A Multi-Challenge Clustering Benchmark Dataset Embedding Large Differences in Spatial Extent

This artificial clustering benchmark dataset was designed manually and draws its inspiration from structural aspects that can be seen in principal component plots of hyperspectral image data. Distance-separated, density-separated, gradient-separated as well as connected clusters have been placed into the dataset. Following the notion that clusters may vary significantly with respect to their spatial extent the respective separability problems are scaled at different levels and only become visible by magnifying certain parts of the dataset. Another special aspect of this dataset is that cluster borders have been kept rather ambiguous which, in our opinion, better resembles the situation in spectroscopic data.