Divergent impacts of land-use change on high and low river flow revealed by explainable machine learning

Quantifying impacts of land-use change on streamflow extremes is challenging, primarily due to the masking effects of other environmental processes. Our current understanding of these impacts on streamflow extremes remains incomplete. Here, we use explainable machine learning techniques to analyse over 1.5 million seasonal 7-day low-flow and high-flow events across 10,717 catchments worldwide between 1982 and 2023. Our model incorporates antecedent meteorological conditions, annual change of six land-use categories, and catchment characteristics (hydrogeological, anthropogenic, and topographic) as explanatory variables. The Shapley additive explanations technique is employed to quantify the contributions of the predictors to low and high flows. Our results indicate that all categories of land-use change exert a greater influence on high flows compared to low flows, although the overall contribution of land-use change to streamflow extremes is far smaller (< 2%) than that of antecedent meteorological conditions (32%–48%) and hydrologic signatures (35%–52%). Contrary to previous studies, our results indicate that land-use impacts are largely independent of catchment size. Notably, urbanization exhibits diverging effects on low flows: enhancing them in arid regions, reducing them in tropical regions, and minimally impacting them in temperate regions. Urbanization nearly always amplifies high flows, except in minimally urbanised catchments of arid regions. Areas with higher forest cover consistently have smaller low flows across all climate zones, and high flows appear generally insensitive to afforestation. Low flows generally are insensitive to cropland expansion but areas with more cropland typically have smaller high flows.