Statistical Models for Snow Detection in the Estimation of Surface Solar Irradiance using the Heliosat Method

Abstract

The aim of this work is to develop a new snow detection classifier to improve the estimation of solar radiation using data of the currently operational MSG-4. Instead of a rulebased method, several statistical classifiers are tested for snow detection, including logistic regression, k-nearest-neighbors, support vector classifier and support vector machine. These statistical methods aim to optimize the classification on a set of training data. In contrast to a stepwise classification like the SAMSAT algorithm, these models consider all the features at once and do not depend on manually chosen thresholds. The classifiers will be trained and evaluated on the basis of manually labeled data and the best classifiers will be integrated into the Heliosat method. The estimation of solar radiation with snow detection will be compared with ground measurements.