SEMI-SUPERVISED CLASSIFICATION OF HYPERSPECTRAL IMAGE BASED ON SPECTRAL AND EXTENDED MORPHOLOGICAL PROFILES

The contradiction between high dimensional data and limited training samples is the main problem in hyperspectral remote sensing images classification. How to obtain high classification accuracy with limited labeled samples is an urgent issue. We propose a semi-supervised classification algorithm SSP_EMP for hyperspectral remote sensing images based on spectral and spatial information. The spatial information is extracted by building extended morphological profiles (EMP) based on principle components of hyperspectral image. Utilize spectral and EMP from two view to enrich knowledge, and integrate the useful information of unlabeled data at the most extent to optimize the classifier. Pick high confident samples to augment training set and retrain the classifier. This process is performed iteratively. The proposed algorithm is tested on AVIRIS Indian Pines. Experimental results show significant improvements in terms of accuracy and kappa coefficient compared with the classification results based on spectral, EMP and the combination of spectral and EMP.