On the performance improvement for linear discriminant analysis-based hyperspectral image classification

In this paper, we present a strategy to improve the performance of Fisher's linear discriminant analysis (FLDA) in dimensionality reduction for hyperspectral image classification. The practical difficulty of applying FLDA to hyperspectral imagery includes the unavailability of enough training samples and unknown information for all the classes including background. The original FLDA has been modified to avoid the requirements of training samples and complete class knowledge, which needs the desired class signatures only. The modified FLDA (MFLDA) can better preserve class information in the low-dimensional space. However, for an image scene with p known classes, the data dimensionality after FLDA and MFLDA transform is p-1. The class-separability performance of FLDA and MFLDA may be significantly improved if the transformed data dimensionality is p instead of p-1. An approach is proposed for this purpose and experimental results demonstrate its advantage.