Enhancement of Land Cover Classification by Training Samples Clustering

Abstract

In this study, a hybrid approach is proposed to enhance land cover classification accuracy by clustering training samples into homogenous subclasses. The proposed approach implies the integration of both supervised and unsupervised classification methods into a holistic framework. A criterion of training sample separability is developed as separability index of training samples. The approach was applied to enhance the land cover classification of the highly heterogeneous natural landscapes by the case of the Shatsky National Natural Park.