A preliminary assessment of glacier melt-model parameter sensitivity and transferability in a dry subarctic environment

Abstract. Efforts to project the long-term melt of mountain glaciers and ice-caps require that melt models developed and calibrated for well studied locations be transferable over large regions. Here we assess the sensitivity and transferability of parameters within several commonly used melt models for two proximal sites in a dry subarctic environment of northwestern Canada. The models range in complexity from a classical degree-day model to a simplified energy-balance model. Parameter sensitivity is first evaluated by tuning the melt models to the output of an energy balance model forced with idealized inputs. This exercise allows us to explore parameter sensitivity both to glacier geometric attributes and surface characteristics, as well as to meteorological conditions. We then investigate the effect of model tuning with different statistics, including a weighted coefficient of determination (wR²), the Nash-Sutcliffe efficiency criterion (E), mean absolute error (MAE) and root mean squared error (RMSE). Finally we examine model parameter transferability between two neighbouring glaciers over two melt seasons using mass balance data collected in the St. Elias Mountains of the southwest Yukon. The temperature-index model parameters appear generally sensitive to glacier aspect, mean surface elevation, albedo, wind speed, mean annual temperature and temperature lapse rate. The simplified energy balance model parameters are sensitive primarily to snow albedo. Model tuning with E, MAE and RMSE produces similar, or in some cases identical, parameter values. In twelve tests of spatial and/or temporal parameter transferability, the results with the lowest RMSE values with respect to ablation stake measurements were achieved twice with a classical temperature-index (degree-day) model, three times with a temperature-index model in which the melt parameter is a function of potential radiation, and seven times with a simplified energy-balance model. A full energy-balance model produced better results than the other models in nine of twelve cases, though the tuning of this model differs from that of the others.