GAN-NL: Unsupervised Representation Learning for Remote Sensing Image Classification

Recently, deep learning methods have greatly enhanced the classification performance because of their strong representation ability in the local receptive field. However, the non-local spatial information always exist in the images. Moreover, the limited amount of the labeled data imposes great challenges on the supervised representation learning model, especially the remote sensing images. With the consideration, we propose a generative adversarial network with non-local spatial information (GAN-NL) for remote sensing image classification. Specifically, a non-local layer is incorporated into a generative adversarial network for unsupervised representation learning. Then, a classification network is designed to infer the labels of the images. The classification results on the challenging NWPU-RESISC45 remote sensing image dataset show that our proposed method performs favorably against the state-of-the-art methods in terms of the classification accuracy without any pre-training.