When the past matters: How model initialization can lead to surprises in long-term simulations in glaciated environments

Hydrological simulations are crucial for understanding water resource changes and developing climate change adaptation strategies. Recent modeling efforts go towards transient, fully distributed simulations to shed light on the long-term evolution of hydrological systems rather than just a glimpse into future streamflow. This development comes, however, with a hitherto understudied challenge: the role of model initialization and evolution of model topography on distributed long-term simulation results. Here, based on an example from a high alpine catchment in Austria, we emphasize the importance of proper surface and subsurface water storage initialization or update (e.g. glacier area and volume, soil moisture content). We show that model initialization and the inability of the model to deal with landscape evolution could cause significant errors in streamflow simulation. Due to initial-condition dependence and difficulty in external state-updating, external coupling between distributed models (e.g. between a glacier model and a hydrological model) can be inadequate for non-stationary systems. Based on our results, we underline the importance of long (>100 years) exploratory simulations run in various slice lengths to identify potential modeling errors arising from model initialization. Most importantly, the hydrological models should be run to steady state, similar to hydrogeological modeling practice.