Category-Oriented Adversarial Data Augmentation via Statistic Similarity for Satellite Images

Deep learning is one of the essential technologies for remote sensing tasks, which heavily depends on the quantity of training data. However, it is difficult to obtain or label the remotely sensed images in their non-cooperative imaging mode. Data augmentation is a viable solution to this issue, but most of the current data augmentation methods are task specific or dataset specific, which are not as applicable as a generalized solution for the remotely sensed images. In this paper, we propose a category-oriented adversarial data augmentation method using statistic similarity cross categories, which formulates the common appearance-based statistic factors in the object detection into a combination index, to depict the statistic similarity between different categories and to generate new adversarial samples between similar categories with more reliable physical significance. Experimental results demonstrated that, taking the most advanced RT method as a baseline, the total mAP can be increased by 2.0% on the DOTA dataset for the object detection task by using our proposed method.