groundwater science for data interpolation and prediction, especially when point measurements are sparse and multiple predictors are available.