In recent years, event-based deep learning has garnered widespread attention among researchers. Compared to conventional visual data, event data exhibits unique advantages in terms of temporal resolution and adaptability to dynamic lighting conditions. However, the extreme sparsity of event streams over prolonged periods limits effective feature extraction, while existing data augmentation methods fail to adequately exploit latent spatiotemporal patterns. To address this issue, three data augmentation strategies were proposed: temporal chunking, random beginning time, and geometric dimension-wise random drop. These methods comprehensively augment event data along both temporal and geometric dimensions, thereby mitigating the limitations of sparse event data in latent feature extraction. Experimental results on three event-based datasets including DVS-Gesture, DVS128-Gait-Day, and DVS-CIFAR10 demonstrate that the proposed methods enable models to achieve superior recognition performance under short input time windows, significantly improving the utilization efficiency of event data. Compared to baseline methods, the classification performance is enhanced by 17.62%, 18.5%, and 9.7%, respectively.