Multi-view feature learning for VHR remote sensing image classification

摘要

Learning high-level semantic information is important for the task of remote sensing(RS) image scene classification. Due to the great intraclass diversities and the interclass similarities, many researchers have explored the convolutional neural network(CNN) to handle this task recently. However, RS images usually have confusing backgrounds, such as the relevant objects, and features only derived from the whole RS images can not achieve satisfying results. Additionally, the great intraclass diversities also increase the difficulty of recognizing the RS images correctly. To solve the problem, the multi-view feature learning network(MVFLN) is proposed to obtain three domain-specific features for the scene categorization task. FC layers in the VGGNet are replaced by the channel-spatial branch and the other multiple metric branchs. The channel-spatial branch is utilized to localize and learn discriminative regions while the triplet metric branch and the center metric branch are used to enlarge the distance between different classes and reduce the distance of samples belonging to the same class, respectively. In this situation, the proposed MVFLN conducts in a concise way without extra SVM classifiers, achieving better performance. Experiments conducted on the AID, NWPU-RESISC45 and UC Merced datasets evaluate its effectiveness.