In this work, an investigation of the capability of a statistical cloud detection scheme, implemented through the use of a multilayer feed-forward neural network, is assessed. The whole methodology is applied to a set of IASI L1C spectral radiances, covering the period January 2016-November 2016 and related to Eastern Europe and tropical areas. From a subsampled training dataset where the sky conditions are “certainly” known, we have performed the supervised learning of statistical features of the cloudy- and clear- sky conditions, where truth data have been taken from a cloud mask product of the Advanced Very High-Resolution Radiometer (AVHRR). Also, to improve the neural network classification performances: i) Principal Component Analysis (PCA) of IASI spectra and ii) neural network learning regularization techniques, have been used. Finally, the neural network classification analysis, evaluated during the training with a validation dataset and then with a test dataset, shows very good performance in detecting clouds, with an accuracy of about 93%.
|