Brain-Computer Interfaces (BCI) try to interpret brain signals, such as EEG, to issue some command or to characterize the cognitive states of the subjects. A strong limitation is that BCI tasks require a high concentration of the user, de facto limiting the length of experiment and the size of the dataset. Furthermore, several BCI paradigms depend on rare events, as for event-related potentials, also reducing the number of training examples available. A common strategy in machine learning when dealing with scarce data is called data augmentation; new samples are generated by applying chosen transformations on the original dataset. In this contribution, we propose a scheme to adapt data augmentation in EEG-based BCI with a Riemannian standpoint: geometrical properties of EEG covariance matrix are taken into account to generate new training samples. Neural network are good candidates to benefit from such training scheme and a simple multi-layer perceptron offers good results. Experimental validation is conducted on two datasets: an SSVEP experiment with few training samples in each class and an error potential experiment with unbalanced classes (NER Kaggle competition).