Generalized Manifold Learning for High Resolution Remote Sensing Image Object Classification

It is not adequate to use classical manifold learning techniques to reduce the dimension of covariance descriptors lied on Riemannian manifold. A generalized manifold learning method named Log-Euclidean Riemannian kernel-based adaptive semi-supervised orthogonal locality preserving projection (LRK-ASOLPP) is proposed, and successfully applied to the high resolution remote sensing image classification issue. Firstly, geometric features of each pixel in the image are extracted, and covariance descriptor of each image is calculated. Secondly, the covariance descriptors are mapped into the reproducing kernel Hilbert space by using the Log-Euclidean Riemann kernel. Thirdly, the model of semi-supervised orthogonal locality preserving projection algorithm on Riemannian manifold is constructed based on manifold learning theory. Fourthly, by using the alternating iteration optimization algorithm to solve the objective function, the similarity weight matrix and low dimensional projection matrix are obtained simultaneously. Finally, low dimensional projections of test samples are computed by using the low dimensional projection matrix, and then classifiers such as K-NN, support victor machine (SVM), etc. are used to classify them. Experiment results on three high-resolution satellite images datasets demonstrate the feasibility and effectiveness of the proposed algorithm. Copyright © 2019 Acta Automatica Sinica. All rights reserved.