Mapping groundwater potential through an ensemble of big data methods.

Groundwater resources are crucial to safe drinking supplies in sub Saharan Africa, and will be increasingly relied upon in a context of climate change. The need to better understand groundwater calls for innovative approaches to make the best out of the existing information. A methodology to map groundwater potential based on an ensemble of machine learning classifiers is presented. A large borehole database (n = 1848) was integrated into a GIS environment and used to train, validate and test twelve machine learning algorithms. Each classifier predicts a binary target (positive or negative borehole) based on the minimum flow rate required for communal domestic supplies. Classification is based on a number of explanatory variables, including landforms, lineaments, soil, vegetation, geology and slope, among others. Correlations between the target and explanatory variables were then generalized to develop groundwater potential maps. Most algorithms attained success rates between 80% and 96% in terms of test score, which suggests that the outcomes provide an accurate picture of field conditions. Statistical learners were observed to perform better than most other algorithms, excepting random forests and support vector machines. Furthermore, it is concluded that the ensemble approach provides added value by incorporating a measure of uncertainty to the results. This technique may be used to rapidly map groundwater potential for rural supply or humanitarian emergencies in areas where there is sufficient historical data but where comprehensive field work is unfeasible.