In this paper, we present a deep learning approach for automatic categorization of age-related macular degeneration (AMD). Faced with the deficiency of training data, we propose a solution to combine additional data to effectively assist the classification task. During training process, the retinal fundus images from two datasets are mapped into a common feature space with adversarial domain adaptation to reduce domain discrepancy. Moreover, we introduce center loss to increase the intra-class compactness of the extracted features to further improve the classification performance. Experiments are conducted on three public fundus image datasets: STARE, ODIR and iCHALLENGE-AMD (hereinafter referred to as iAMD). Our method outperforms three state-of-the-art classification models as well as other augmentation approaches. The proposed approach provides a general framework to handle the issue of training samples with domain discrepancy.