Estimating the Intrinsic Dimensionality of Hyperspectral Remote Sensing Imagery with Rare Features

Estimating the intrinsic dimensionality of hyper spectral remote sensing imagery is an essential step in processing this kind of data. A novel estimation algorithm is proposed, which can preserve both abundant and rare features in original data. First of all, the QR decomposition of an original data matrix is carried out in order to decrease computational complexity, and a sliding noise detection window is applied to noise reduction for improving the accuracy of dimensionality estimation. Furthermore, a manifold learning method is used to determine a limit of intrinsic dimensionality and finally, intrinsic dimensionality is estimated through the singular value decomposition and $l_{2,\infty }$-norm theory. The experimental results of simulated and real data are presented, which shown our proposed algorithm outperforms some classical algorithms.