Cloud detection using machine learning techniques with application to IASI measurements

"Cloud detection using machine learning techniques with application to IASI measurements" Cloud detection is a crucial component of operational satellite data analysis, especially for atmospheric monitoring and reconstruction. This study investigates the application of machine learning algorithms, specifically Support Vector Machines (SVM), to detect clouds using satellite radiance data. The dataset consists of spectral radiances from the Infrared Atmospheric Sounding Interferometer (IASI) Level 1C product, spanning the spectral range of 645 to 1600 cm^(−1). Three distinct approaches to dataset partitioning were implemented, each employing different strategies for dividing the training and test datasets. The Soil Type index from the ERA5 database, part of the fifth-generation ECMWF reanalysis, is used for data comparisons. The three setups in which the tests were conducted are as follows: the first analysis included global satellite observations without any additional regional subdivisions; the second grouped the observations into five distinct climatic zones; and the third further subdivided these climatic zones longitudinally into three regions. The SVM algorithm solves a quadratic optimization problem to find a hyperplane that maximizes the margin between cloudy and clear-sky classes. Spectral radiance values are the feature vectors, and cloud classification is the target label. Kernel functions map data to higher-dimensional spaces for non-linear decision boundaries. The results of this study demonstrate good results in cloud detection based solely on radiance data, without relying on physical models or prior domain specific knowledge. This underscores the efficiency of machine learning techniques, particularly SVM, in simplifying and enhancing cloud detection workflows in satellite meteorology. While this research is ongoing, the preliminary outcomes are highly promising, marking a significant advancement in leveraging data driven approaches for atmospheric studies. This work establishes a foundation for further refinement and application of these techniques in operational and research settings.