Spatial prediction of borehole yield in southern Mali using machine learning classifiers

Abstract

Groundwater plays a vital role in drinking water supply, food security and ecosystem services. Approximately 2.5 billion people worldwide rely exclusively on groundwater to meet their daily needs, while hundreds of millions of farmers depend on groundwater resources to sustain their livelihoods. Groundwater potential mapping based on machine learning (ML-GPM) can be used to support groundwater exploration, planning and management practices. Most ML-GPM studies aim to predict a positive or negative outcome, that is, to identify areas of high or low groundwater potential. This work takes this conventional bivariate outcome approach one step further by predicting borehole yields and applying a multiclass approach. The method is illustrated through an application over a study area of 21,000 km2, including the administrative region of Bamako and the municipalities of Kati and Kangaba in the Koulikoro region of southern Mali. Logistic Regression, Gradient Boosting and Extra Trees classifiers were trained on an imbalanced multiclass database of 483 boreholes and 20 explanatory variables. The explanatory variables include information on lithology, geomorphology, soil, land use/land cover, topography, drainage and slope-related variables and rainfall, among others. All models returned prediction scores between 0.80 and 0.87. The most important variables include elevation, vegetation cover, basement depth and geology. The alluvial sediments of the Niger river banks, especially in the southern and northern sectors, are clearly associated with the most productive class. In contrast, the Mandingue plateau has the lowest groundwater potential. The piedmont areas present an intermediate groundwater perspective. These maps could be used to inform water supply policy at a regional scale.