This paper presents a novel data fusion technique for improving the snow cover monitoring for a mesoscale Alpine region, in particular in those areas where two information sources disagree. The presented methodological innovation consists in the integration of remote-sensing data products and the numerical simulation results by means of a machine learning classifier (support vector machine), capable to extract information from their quality measures. This differs from the existing approaches where remote sensing is only used for model tuning or data assimilation. The technique has been tested to generate a time series of about 1300 snow maps for the period between October 2012 and July 2016. The results show an average agreement between the fused product and the reference ground data of 96%, compared to 90% of the moderate-resolution imaging spectroradiometer (MODIS) data product and 92% of the numerical model simulation. Moreover, one of the most important results is observed from the analysis of snow cover area (SCA) time series, where the fused product seems to overcome the well-known underestimation of snow in forest of the MODIS product, by accurately reproducing the SCA peaks of winter season.

This paper presents a new concept to derive the snow water equivalent (SWE) based on the joint use of snow model (AMUNDSEN) simulation, ground data, and auxiliary products derived from remote sensing. The main objective is to characterize the spatial-temporal distribution of the model-derived SWE deviation with respect to the real SWE values derived from ground measurements. This deviation is due to the intrinsic uncertainty of any theoretical model, related to the approximations in the analytical formulation. The method, based on the k-NN algorithm, computes the deviation for some labeled samples, i.e., samples for which ground measurements are available, in order to characterize and model the deviations associated to unlabeled samples (no ground measurements available), by assuming that the deviations of samples vary depending on the location within the feature space. Obtained results indicate an improved performance with respect to AMUNDSEN model, by decreasing the RMSE and the MAE with ground data, on average, from 154 to 75 mm and from 99 to 45 mm, respectively. Furthermore, the slope of regression line between estimated SWE and ground reference samples reaches 0.9 from 0.6 of AMUNDSEN simulations, by reducing the data spread and the number of outliers.