Hochstuhl, SylviaSylviaHochstuhlHammer, HorstHorstHammerThiele, AntjeAntjeThieleHinz, StefanStefanHinz2023-10-302023-10-302023https://publica.fraunhofer.de/handle/publica/45240410.1109/igarss52108.2023.10281658Machine learning methods have proven to be a powerful tool for the classification of Polarimetric Synthetic Aperture Radar (PolSAR) images. Since well-established pixel- or image-based classifiers expect real-valued input data, the information content of complex-valued PolSAR data needs to be represented by real-valued descriptors. This paper proposes a method to enable the generation of suitable descriptors without the explicit extraction of hand-crafted and preselected polarimetric features. For this purpose, the neighbor graph based dimension reduction method Uniform Manifold Approximation and Projection (UMAP) is applied to translate pixel similarity measured by the revised Wishart distance to real-valued 3-dimensional feature descriptors. Land cover classification results performed on a real-world dataset, show that the resulting 3-dimensional feature descriptor which allows a fast and memory efficient training, provides a similar level of information compared to an exhaustive feature set composed of 30 polarimetric features.enconference paper