We have proposed a new multi-manifold learning algorithm, which combines semi-supervised multimanifold modeling, nonlinear feature extraction and a new semi-supervised discriminant analysis method to achieve better performance in geodesic feature extraction and classification tasks.