**4. Discussion**

The first part of the results evidenced that automated solutions provide FSC estimations compatible to the supervised solutions available in literature. The major advantage of automated methods consists in the reduction of time consumption and, consequently, in the opportunity of processing long time series of terrestrial images. We described an automated approach based on the concept of Spectral Similarity [23], which could prevent artifacts under particular illuminating conditions. While a small training dataset supported the training of an SS-based algorithm, the ten-year dataset, with about 8000 images, showed a better performance compared to a state-of-the-art automated method BT described by [26]. The trend of FSC underestimation (about 10%) outlined by the small training dataset was confirmed by the large decadal comparison. The observed statistically significant differences were limited in terms of pixel number (less than 1%), but these discrepancies were important in terms of surface. The projection of each pixel on the surface could increase consistently from closer to faraway pixels and from this perspective; the impact of omissions and false discoveries on the projected area could be high. Furthermore, the ability to analyze the "difficult" conditions affecting the BT performance [10] was confirmed by statistically significant differences detected between the

two data series. The limitations of BT retrievals can be associated with poor illuminating conditions (low Sun elevation or heavy cloud coverage) and surface roughness. While low Sun elevation can occur in the early morning or in the late afternoon, surface roughness and cloud coverage are not time dependent. Furthermore, while the illuminating conditions can alter the reflective behavior of snow in response to a more blueish incident light, the roughness can imply the presence of shadowed surfaces that BT cannot discriminate compared to SS. While BT tends to separate shadowed and illuminated areas, SS can be trained to integrate both types since the spectral angle is similar and its only variation is the spectral distance. While BT can generally provide good results between 11:00 am and 3:00 pm local time, SS can enlarge the range of performing conditions in terms of both Sun elevation and cloud cover. These preliminary results concerning the SS approach represent a first step towards the development of a machine learning strategy aimed to analyze routinely ground-based images. Artifacts associated with purely-BT classification [19,26,28], which are well documented in literature [19,27,28], were reduced and the need to consider all the information present in a RGB composite image [27] was followed. Differently from [27], which combined principal component analysis to BT, SS is independent from BT and considers all the bands at the beginning of the classification step obtaining a discrimination between surface types based directly on the spectral behavior of each classified feature. Furthermore, SS considers the color variations induced by illumination conditions and the probability to separate different surface types is associated with statistical measurements such as the Mahalanobis distance.

Finally, the FSC estimated by terrestrial photography and satellite products evidenced different aspects to be considered: the spatial resolution and the cloud screening. The cloud screening is a critical step present in all of the data chain considered in this study. Our data demonstrated, in fact, that a large number of satellite omissions were associated with a wrong detection of clouds. In addition to those exclusions, different situations evidenced an underestimation of FSC affected by the presence of cloud shadows that reduce the reflection of light from the surface. Although the different data chains [6,8,13] of course, consider these anomalies, the contribution of terrestrial photography, in this case, could support for the validation of remotely sensed retrievals. Moving to the spatial resolution, we considered data ranging from a 30 m resolution (Landsat), to 500 m (MODIS), to 1 km (GlobSnow SE) in order to test different data chains with different spatial and time resolutions. The spatial resolution had, of course, an impact and we found a more reliable relation with Landsat data than with those characterized by a coarser resolution. While the projected area of the camera view is five times the surface covered by a single Landsat pixel, it represents the 2% of a MODIS pixel the 0.5% of a single GlobSnow grid element. This implies that the surface morphology can affect the final estimates due to the presence of hills and small valleys.

This framework outlines the potentiality of facilities where different satellite snow products can have a common term of comparison such as terrestrial cameras. Ground-based images represent a good proxy, useful for validating the coherence between different products. On the one hand, this data-source can support the reconstruction of long time series useful for climate change studies. On the other one, this kind of proxy can assist the definition of site-specific relation between FSC and the optical behavior of the surface.
