Neural network and data-driven correction methods for improving the simulation of alpine snowpack

This study investigates on the internal structure of snowpack, metamorphism, and their evolution during the winter season using numerical models. A customized snowpack forecast model is used for this scope, enabling to compare simulation's results obtained with different initialization choices and to evaluate methods for improving the accuracy of the snowpack simulation. Specifically, meteorological data from automatic weather stations (AWS), reanalysis and outputs from numerical weather models (NWMs) are respectively used to simulate snowpack characteristics as an input to the SNOWPACK model, a one-dimensional finite element model developed for avalanche forecasting and snow behavior studies. The model is validated using a specially built statistical verification method to assess its performance. In particular, it is evaluated how various meteorological inputs influence snowpack simulations and to identify the best dataset for accurate snowpack representation. Special emphasis is placed on bias correction techniques and the integration of neural networks to enhance model performance. Results indicate that neural network approaches, better than bias-corrected meteorological data, offer improved predictions of snow height and internal snow structure when compared to traditional NWM data. The promising results of this study support new and encouraging possibilities for understanding the processes related to the snow and improving their predictability over the winter season. Although applied to a specific case study, both bias correction and neural networks methods are general enough to be applied to other location of interest for the snow science. Nonetheless, great improvements in the accuracy of both the snowpack internal structure and the snow quantity are produced by these methods. A similar approach can be used over forecast data to improve the simulations of the snowpack conditions to come.