Groundwater Recharge Variability in a Semi-Arid South African Catchment under Climate Change: Insights from Long-Term Observations and Machine Learning

In semi-arid regions characterized by little and erratic precipitation and ephemeral river flow, groundwater is commonly the only perennial source of freshwater sustaining ecosystems and freshwater withdrawals for agricultural, domestic and industrial uses. However, the renewability of groundwater in these regions is associated with substantial uncertainty. Focused recharge, groundwater replenishment via seepage from surface drainage during high river flow, has been shown to contribute substantially to groundwater storage. Yet, the relative contributions of focused and diffuse recharge, as well as their dependence on rainfall variability and climate change, remain underexplored at catchment scale. This study employs a data-driven approach to estimate annual groundwater recharge in the semi-arid Hout/Sand catchment (7,722 km²), Limpopo, South Africa, utilizing data from 105 boreholes spanning 1955-2023. The Water Table Fluctuation method is used to derive annual recharge estimates from individual groundwater hydrographs. The recharge estimates are used to train a Light Gradient-Boosting Machine (LightGBM) model employing physiographic and climatic predictors, which generates a fully distributed annual recharge map at a 100 m resolution for each year of the 69-year study period. The results show recharge rates exceeding 1,000 mm/year in wells near riverbeds, highlighting the dominance of focused recharge. Annual recharge maps demonstrate significant spatial variability, with high recharge values concentrated along river networks. Among the predictors in the LightGBM model, proximity to rivers emerged as the most critical factor. Total annual recharge exhibits strong inter-annual variability, closely correlated with total annual rainfall. However, preliminary findings indicate a decline in annual recharge after 2015 despite increasing annual rainfall, suggesting a decoupling of the recharge-rainfall relationship. A key limitation of the study is the bias introduced by the high concentration of wells near riverbeds characterized by high recharge rates. To address this, we aim to incorporate synthetic data points representing diffuse recharge into the model training. Focused recharge may provide a buffering effect against climate change, as more intense rainfall events could enhance recharge along the river networks. Future work will focus on quantifying the relative contribution of focused recharge to total recharge at the catchment scale, its temporal evolution, and its correlation with rainfall variability to assess the impact of climate change on groundwater recharge.