A Machine Learning - Explainable AI approach to tropospheric dynamics analysis using Water Vapor Meteosat images

Abstract

Water vapor in the atmosphere plays a crucial role in the energy balance and weather, being responsible for half of the greenhouse effect. Meteorological satellites detect the water vapor and represent it on images, providing important information to understand and forecast the flow dynamics of the General Atmospheric Circulation System. A collection of computational intelligence techniques was used to investigate the structure of a large series of Meteosat (ESA) water vapor band (WV6.2) hourly images from 2009 to 2020. These techniques include the Visual Information Fidelity image quality measure, unsupervised and supervised machine learning and explainable AI methods. Explainable AI methods (XAI) like Permutational Variable Importance, Local Interpretable Model-Agnostic Explanations, Shapley Additive Explanations and Ceteris Paribus profiles, were able to discover temporal variations and changes on the water vapor patterns. The results obtained demonstrate the great potential of ML and XAI in the domain of atmosphere dynamics and weather evolution.