Paper regarding feature clustering of high dimensional feature spaces and subsequent ranking of the features published...
The paper "Analyzing Hyperspectral and Hypertemporal Data by Decoupling Feature Redundancy and Feature Relevance" has been published in IEEE Geoscience and Remote Sensing Letters by Held, M., Rabe, A., Senf, C., van der Linden, S., Hostert, P.
The researchers show that a clustering of features of highly correlated high dimensional data (e.g. bands in hyperspectral data or time steps in hypertemporal data, e.g. MODIS) is a reasonable step prior to the classification. We cluster (1) 114 hyperspectral features into 15 clusters and (2) 30 MODIS time steps into 6 clusters and use cluster representatives thereof, which achieved similar results in the support vector classification to using all features. Moreover, these clusters are then ranked according to their relevance for the application (by feature forward selection), here classification, which provides information about the dataset and the spectral or temporal segments relevant for the application. Finally, since the processing time of the feature forward selection grows to the second power with respect to the number of features, the time needed for this step is greatly reduced by the prior clustering.