On the proper use of screen-level temperature measurements in weather forecasting models over mountains

Abstract. The near-surface air temperature, considered to be measured at about 2 m above the ground, is a key meteorological parameter with a wealth of uses for humankind. However, its accurate estimation in mountain regions is impeded by persistent limits inherent to atmospheric modeling over complex terrain. In the present study, we analyze the role of structural inhomogeneities of the valleys and mountains observational network in France to highlight their contribution to the misrepresentation of near-surface air temperature over mountain regions in the numerical weather prediction (NWP) system Arome-France. We examine in particular the effects of the disparity in height above ground of the temperature measurements, of the inhomogeneous geographical distribution of stations that are preferentially located in valleys, and of the relief mismatch between station location and model grid points. The consequences of these inhomogeneities are analyzed through their effect on model performance evaluation and on the assimilation, with a focus on the winter season. In France, high-altitude stations usually measure temperature at about 7 m over the snow-free ground and on average 1 to 2 m lower when the ground is snow-covered in winter. We show that this height difference with respect to standard stations measuring at 2 m should be considered both when evaluating the model performances and in assimilation. In terms of scores, model behaviors can be highly different at 2 m vs. 5 m so that confounding the two levels can lead to a strong mischaracterization of model biases. This confusion additionally makes the assimilation of high-altitude stations detrimental to the analysis for the Arome-France NWP system. We also show that due to the current 3DVar assimilation system, the assimilation of valley stations affects the near-surface temperature analysis at all altitudes in the mountains. On the other hand, the altitude mismatch between observation points and model grid points does not play an important role, probably in part due to its relatively marginal occurrence in an NWP system with 1.3 km grid spacing. In summary, this study describes new methods and provides guidelines for comparing models with mountain observation data in terms of both assimilation and performance assessment.