A regime-aware framework for runoff prediction in ungauged basins via self-supervised learning of hydrometeorological...

Runoff prediction in ungauged basins remains a fundamental challenge due to the absence of discharge observations and the limited transferability of data-driven models across heterogeneous hydrological conditions. This study proposes a regime-aware framework that models hydrological processes from the perspective of shared driver-induced regimes. Instead of relying on runoff observations, a view-based contrastive self-supervised learning strategy is employed to extract transferable features from hydrometeorological drivers. Specifically, augmented temporal-window views generated through temporal masking and feature perturbation are contrasted to learn runoff-independent representations, enabling the identification of cross-basin hydrological regimes via unsupervised clustering. Regime-specific prediction models are then constructed and adaptively combined for target basins based on regime similarity. Experimental results show that the proposed method achieves a median Nash-Sutcliffe efficiency of 0.5 across ungauged test basins, demonstrating improved prediction performance under ungauged conditions. These findings highlight the effectiveness of learning shared hydrological regimes from driver space and provide a new perspective for robust runoff prediction in ungauged basins.