A new post-classification and band selection frameworks for hyperspectral image classification

Hyperspectral image classification has been an active field of research in recent years. The high dimensionality of spectral bands and the small number of training pixels cause the Hugh phenomenon and reduce significantly the classification results quality. In this paper, we introduce a new framework for hyperspectral images classification. The proposed approach is composed of three steps. First, the problem of band selection is considered. We propose to merge the adjacent bands that are highly correlated and to select the bands that maximize the class separability using the Jeffries-Matusita distance. The second step consists to use the bagger algorithm, SVM and KNN to classify the pixels. Finally, a post-classification algorithm of misclassified pixels namely Classification Errors Correction (CEC) is developed. The algorithm consists to correct the label assigned by the classifier system for each pixel by exploiting the labels of neighbors and the spectral information around the pixel according to certain transitions. Experimental results show that the proposed approach improves considerably the classification quality. The band selection approach and the CEC algorithm enable us to achieve a high classification accuracy rate even when the number of training pixels is very small.