Along with the development of deep learning, object detection and classification performance in the field of computer vision has made a breakthrough. In order to classify objects with high accuracy, it is important to select an optimal deep learning model for each environment. Especially for non-typical environments, there is a high probability that the deep learning model will perform differently than expected. In this paper, we compare and evaluate the ship classification performance of the existing deep learning model for the marine environment where it is difficult to acquire the data set, and experimentally change the performance according to the dataset change.