A Comparison of Ensemble and Deep Learning Algorithms to Model Groundwater Levels in a Data-Scarce Aquifer of Southern Africa

Machine learning and deep have demonstrated usefulness in modelling various groundwater phenomena. However, these techniques require large amounts of data to develop reliable models. In the Southern African Development Community, datasets are generally poorly developed. Hence, question arises as whether machine can be a tool support management data-scarce environments Africa. This study tests two algorithms, gradient-boosted decision tree (GBDT) long short-term memory neural network (LSTM-NN), model level (GWL) changes Shire Valley Alluvial Aquifer. Using from boreholes, Ngabu (sample size = 96) Nsanje 45), we predictive scenarios: (I) predicting change current month’s level, (II) following level. For borehole, GBDT achieved R2 scores 0.19 0.14, while LSTM 0.30 0.30, experiments I II, respectively. −0.04 −0.21, 0.03 −0.15, The results illustrate that performs better than model, especially regarding slightly greater time series extreme GWL changes. closer inspection reveals where relatively small (e.g., Nsanje), may more efficient, considering cost required tune, train, test model. Assessing full spectrum results, concluded sample sizes might not sufficient generalised