Efficient multi-objective calibration and uncertainty analysis of distributed snow simulations in rugged alpine terrain

Abstract. In steep and complex mountainous terrain, robust simulations of snow accumulation and ablation are crucial to a wide range of applications, especially those related to hydrology and ecology. Whilst new opportunities exist to integrate high-resolution spatio-temporal observations in the estimation of uncertain parameters in (a.k.a. “calibration” of) sophisticated, process-rich snow models, they have not yet been fully exploited. Here, with a view towards improving representations of snow and ultimately meltwater dynamics in rugged topography, a novel approach to the calibration of a high-resolution energy balance-based snow model that additionally accounts for gravitational snow redistribution is presented. Several important but uncertain parameters are estimated using an efficient, gradient-based method with respect to two complementary types of snow observations – snow extent maps derived from Landsat 8 images, and snow water equivalent (SWE) time-series reconstructed at two contrasting locations. When assessed on a per-pixel basis over 17 days that together encompass practically the full range of possible snow cover conditions, snow patterns were reproduced with a mean accuracy of 85 %. The spatial performance metrics obtained compare favourably with those previously reported, whilst the temporal evolution of SWE at the stations was also satisfactorily simulated. Uncertainty and data worth analyses revealed that: i) the propensity for model predictions to be erroneous was substantially reduced by the calibration process, ii) pre-calibration uncertainty was largely associated with two parameters that were introduced to modify the longwave component of the energy balance, but this uncertainty was greatly diminished by calibration, and iii) the lower elevation SWE time-series was particularly valuable despite the comparatively small number of observations at this site. Alongside a gridded snowmelt dataset, commensurate estimates of firn melt, ice melt, liquid precipitation, and potential evapotranspiration were also produced. Our study demonstrates the growing potential of combining observation technologies and state-of-the-art inverse approaches to both constrain and quantify the uncertainty associated with simulations of alpine snow dynamics.