Manifold Regularized Distribution Adaptation for Classification of Remote Sensing Images

We perform unsupervised domain adaptation for the classification of remote sensing data by learning a shared subspace in this paper. Maximum mean discrepancy (MMD) is applied to each class, making the approach able to minimize the domain shift on a per-class basis. Furthermore, manifold regularization is employed to constrain the data manifold of both the source and target data to be preserved in the subspace. The manifold regularization in conjunction with the per-class MMD strategy is called manifold regularized distribution adaptation (MRDA) algorithm. Since the class mean of target data should be estimated by the predicted labels, we integrate spatial information and overall mean coincidence (OMC) method to improve the prediction accuracy, resulting in Spa_OMC_MRDA approach. Experimental results on both multispectral and hyperspectral remote sensing data indicated the good performances of the proposed approach.