Retrieving fractional snow cover in Central Apennines from Sentinel 2 and 3 visible-infrared spectroradiometer data and random forest learning techniques

Seasonal snow cover is the largest cryospheric component of in areal extent, covering more than 50 million square kilometers of the Earth surface (more than 31% of its land area) every year. Snow cover area (SCA) and its local properties, in terms of snowpack height and snowpack density, are the main parameters characterizing the snow accumulation in mountainous regions. Such parameters result in particular importance in meteorology, hydrology, and climate monitoring applications. Anyway, in the general case, the considerable geographical extension of snow layers and their typical spatial heterogeneity makes it impractical to monitor the above three parameters regularly (i.e., with a high spatial and temporal resolution) by means of direct or indirect in situ measurements, suggesting the exploitation of satellite technologies for the provision of such data. Snow cover patterns are governed by the effects of topography, land cover, wind redistribution, solar irradiance, and air temperature. On the other hand, in the last few decades, a general back-scaling of snow observation networks occurred worldwide. Based on the above considerations, space-borne SAR sensors are particularly suitable for the analysis of snow deposits, providing data with resolutions up to some meters, with global coverage and a few days revisit time.

In this study, we introduce a satellite-based technique for mapping snow cover fraction balancing the requirements between spatial and temporal resolution, and using data from the European Sentinel constellation. The available current fractional snow cover (FSC) products, provided by Sentinel-2 MSI (Multispectral imager) cloud-gap-filled (CGF) products and Terra MODIS (Moderate-resolution Infrared Spectroradiometer) snow cover products, may suffer either of relatively poor spatial resolution and/or temporal resolution (e.g., FSC at 25-m spatial resolution every 5 days from Sentinel-2 MSI products or 500-m spatial resolution every day from Terra MODIS). For this purpose, we explore the use of the Sentinel-3 optical sensors, OLCI (Ocean Land Color Imager), and SLSTR (Sea-Land Surface Temperature Radiometer), showing a 300-m and 500-m spatial resolution with 2-3 and 1-2 days temporal resolution.

Using as a reference the Sentinel-2 FSC product and employing a DEM (Digital Elevation Model) at 90 m spatial resolution, a machine learning Snow-Cover-Area Random Forest (SCARF) approach has been developed. The proposed algorithm takes, as inputs, both DEM as well as OLCI and SLSTR data, linearly up-sampled at 90-m, and can provide as output FSC product at 90-m spatial resolution every 1-2 days. Input data are derived from NASA SRTM 3-arc-second DEM, OLCI multi-band reflectances, and SLSTR multi-band reflectance and brightness temperatures at nadir and oblique view. After creating 2 datasets (nadir and oblique), we have introduced a distinction between the complete dataset and a subset leaving only the pixel with an elevation higher than 1000 m. As a classification method, we used an RF gradient-boosting classifier (called XGBoostClassifier). In this work, we will illustrate the results of the proposed SCARF algorithm using area-of-interest the Italian Central Apennines and period-of-interest winter 2019-20. Statistical performances, potential developments, and critical issues of the SCARF algorithm will also be discussed.