A Discriminative Feature Learning Approach for Remote Sensing Image Scene Classification

To overcome the challenging task of remote sensing image scene classification, deep convolutional neural network (DCNN) has been widely adopted in remote sensing image scene recognition for their excellent performance. Compared with the hand-crafted feature method and the significantly improved performance by using DCNN, it is a considerable challenge to identify various remote sensing image scenes due to the problems of similarity of intraclass features and dissimilarity of interclass features. In this paper, an effective method called the center metric method to address this problem is proposed. Compared with existing DCNN models, we improve the objective function by adding a center metric regularization term, which consists of a cross entropy loss term, a center metric loss term and a weight decay term. The center metric regularization term consists of two parts: one part enforces mapping similar scene classes closely to their center; the other restricts the minimum distance between the sample and the center of other classes. The result on two datasets demonstrate that the improved loss function can make the feature more discriminative.