Mapping fecal pollution in rural groundwater supplies by means of artificial intelligence classifiers

Abstract

Fecal contamination poses a threat to groundwater supplies in low income regions. This is often due to the coexistence of pit latrines with domestic wells in densely populated areas. In this context, developing alternative methodologies to map fecal pollution in shallow wells is needed. A thorough survey of over 240 domestic wells and 570 pit latrines was conducted in a rural town of southern Mali. Water samples were collected from all wells and tested for temperature, pH, electric conductivity, total dissolved solids, turbidity and thermotolerant coliforms. The outcomes of the field survey were incorporated into a GIS database. Thirteen machine leaning classifiers, including different statistical algorithms, instance-based learners and tree-based models, were used to determine the spatial distribution of fecal pollution as per five explanatory variables (latrine density, distance to the closest latrine, borehole yield, water table depth and population density). The best performing classifiers, selected on test scores, were then used to develop predictive maps. Random forest and logistic regression rendered prediction scores for fecal pollution in excess of 0.90. Multilayer perceptrons, support vector machines and quadratic discriminant analyses also proved adept at forecasting fecal pollution. Ensemble mapping shows that 30–50 m buffers around domestic wells may be sufficient to prevent contamination of domestic supplies in most instances. This demonstrates that machine learning may provide a versatile methodological alternative to traditional Darcian approaches. On the other hand, the practical difficulties involved in maintaining wellhead protection areas suggests the need to implement piped water supplies.