A study on the performances of different metrics of a machine learning cloud classificator

Detecting clouds and clouds properties is essential for meteorological research, climate modeling and weather forecasting. CIC (Cloud Identification and Classification) is a recently proposed, innovative machine learning algorithm adopted as the cloud identification code in the ESA Far-infrared Outgoing Radiation Understanding and Monitoring (FORUM) End2End simulator (FE2ES). CIC performs a classification by defining an eigenvectors-based Similarity Index that measures the information content brought into a Training Set when it is concatenated with a new observation. In this thesis work, a new metric called the eigenvalues-based Similarity Index is proposed and studied to quantify the change in information content. Additionally, a novel methodology is developed within the CIC algorithm framework, allowing the simultaneous utilization of multiple Similarity Indices. A cross-validation procedure is conducted using downwelling radiance spectra collected on the Antarctic Plateau to test and compare three configurations: classical eigenvectors-CIC, eigenvalues-CIC, and double-CIC. The double-Similarity-Index CIC demonstrates superior performance and is selected for the classification of a comprehensive dataset of spectra obtained by the REFIR-PAD instrument on the Antarctic Plateau from 2013 to 2020. The analysis yields statistically consistent results with previous studies. Finally, the double-CIC has been used for the challenging analysis of a large dataset, containing the simulated radiance of the Thermal Infra-Red Spectrometer (TIRS) of the PREFIRE mission. TIRS is characterized by only 64 wavenumber channels, and the synthetic data have global coverage. Very good cloud detection performances (more than 93% of clear and cloudy sky spectra correctly classified) are obtained in this study. This constitutes an important result that witnesses the classification power of CIC-like algorithms and their suitability for satellite missions.