Laplacian Regularized D-Optimal Design for Remote Sensing Image Classification

Abstract

Obtaining training sample for remote sensing image classification is time consuming and expensive especially for relatively inaccessible locations. Therefore, determining which unlabeled samples would be the most informative if they were labeled and used as training samples is the most delicate phase. Particularly, we consider the problem of active learning in remote sensing image classification. However, Classical optimal experimental design approaches are based on least square errors over the labeled samples only. They fail to take into account the unlabeled samples. In this paper, a manifold learning technique which is performed in the sample space by using graph Laplacian is applied to reflect the underlying geometry of the sample. By minimizing the least square error with respect to the optimal classifier, we can select the most representative and discriminative sample for labeling. The effectiveness of the proposed method is evaluated by comparing it with other active learning techniques existing in the literature. Experimental results on data set confirmed the effectiveness of the proposed technique.