Application of data mining and machine learning techniques for geohydrological datasets in South Africa

Abstract

A desktop study was conducted to research data-driven modelling techniques to classify relationships between borehole parameters and the relevant geological setting. Borehole surveying and drilling is a costly endeavour and by applying data mining and machine learning techniques to national groundwater databases and other available national datasets such as spatial data, better insight and improvements on management of groundwater resources can result. Five machine learning algorithms were tested on a consolidated dataset and their performances compared in order to establish which algorithm yielded the most accurate results. It was established that Random Forest Regression and Classification could be used to model yield, and Support Vector Regression and Random Forest Classification could model static water levels. The algorithm was tested on three case study areas, based on Vegter regions. The results indicated that static water levels could be modelled with high rates of accuracy, but yield modelling was not as successful, and a lot of uncertainty still remains as to the drivers behind water strike yield.