Author(s): Mohammad Zeynoddin; Tadros Ghobrial; Hossein Bonakdari; Gregory Langston; Jean-Pascal Faubert
Linked Author(s): Tadros Ghobrial, Hossein Bonakdari
Keywords: River ice; Backwater factor; ERA5; Reanalysis; Ice-affected discharge
Abstract: Ice cover alters open-channel hydraulics by introducing temporally variable resistance, disrupting the monotonic stage-discharge relationship and invalidating open-water rating curves during winter. Hydrometric agencies address this through a daily backwater factor (dbf), defined as the ratio of under-ice discharge to its open-water equivalent. In practice, dbf estimation relies on sparse winter measurements and subjective interpolation, limiting reproducibility and scalability. This study develops a structured, data-driven framework for continuous dbf estimation. Seasonal dbf behaviour is quantified at six Canadian stations, and predictive models are developed using hydrometric observations and ERA5 reanalysis drivers. Three strategies are compared: a LOESS benchmark, site-specific Extreme Learning Machines (ELM), and a common-key ELM with uniform structure across basins. LOESS exhibits weak predictive skill, with a mean test correlation of 0.16 and limited error reduction, particularly at data-sparse Quebec sites, and fair estimates of 60% correlation at other sites with more data. Site-specific ELM configurations increase test correlation to approximately 0.90, with error reductions of 22-93% relative to LOESS. The common-key ELM captures broad seasonal structure but shows reduced stability and lower correlation at a few sites. These findings demonstrate that reanalysis-driven machine learning can recover robust winter discharge patterns, while highlighting remaining limits to cross-site transfer without additional hydraulic descriptors.
Year: 2026