Classification of Hyperspectral Images Using Cluster Space Linear Discriminant Analysis and Small Training Set

The hyperspectral images allow us to discriminate between different classes with more details. There are lots of spectral bands in hyperspectral images. On the other hand, the limited number of available training samples causes difficulties in classification of high dimensional data. Since the gathering of training samples is hard and time consuming, feature reduction can considerably improve the performance of classification. So, feature extraction is one of the most important preprocessing steps in analysis and classification of hyperspectral images. Feature extraction methods such as LDA have not good efficiency in small sample size situation. A supervised feature extraction method is proposed in this paper. The proposed method, which is called cluster space linear discriminant analysis (CSLDA), without obtaining the label of testing samples and just with doing a clustering on testing data, finds the relationship between training and testing samples. Then, it uses the power of unlabeled samples together with training samples for estimation of within-class and between-class scatter matrices. The CSLDA improves the classification accuracy particularly in multimodal hyperspectral data. The experimental results on urban and agriculture hyperspectral images show the better performance of CSLDA compared to popular feature extraction methods such as LDA, GDA, and NWFE using limited number of training samples.