Semi-supervised manifold learning and its application to remote sensing image classification

To improve the remote sensing image classification accuracy by incorporating labeled and unlabeled samples, this paper proposes a new manifold learning method called Semi-supervised Manifold Discriminant Embedding (SSMDE). This method uses data point labels to construct two relational graphs, within-class graph and between-class graph, they then are taken to encode the class relation information indicated in the labeled data points and to construct two weighted matrices. The labeled and unlabeled data points are utilized to construct the total scatter matrix to describe all the data points. Finally, the projection matrix of SSMDE is obtained by solving an optimization problem. The SSMDE method can not only take into account the discriminant information of labeled data, but also preserve the global structure of all data points. The experimental results on both synthetic and remote sensing images show that the proposed method can achieve the classification accuracy of 92.32% and the error between the classification results by the SSMDE and the government statistics is less than 5%, which demonstrates the effectiveness of SSMDE.