Dimensionality Reduction with Improved Local Tangent Space Alignment for Hyperspectral Imagery Classification

Abstract:

The paper proposes a new version of local tangent space alignment (LTSA), named multi strategies upgraded local tangent space alignment (MSU LTSA), to solve the problem of computational complexity in dimensionality reduction of hyperspectral imagery (HSI) for classification. First, random projection is introduced into the new method to reduce the number of HSI bands. That decreases the computational complexity of k nearest neighbors (KNNs) construction and local tangent space construction of each pixel. Then, the recursive lanczos bisection algorithm is utilized to construct the fast approximate KNNs graph and it reduces the computational time of regular approach. Finally, when finishing constructing the global alignment matrix, the new method uses the fast approximate singular value decomposition to promote the computational speed of the regular eigenvalue decomposition of global alignment matrix. With two different HSI datasets, four groups of experiments are designed to completely analyze and testify the performance of computation and classification for MSU LTSA. The results show that MSU LTSA speeds up LTSA at least 3 times whereas only degrading about 1% in its overall classification accuracy (OCA).

Key words:multi strategies upgraded local tangent space alignment; local tangent space alignment; random projection; dimensionality reduction; hyperspectral image classification