Snow cover detection in Estonia from SAR images using machine learning methods

Abstract

Usability of optical satellite data for monitoring snow cover can be limited in regions with frequent high cloud coverage. Synthetic aperture radar (SAR) could theoretically be used to monitor snow regardless of clouds or lack of illumination. There are several factors that complicate the task in Estonia such as dense vegetation and quickly changing snow conditions. So far most studies on using SAR for snow detection have been done in mountainous regions and over short time periods. The aim of this study was to test applicability of a method that combines most common features for snow detection extracted from SAR images in a machine learning model. This method had shown good transferability in mountain regions, however the modelling results on Estonian data were unsatisfactory. Analysis of features derived from SAR images revealed poor separability of snow free and snow covered classes. This suggest the main issue is with the feature extraction methods rather than machine learning. Possibly the processing chain could be optimized for Estonia and other regions with flat topography and predominantly dense vegetation. This thesis did not result in a usable model, but could serve as a basis for further studies.