Neural network and geostatistical interpolation of observed near-surface atmospheric variables for snow cover numerical model simulation

The snowpack plays a fundamental role in regulating the global climate thanks to its high albedo and thermal insulation properties, and for many regions of the world it also has very local and important impacts. Indeed, the snow is an important water reservoir, storing the water in solid state during cold months, and releasing it in liquid state during warmer months. But the snow is also the necessary condition for the development of rural places which base their economy on winter sports. However, a certain risk is always associated with snow when it deposits on the ground, since the snow can slide down, creating avalanches which may cause several damages to the local flora, fauna, buildings and infrastructures. Typically, the conditions that allow the occurrence of snow avalanches span from the point scale to the slope scale, and depend on the snowpack properties. Kilometer-resolution numerical models are not able to reproduce the slope-scale variability of the snowpack properties because of the complex interaction between the atmospheric flows and the topography at finer scale, nevertheless running atmospheric models at sub-kilometer resolution is highly computational demanding. To face this limitation, sub-kilometer gridded weather data can also be generated using interpolation and downscaling methods of in situ or remote observations, which require less computational resources compared to numerical weather models. Thus, we develop and validate several algorithms based on geostatistical an machine learning techniques to produce a sub-kilometer gridded dataset of the most important near-surface atmospheric variables. Finally, we use the generated dataset to force a snow cover model and we evaluate the model skills in reproducing the observed snowpack properties.