Semi-supervised Remote Sensing Image Scene Classification Based on Generative Adversarial Networks

Abstract

With the availability of numerous high-resolution remote sensing images, remote sensing image scene classification has been widely used in various fields. Compared with the field of natural images, the insufficient number of labeled remote sensing images limits the performance of supervised scene classification, while unsupervised methods are difficult to meet the practical applications. Therefore, this paper proposes a semi-supervised remote sensing image scene classification method using generative adversarial networks. The proposed method introduces dense residual block, pre-trained Inception V3 networks, gating unit, pyramidal convolution, and spectral normalization into GANs to promote the semi-supervised classification performance. To be specific, the pre-trained Inception V3 network is introduced to extract semantic features to enhance the feature discriminant capability. The gating unit is utilized to capture the relationships among features. The pyramidal convolution is integrated into dense residual block to capture different levels of details to strengthen the feature representation capability. The spectral normalization is introduced to stabilize the GANs training to improve semi-supervised classification accuracy. Extensive experimental results on publicly available EuroSAT and UC Merced datasets show that the proposed method gains the highest overall accuracy, especially when only a few labeled samples are available.