We introduce SegEraser, a novel data augmentation framework that broadens the input space and enhances training data depth for sEMG gesture classification. SegEraser comprises two modules: a segmenting module, which partitions complex movements into linear signal segments, and an erasing module, which applies value erasing with contextual information. Unlike fixed-interval methods that fail to capture dynamic gesture changes and potentially produce redundant data, the segmenting module creates dynamic intervals and decomposes movements into finer temporal structures for learning intricate patterns. The erasing module masks randomly selected intervals with contextual information to generate positive training samples, overcoming the challenge of handling the theoretically infinite range of signal values. Furthermore, our experiments demonstrate SegEraser’s effectiveness: it reduces validation loss by 6.67%, lowers error rates by up to 5.58%, and increases robustness against noise and data loss by up to 46.04%. The comparative experiment results also show that SegEraser outperforms similar methods such as Dropout and Random Erasing.