Brain-computer interfaces (BCIs) have emerged as a promising neurorehabilitation tool for the motor recovery of people with disabilities. A BCI is designed to detect motor execution or motor intention from brain activity, with the aim of promoting functional recovery in individuals affected by chronic neurological conditions. The use of a BCI consists of two distinct stages: calibration and closed loop. During the calibration stage, EEG signals are recorded over several runs and processed to extract features for training a classifier. This classifier is then used to decode the user’s motor intention in real time during the closed-loop stage. However, when designing BCI intended for clinical environments, it is essential to minimize the calibration time to avoid patient fatigue and ensure clinical feasibility within therapeutic sessions. To address this limitation, the present study explores data augmentation strategies based on polyphase decomposition and empirical mode decomposition (EMD) to increase the number of data required to train the classifier. EEG signals recorded during upper limb motor tasks of healthy subjects were preprocessed using FIR filtering and temporal segmentation. The data augmentation strategies were then applied to the calibration data. From the resulting augmented signals, power spectral density features within the 8–30 Hz range were extracted and used to train a regularized linear discriminant analysis classifier. The results suggest that both augmentation strategies enhance classifier performance, with EMD proving to be the most effective. Notably, EMD enabled comparable performance to be achieved with a reduced amount of calibration data, thereby reducing the overall calibration time. These findings have important clinical implications, as they have the potential to minimize patient fatigue while facilitating the implementation of BCIs in rehabilitation.