**1. Introduction**

Snow cover is an important component of the cryosphere that plays a key role for climate dynamics and the resources availability: the seasonality of the snow cover influences, in fact, weather patterns, hydropower generation, agriculture, forestry, tourism, and aquatic ecosystems [1–3]. Remote sensing is the most common tool for the routine estimation of the snow cover extent. However, two different aspects must be considered for the enhancement of the final output: time and spatial resolutions. Both components, using remotely sensed data, are connected to each other, since the higher the spatial resolution (below hundreds of meters), the lower the revisit time interval (more than 1 week) [4].

The state-of-the-art snow products concerning the snow extent are remotely sensed and they are based mainly on multispectral optical sensors. They can investigate the snow cover and give information about the size and the shape of snow grains [5]; the presence of impurity soot; the age of the snow; and the presence of depth hoars. Furthermore, the short-wave infrared signal can support the discrimination between snow and clouds [6]. The estimation of the snow extent from remotely sensed multispectral images is based on the relation between the radiative behavior of the surface and the Fractional Snow Cover (FSC). This parameter describes the percentage of surface covered by snow [7] in a pixel element of a remotely sensed image. Considering that snow-covered surfaces are highly reflective in the visible range and low reflective in the short-wave infrared (swir) [8], it is possible to define an index that enhances the discrimination between snow and not snow in a single pixel. This index, defined as Normalized Difference Snow Index (NDSI), is calculated as follows:

$$\text{NDSI} = \frac{\text{green} - \text{swir}}{\text{green} + \text{swir}} \tag{1}$$

The green and the swir parameters are the bands available for each satellite sensor and their selection includes generally wavelength ranges centered at 500–600 nm (green) and 1500–1600 nm (swir). The relation between the FSC and the Normalized Difference Snow Index (NDSI) represents the most common inference required by remote sensing studies. There are two options for estimating the NDSI—FSC relation: the first one consists in combining satellite products with different spatial resolution [9,10]; and the second one can be approached having a ground truth information. The first solution is based on [8] combining Landsat and MODIS data and a NDSI to FSC relation is defined.

$$\text{FSC} = 1.45 \times \text{NDSI} - 0.01 \tag{2}$$

This knowledge is implemented in the SNOWMAP algorithm [11], which is the core of the MODIS data chain for the definition of remotely sensed snow products. The second solution can be approached defining an empirical reflectance-to-snow-cover model that requires a calibration having a number of reference sites in the satellite image. The most important example is the so-called Norwegian Linear Reflectance-to-snow-cover algorithm (NLR) [12] that is the core of the GlobSnow Snow Extent (SE) data chain [13]. From this perspective, the availability of webcam networks is an important data source for calibration and validation processes. The attention of the scientific community of this proxy is increasing, and the literature about this topic is growing [9,14–17]. Furthermore, several tools (for example, FMIPROT and PRACTISE) can be considered for research purposes [18–20]. The solutions available for the analysis of webcam imagery are commonly based on two different processes: orthorectification and classification. While the geometrical issue is based on the mathematical solution of the relationship between pixel elements and the ground surface, the detection of snow cover represents the real cognitive gap. The classification issue can be approached, following the applications available for the remote sensing imagery, using supervised, unsupervised or object-oriented methods [21], depending on the number of images that must be processed.

The focus of this paper is to investigate the contribution of the terrestrial photography to define site-specific thresholds useful for studying the snow cover with remotely sensed data. The expected outcomes are: (i) the description of an automated procedure able to process long time series of ground-based images; (ii) the comparison between automated approaches and supervised methods; (iii) and the evaluation of the potential contribution of terrestrial photography to the snow cover retrieval from remotely sensed data.
