The current deep learning studies on remote sensing mainly focused on deep learning algorithms rather than deep learning datasets. This study proposes a method of dataset consistency assessment based on deep learning, in which the similarity among various types of ships from different sources (such as satellite remote sensing, 3D modeling, and web crawler) is evaluated and then used to characterize the consistency of the ship dataset. The results show that when the consistency of the dataset is the highest, the consistency by the proposed method is 1. When the consistency of the datasets is gradient, the consistency also changes. Images with similar data sources can be considered as same class, and images with greatly differences cannot be merged. Thus, the proposed method can assess the dataset consistency properly, and provide a suggestion to build an image dataset for deep learning training.