Hyperspectral hierarchical representation based on density peak ranking

An unsupervised hyperspectral hierarchical representation (HHR) framework is proposed by estimating density peak index (DPI). To accelerate our proposed algorithm, we first aggregate adjacent homogenous pixels into extended-pixels as basic initial units. Next, to rank the importance of initial units, both local density and interunit distance of each extended-pixel are taken into account to calculate the DPI. Then, a local–global merging procedure is conducted to construct the hierarchical structure. Selecting the extended-pixels with high DPI values as centers, the related adjacent extended-pixels are merged into local-regions. To generate final HHR, the local-regions are grouped globally based on their updated DPI rank. Considering the representative property of HHR, we also explore a hyperspectral classification strategy. Different from traditional segmentation-based classification methods, our classification scheme only focuses on the classification of the representative subregions, then the labels of other subregions directly refer to the results of representative subregions in the same HHR group. Thus, the computational load for classification can be significantly decreased than for pixel-wise classification. The experiments on three hyperspectral subsets show that the proposed HHR framework can yield remarkable performance in unsupervised clustering, and the classification scheme can obtain equivalent classification accuracy with distinctly limited calculation.