Deep learning error post-processing improves stochastic watershed modeling

Abstract

Hydrologic extremes, including floods and droughts, pose substantial societal risks that are expected to intensify with climate change. Deterministic watershed models (DWMs) remain a mainstay for modeling these extremes, but lack explicit representation of uncertainty, limiting their utility for risk-informed planning. Stochastic watershed models (SWMs) address this limitation by generating ensembles of streamflow via models of observed DWM residuals. However, most SWMs struggle with the complex dependence between DWM residuals and the underlying hydrologic state, which can complicate stochastic simulations under nonstationary climates. Deep learning (DL) models, whether used as standalone models or post-processors for process-based DWMs, offer a pathway to address this challenge by reducing conditional dependence. In this study, we evaluate SWMs applied to seven models: three process-based models (PRMS, Hymod, and HBV), their hybrid process-DL counterparts, and a pure DL DWM, focusing on daily simulations and extremes under both historical conditions and synthetic climate change scenarios. Results for a case study watershed in Massachusetts show that SWMs applied to hybrid or pure DL DWMs consistently outperform those applied to process-based DWMs. However, an SWM applied to the pure DL model exhibits weaknesses at low flows for this study basin, underscoring the value of hybrid approaches. Extending the analysis across 73 additional basins demonstrates that these improvements are robust and generalizable statewide. This work highlights the potential of a DL-enhanced stochastic watershed modeling framework to advance hydrologic risk prediction under changing climate conditions, offering a scalable methodology for integrating uncertainty into watershed modeling for long-term planning.