Cloud Identification and Classification from Ground Based and Satellite Sensors on the Antarctic Plateau

Cloud identification from satellites is considerably challenging in polar environments due to the similar radiative properties of surface and ice clouds, making it difficult to detect and distinguish cloud features. CIC (Cloud Identification and Classification) is a machine learning algorithm adopted as the official software in the ESA Far-infrared Outgoing Radiation Understanding and Monitoring (FORUM) (Palchetti et al., 2020) End2End simulator (FE2ES). CIC is based on Principal Component Analysis and performs cloud detection and multi-scene classification. It is adaptable to every type of sensor and is particularly suitable when a small number of elements are available for the Training Set. Assessment studies have already been conducted to evaluate the performances of the algorithm in multiple conditions. In Maestri et al. (2019), CIC was applied to simulated radiance all over the globe, while Magurno et al. (2020) used the algorithm to analyze airborne interferometric spectra. Finally, in Cossich et al. (2021) the algorithm was tested on downwelling radiances collected at Dome-C in Antarctica. In this work, CIC is applied to high spectrally resolved data taken from ground and, for the first time, from satellites. Ground-based data are collected by the REFIR-PAD sensor (Di Natale et al., 2020), covering the far and mid-infrared part of the spectrum. Collocated satellite data are measured by IASI (Infrared Atmospheric Sounding Interferometer) which collects upwelling radiance between 3.4 and 15.5 μm. The period under study spans from 2012 to 2022. CIC results applied to ground-measured spectra are compared to IASI’s L2 classification products. Large discrepancies between the two classifications are observed, indicating an overestimation of the cloud occurrence in case of IASI. A verification is obtained using collocated ground-based LIDAR measurements, which are available for subsets of the collocated radiances. Finally, the CIC algorithm is trained with a subset of IASI data collocated with REFIR-PAD and LIDAR measurements. The training set is defined also with the help of the Advanced Very High Resolution Radiometer (AVHRR) on board of MetOp satellites. The AVHRR has 1 km resolution (at the nadir) and its collocated measurements are used to evaluate the scene homogeneity in the satellite field of view. Statistical analyses are then performed on IASI spectra using the CIC classification. Results indicate a much better agreement with ground-based data, improving the cloud occurrence provided in IASI L2 products.