Accidental infrastructure for groundwater monitoring in Africa.

A data deficit in shallow groundwater monitoring in Africa exists despite one million handpumps being used by 200 million people every day. Recent advances with smart handpumps have provided accelerometry data sent automatically by SMS from transmitters inserted in handles to estimate hourly water usage. Exploiting the high-frequency noise in handpump accelerometry data, we model high-rate wave forms using robust machine learning techniques sensitive to the subtle interaction between pumping action and groundwater depth. We compare three methods for representing accelerometry data (wavelets, splines, Gaussian processes) with two systems for estimating groundwater depth (support vector regression, Gaussian process regression), and apply three systems to evaluate the results (held-out periods, held-out recordings, balanced datasets). Results indicate that the method using splines and support vector regression provides the lowest overall errors. We discuss further testing and the potential of using Africa's accidental infrastructure to harmonise groundwater monitoring systems with rural water-security goals. A data deficit exists in shallow groundwater monitoring in Africa.Our smart handpump has low-cost accelerometers mounted in the handle.We show that machine learning methods applied to the accelerometry can estimate aquifer depth.We demonstrate that we can use the accidental infrastructure of handpumps for estimating groundwater levels.