Sensitivity of snow models to the accuracy of meteorological forcings in mountain environments

Abstract

Abstract. Snow models are usually evaluated at sites providing high-quality meteorological data, so that the uncertainty in the meteorological input data can be neglected when assessing model performances. However, high-quality input data are rarely available in mountain areas and, in practical applications, the meteorological forcing used to drive snow models is typically derived from spatial interpolation of the available in situ data or from reanalyses, whose accuracy can be considerably lower. In order to fully characterize the performances of a snow model, the model sensitivity to errors in the input data should be quantified. In this study we test the ability of six snow models to reproduce snow water equivalent, snow density and snow depth when they are forced by meteorological input data with gradually lower accuracy. The SNOWPACK, GEOTOP, HTESSEL, UTOPIA, SMASH and S3M snow models are forced, first, with high-quality measurements performed at the experimental site of Torgnon, located at 2160 m a.s.l. in the Italian Alps (control run). Then, the models are forced by data at gradually lower temporal and/or spatial resolution, obtained by (i) sampling the original Torgnon 30 min time series at 3, 6, and 12 h, (ii) spatially interpolating neighbouring in situ station measurements and (iii) extracting information from GLDAS, ERA5 and ERA-Interim reanalyses at the grid point closest to the Torgnon site. Since the selected models are characterized by different degrees of complexity, from highly sophisticated multi-layer snow models to simple, empirical, single-layer snow schemes, we also discuss the results of these experiments in relation to the model complexity. The results show that, when forced by accurate 30 min resolution weather station data, the single-layer, intermediate-complexity snow models HTESSEL and UTOPIA provide similar skills to the more sophisticated multi-layer model SNOWPACK, and these three models show better agreement with observations and more robust performances over different seasons compared to the lower-complexity models SMASH and S3M. All models forced by 3-hourly data provide similar skills to the control run, while the use of 6- and 12-hourly temporal resolution forcings may lead to a reduction in model performances if the incoming shortwave radiation is not properly represented. The SMASH model generally shows low sensitivity to the temporal degradation of the input data. Spatially interpolated data from neighbouring stations and reanalyses are found to be adequate forcings, provided that temperature and precipitation variables are not affected by large biases over the considered period. However, a simple bias-adjustment technique applied to ERA-Interim temperatures allowed all models to achieve similar performances to the control run. Regardless of their complexity, all models show weaknesses in the representation of the snow density.