Bridging the Gap: How Watershed Discretization Scale Affects Flood Forecasting Accuracy

Accurate flood forecasting is critical for mitigating risks and safeguarding infrastructure and communities. Usually, we perform these forecasts by using distributed hydrological models, often calibrated at gauged watersheds and extrapolated to ungauged regions. However, depending on the model discretization, forecasts may suffer from significant performance degradation due to inadequate spatial discretization scales (DS), particularly in representing river networks. This study investigates the effects of discretization on watershed geomorphology and hydrological simulations using the Hillslope Link Model (HLM) applied to the Smooky Hills watershed in Kansas, U.S. We analyzed six DS ranging between 0.1 (benchmark DS-BDS and closer to observable network) and 70km2 (closer to USGS HUCs 12 and Hydrosheds). We assessed changes in geomorphological features such as width functions, saturated hydraulic conductivity, slope distributions, and hillslope travel times. We forced a constant runoff HLM formulation with 100 uniform rainfall patterns obtained from MRMS observations for the hydrological simulations. We compared our results at 400 control points distributed along the river network, covering scales between 1 and 50,000 km2 (outlet). The simulations include a version in which all the DS share the same routing parameters and another in which we select the routing parameters with the best performance at each control point and event. Our results show a loss of geomorphological and topological information as we use coarser DSs.Features such as the estimated hillslope travel times and the width function exhibit significant changes compared to the 0.1 km2 DS. Hydrological simulations revealed that coarser DSs result in decreased peak flows and delayed times to peak, highlighting the sensitivity of model performance to spatial scale. Parameter calibration further demonstrated that optimal model parameters vary across discretization scales and locations within the watershed, underscoring the limitations of universal calibration strategies. Our results also suggest a connection between the loss of geomorphological features and the simulations that could eventually be used to explain and overcome the limitations due to inadequate DSs. However, we are still unable to generalize this connection. These findings emphasize the importance of scale-sensitive modeling approaches and caution against the indiscriminate application of coarse discretization in flood forecasting for ungauged basins. By addressing the impacts of spatial resolution on predictive accuracy, this work contributes to advancing process representation in hydrological modeling.