Feature-independent classification of hyperspectral images by projection-based random forests

The automatic classification of land cover types from hyper-spectral images often relies on expert knowledge and focuses on the extraction of meaningful features that enable a subsequent classifier to distinguish between different land cover classes. In this work Projection-Based Random Forests are applied to hyperspectral images, which completely overcome the need of an explicit feature extraction or any other kind of (semi-) automatic preprocessing. Instead the classifier works on the data as it is, selects meaningful spectral bands automatically and only implicitly computes features. The proposed method is applied to four challenging hyperspectral datasets and achieves state-of-the-art performance. A deeper insight into the classification task is provided as information about which bands are useful and which are not.