Annual Streamflow Modelling Using Large-Sample Datasets: Insights from Hybrid Models and Seasonal Synchronicity

Large-sample datasets of catchments offer opportunities to explore the hydroclimatic and physiographic controls on hydrological processes across various spatial and temporal contexts. This study leverages a global dataset of over 4,000 catchments to investigate how annual precipitation, potential evapotranspiration, and seasonal synchronicity influence streamflow dynamics. Seasonal synchronicity, reflecting the temporal alignment of precipitation and evapotranspiration, is identified as a key factor in shaping hydrological responses and improving our understanding of inter-annual variability.

We use a hybrid-modelling framework where a dense neural network, trained on catchment descriptors, is employed to parameterize a simple annual hydrological model. The hydrological model is characterized by three easily interpretable coefficients, each representing the sensitivity of annual streamflow to precipitation, evapotranspiration, and their synchronicity. By systematically evaluating regionalization across spatial, temporal, and spatiotemporal contexts, we demonstrate the potential for transferring insights and functional understanding from data-rich to data-scarce catchments.

This work contributes to advancing hydrological synthesis by linking catchment descriptors with dominant hydrological controls and exploring the representativeness of global catchment datasets. Our findings underline the importance of harmonized large-sample datasets and systematic workflows for uncovering annual hydrological processes and enabling robust predictions in ungauged basins.