Evaluation of ERA5-Land and HARv2 Reanalysis Data at High Elevation in the Upper Dudh Koshi Basin (Everest Region, Nepal)

Abstract

We present a multisite evaluation of meteorological variables in the Everest region (Nepal) from ERA5-Land and High Asian Refined Analysis, version 2 (HARv2), reanalyses in comparison with in situ observations, using classical statistical metrics. Observation data have been collected since 2010 by seven meteorological stations located on or off glacier between 4260 and 6352 m MSL in the upper Dudh Koshi basin; 2-m air temperature, specific and relative humidities, wind speed, incoming shortwave and longwave radiations, and precipitation are considered successively. Overall, both gridded datasets are able to resolve the mesoscale atmospheric processes, with a slightly better performance for HARv2 than that for ERA5-Land, especially for wind speed. Because of the complex topography, they fail to reproduce local- to microscale processes captured at individual meteorological stations, especially for variables that have a large spatial variability such as precipitation or wind speed. Air temperature is the variable that is best captured by reanalyses, as long as an appropriate elevational gradient of air temperature above ground, spatiotemporally variable and preferentially assessed by local observations, is used to extrapolate it vertically. A cold bias is still observed but attenuated over clean-ice glaciers. The atmospheric water content is well represented by both gridded datasets even though we observe a small humid bias, slightly more important for ERA5-Land than for HARv2, and a spectacular overestimation of precipitation during the monsoon. The agreement between reanalyzed and observed shortwave and longwave incoming radiations depends on the elevation difference between the station site and the reanalysis grid cell. The seasonality of wind speed is only captured by HARv2. The two gridded datasets ERA5-Land and HARv2 are applicable for glacier mass and energy balance studies, as long as either statistical or dynamical downscaling techniques are used to resolve the scale mismatch between coarse mesoscale grids and fine-scale grids or individual sites.