Calibration Strategies for Improved Hydrological Predictions inLake-dominated River Systems

Efficiently and accurately calibrating and modelling watersheds with hundreds of lakes and mixed observational data remains a key challenge for developing reliable forecasting tools. This study assesses a range of calibration strategies for the routing of water through lakes and rivers in the Southeastern region of Ontario. The routing model, implemented within the Raven Hydrological Modelling Framework, is forced with runoff from Environment and Climate Change Canada's GEM-Hydro model and incorporates over 40 streamflow and lake level gauges screened for anthropogenic influence. We compared several global and local calibration strategies, optimizing routing model parameters including bias correction, baseflow parameters, routing coefficients, and lake crest widths. In the global calibrations, we applied a single multi-metric objective function that incorporated discharge and lake water levels, whereas the parameters were calibrated locally for gauge-specific sub-regions in local calibration. In local calibration of the model, the median Kling Gupta Efficiency (KGE) of the river discharge was improved from 0.37 to 0.73 compared to the baseline experiment. Moreover, the local calibration improved discharge predictions at 88.6% of gauges during the calibration period and 67.7% during validation compared to global calibration. Median KGE increased from 0.62 under global calibration to 0.78 with the local approach. Lake level simulations also benefit from individual crest width calibration, with KGE deviation without accounting for bias improving from 0.72 to 0.88. We also found that spatially distributed bias correction of runoff inputs enhances discharge accuracy, with improvements showing some correlation with underlying geology. These findings emphasize the value of localized calibration for capturing site-specific hydrological dynamics and recommend locally updating lake crest width estimates to improve lake parameterization in semi-distributed models.