Evaluation of intrinsic dimensionality methods using residual and change-point analyses

The number of endmembers (NOE) in hyperspectral imagery plays an important role in image analysis applications such as classification, clustering and unmixing. Over the last years, different algorithms have been proposed to estimate the NOE. Nonetheless, each method depends on its own parameters' values, and as a result, leads to different values for intrinsic dimensionality (ID). In this study a statistical-based method is proposed to evaluate different results of ID algorithms. In this method, the reasonable candidates of ID are selected using both residual analysis (RA) and Change-Point analysis (CPA). Different values for ID are then compared with these candidates. If these values are equal or close to these candidates they may be considered as the ID. Although the proposed method can be used for every ID method, here, the results of two new methods, namely, SML, and O-GENE-AH algorithms have been investigated on Pavia Center hyperspectral dataset.