2026 Volume 7 Issue 3 Pages 495-503
In river water-level prediction, this study focuses on the heterogeneity of input structure, in which the presence and number of upstream water-level gauges differ among forecast sites. Under this condition, we examine whether a fine-tuned time-series foundation model, Chronos-2, can transfer to ungauged basins. The training data consist of 7,360 flood-event series extracted from 112 water-level gauges across 10 river basins in Japan, and observed rainfall is used as perfect forecast data for future time steps. The models are evaluated by Leave-One-Basin-Out cross-validation and compared with an LSTM that incorporates static catchment attributes. Chronos-2 attains a mean NSE of about 0.8 across the 12 evaluation sites at a six-hour lead time even for ungauged basins, substantially outperforming the LSTM. Adding upstream water levels improves the accuracy of the foundation model but degrades that of the LSTM. These results suggest that the characteristics of the foundation model, including large-scale pre-training and its architecture, contribute to transfer to ungauged basins even for observation networks with heterogeneous input structure.