Predicting annual base flow index and its trends using a large sample of dataset across the globe

The Base-Flow Indices (BFI), indicating surface and groundwater interaction, play a significant role in the hydrological cycle. The values vary widely across the globe, and are expected to change in the context of a changing climate. Predicting global “natural” BFI is challenging due to limited observations reflecting natural impacts. This study aims to predict annual BFI and their trends for a compiled dataset of annual streamflow and meteorological data covering the period 1982-2020 for more than 2250 small unregulated catchments worldwide. BFI were derived using three digital filtering methods, resulting in a trend of -0.0009 ± 0.01 per decade over the last four decades. To predict annual BFI and their trends in ungauged catchments, a Random Forest Regression approach was employed, incorporating static attributes and meteorological time series data as model inputs. Five-fold cross-validation demonstrated the effectiveness of the Random Forest Regression in predicting both BFI and their trends. The Quantile Regression Forest method was utilized to quantify the uncertainty, achieving a relatively low range for both BFI and their trends. Soil conditions and maximum temperature emerged as the most crucial variables for predicting BFI, while temperature-related variables also proved essential for predicting the BFI trends. The goal is to extend the understanding of "natural" BFI and their trends in ungauged catchments, as the Random Forest Regression model was trained under unregulated conditions. This study offers the possibility to predict "natural" BFI and their trendsacross the globe. This could support water authorities in managing water resources, particularly concerning base flow.