Process guided graph-based transformer learning for streamflow predictions in data-sparse river basins

Abstract

Streamflow prediction in data-sparse regions has traditionally relied on process-based hydrological models (PBMs) or hybrid approaches that combine physical modeling with data-driven learning. However, these methods typically require extensive calibration using local observations and detailed knowledge of internal model parameters, which limits their transferability and interpretability when applied to ungauged basins. To address these challenges, we introduce a novel Process-Guided Graph–Transformer (PGT) framework that surrogates PBMs and enhances streamflow prediction in ungauged basins (PUB) using uncalibrated PBM parameters without explicit calibration or optimization. The proposed framework integrates uncalibrated PBM outputs, river network topology, Graph Neural Networks, and Transformers for spatio-temporal learning, and is evaluated across three hydrologically contrasting river basins in India, Australia, and the United States. Results show that PGT consistently improves predictive skill, increasing median KGE from −0.5 to 0.4 for baseline models (SWAT and SWAT-SUFI2) to 0.50–0.75, while maintaining PBIAS within±10%, compared with baseline overprediction frequently exceeding 50%. PUB generalization experiments further demonstrate robust skill transfer from gauged to ungauged subbasins, with ΔKGE remaining near 0.0±0.2 across hydrological distances of 0–300km. Additional analyses indicate improved extreme-event detection (Critical Success Index ≈ 0.8–0.9 compared to 0.4–0.6 for baselines), reliable uncertainty characterization (p-factor: 0.5–1.0 compared to 0.25–0.6 for baselines), and an 8–40×reduction in computational cost relative to traditional calibration-based approach (SWAT-SUFI2). Overall, these results demonstrate that PGT provides a scalable and process-consistent solution for multi-site streamflow prediction in data-sparse and ungauged regions. © 2026 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.