Predicting Groundwater Depth Using Handpump Accelerometry Data

Safe and reliable access to groundwater is a necessity for people in rural Africa. Many rely on handpumps, for which there is often no system of maintenance in place, leading to around one third being out of use at any given time. Many places are experiencing increased demand for groundwater, leading to the need for a system to monitor groundwater levels. This paper presents a novel approach to monitoring ground water levels using the “accidental infrastructure” of handpumps. We place accelerometers on the handles of handpumps and record tri-axial data when the pump is in use. The accelerometry data represents both the gross movement and the vibrations of the handle. We show how this data can be used to predict the depth of the water at the pump. We create feature vectors using wavelet-, smoothing spline-, and heteroscedatic Gaussian process-based models of the accelerometery data, and subsequently perform regression using support vector regression or Gaussian process regression. We find that we can obtain low errors low errors of groundwater depth, but that there are biases among the predictions. In addition, we investigate the effect of the tide on groundwater levels and show how it can be modelled with linear regression.