Hierarchical Manifold Learning With Applications to Supervised Classification for High-Resolution Remotely Sensed Images

Manifold learning is one of the representative nonlinear dimensionality reduction techniques and has had many successful applications in the fields of information processing, especially pattern classification, and computer vision. However, when it is used for supervised classification, in particular for hierarchical classification, the result is still unsatisfactory. To address this issue, a novel supervised approach, namely hierarchical manifold learning (HML) is proposed. HML takes into account both the between-class label information and the within-class local structural information of the training sets simultaneously to guide the dimension reduction process for classification purpose. In this process, we extract sharing features to represent the parent manifold´s information, and better solve the out-of-sample problem of manifold learning by using the generalized regression neural network at considerably lower computational cost, thereby making the proposed HML more suitable for supervised classification. Experimental results demonstrate the feasibility and effectiveness of our proposed algorithm.