Fine-grained visual categorization (FGVC), which aims at classifying objects with small inter-class variances, has been significantly advanced in recent years. However, ultra-fine-grained visual categorization (ultra-FGVC), which targets at identifying subclasses with extremely similar patterns, has not received much attention. In ultra-FGVC datasets, the samples per category are always scarce as the granularity moves down, which will lead to overfitting problems. Moreover, the difference among different categories is too subtle to distinguish even for professional experts. Motivated by these issues, this paper proposes a novel compositional feature embedding and similarity metric (CECS). Specifically, in the compositional feature embedding module, we randomly select patches in the original input image, and these patches are then replaced by patches from the images of different categories or masked out. Then the replaced and masked images are used to augment the original input images, which can provide more diverse samples and thus largely alleviate overfitting problem resulted from limited training samples. Besides, learning with diverse samples forces the model to learn not only the most discriminative features but also other informative features in remaining regions, enhancing the generalization and robustness of the model. In the compositional similarity metric module, a new similarity metric is developed to improve the classification performance by narrowing the intra-category distance and enlarging the inter-category distance. Experimental results on two ultra-FGVC datasets and one FGVC dataset with recent benchmark methods consistently demonstrate that the proposed CECS method achieves the state of-the-art performance.