Semi-supervised learning in habitat classification from remotely-sensed imagery

Abstract

Remote sensing helps monitor and evaluate the state of ecosystems, covering also wilderness areas that can be hard to access for field observations. Wilderness areas, such as the ones in northern Lapland, are home to endangered species and habitat types. Automatic detection and classification of habitats is a difficult task, as target class distributions are long-tailed, fine-grained, and have semantic properties that can be difficult to distinguish even for humans and especially from limited remotely sensed imagery. Training data for building models is often sparse, point-like, and limited to areas accessible by foot. This thesis presents methods for habitat classification from limited data using supervised, unsupervised, and semi-supervised methods. The presented approaches take advantage of the large amounts of unannotated and weakly annotated source data that is available. Convolutional neural networks and random forests are compared and an ensemble model combining both approaches is shown to increase classification performance. Convolutional neural networks are also used to produce fully unsupervised segmentation maps. The classification and segmentation maps are produced for the entire northern Lapland area.