Flood Forecasting in Data-Scarce Basins Using Event-Based Records and Multi-Paradigm Modeling with Similarity-Based Transfer

Accurate flood forecasting in ungauged or data-scarce basins remains a persistent challenge in hydrology, particularly in regions like China where many basins offer only intermittent flood records, often limited to short, hourly datasets collected during individual events. Traditional physically based models, constrained by their reliance on basin-specific calibration and validation, struggle to scale across diverse catchments with limited observations. While deep learning presents a significant opportunity by integrating multi-basin data to uncover latent hydrological patterns, its application in regions reliant on fragmented, event-based records still faces limitations. To address these challenges, this study systematically evaluates the predictive accuracy and transferability of three representative modeling paradigms: a process-based conceptual model, a purely data-driven model, and a hybrid physics-guided differentiable model. Using a dataset of 612 historical flood events across 16 basins in Anhui Province, China, we assess model performance under both local (per-basin) and regional (multi-basin) training schemes. To overcome the limitations of direct model transfer in data-scarce contexts, we explore parameter generalization strategies, including PCA-based regionalization, direct transfer of pre-trained models, and similarity-informed transfer learning with limited local fine-tuning. A key objective is to examine whether similarity-based donor basin selection is essential, or if universal models can be applied across hydrologically consistent basins. Results show that regional models consistently outperform per-basin models, with enhanced generalization and robustness. Among the approaches, the physics-guided differentiable model outperforms the purely data-driven model, particularly in capturing flood peaks under short-duration, event-based data. Moreover, in regions with pronounced inter-basin heterogeneity, models trained on fewer events from the most similar basin may outperform those trained on larger, aggregated datasets from less similar basins. These findings underscore the dual importance of physical constraints and inter-basin similarity in flood modeling.