Study region The Southern African region consists of ten countries including, Angola, Zambia, Malawi, Namibia, Botswana, Zimbabwe, Mozambique, South Africa, Lesotho, and Swaziland. Frequent drought episodes in the arid and semi-arid parts of the region suggest the impact of climate change and the need for a sustainable framework for groundwater level assessments.

Study focus In this study, we developed a machine learning modelling framework based on the deep belief network (DBN) to predict the changes in monthly groundwater levels (CGWLs) at 1–5 month time scales for 27 groundwater wells over the southern Africa region. With a predictor dataset constituted by hydrological parameters, groundwater level estimates, and global climate indices, we ascertain the possibility of forecasting changes in groundwater levels (CGWL) up to 5-month lead times at most locations in the study region. Using the quantile regression technique, the strength of the DBN network at 90% and 95 % confidence levels were tested, and this helped to determine the accuracy of our forecasts at five months lead times over the four representative wells.

New hydrological insights for the region Deep learning offers new capabilities in evaluating non-linear hydrological systems, including groundwater analysis. The re-injection procedure of the DBN network, which allows the prior CGWL estimates to serve as inputs for the next estimate was key in maintaining the forecast accuracy in the 4th and 5th month lead times. This was evidenced in the accuracy of the four representative wells (r = 0.91, 0.87, 0.81, 0.75), which formed part of the test samples in CGWL analysis. We also observed that the DBN is highly susceptible to local climate variables and global climate indices, which are important drivers of climate change, and can have strong impact on groundwater level fluctuations. Therefore, the DBN proved to be a robust algorithm in our general assessments of groundwater level fluctuations.