A Modified Locality-Preserving Projection Approach for Hyperspectral Image Classification

Abstract

Locality-preserving projection (LPP) is a typical manifold-based dimensionality reduction (DR) method, which has been successfully applied to some pattern recognition tasks. However, LPP depends on an underlying adjacency graph, which has several problems when it is applied to hyperspectral image (HSI) processing. The adjacency graph is artificially created in advance, which may not be suitable for the following DR and classification. It is also difficult to determine an appropriate neighborhood size in graph construction. Additionally, only the information of local neighboring data points is considered in LPP, which is limited for improving classification accuracy. To address these problems, a modified version of the original LPP called MLPP is proposed for hyperspectral remote-sensing image classification. The idea is to select a different number of nearest neighbors for each data point adaptively and to focus on maximizing the distance between nonnearest neighboring points. This not only preserves the intrinsic geometric structure of the data but also increases the separability among ground objects with different spectral characteristics. Moreover, MLPP does not depend on any parameters or prior knowledge. Experiments on two real HSIs from different sensors demonstrate that MLPP is remarkably superior to other conventional DR methods in enhancing classification performance.