Development of unsupervised or weakly supervised Deep Learning methods for cloud detection

Clouds usually represent a big problem for optical remote sensing. A wide range of works have been developed in producing efficient cloud detection algorithms. However, most algorithms still depend on quality assurance data provided by image providers. In addition, most of those algorithms generate a considerable quantity of false positives. Those false positives make it harder to develop automated methods for analyzing cloudy regions, frequently generating misobservation problems on clear areas. One of the biggest challenges to generate an effective method for accurately detect clouds on a worldwide basis is the lack of accurate training data encompassing the wide range of land covers and atmospheric conditions. In that sense, unsupervised and weekly supervised Deep Learning methods have demonstrated a good potential to tackle these types of tasks with a suitable accuracy and requiring considerably less amounts of labeled samples. This MSc topic focus on designing and implementing efficient unsupervised or weekly supervised methods that are able to detect clouds in images from different sensors, acquired in different parts of the world, without the use of massive amounts of training labels.