Orthogonal time–frequency space (OTFS) modulation has emerged as a promising technology to alleviate the effects of the Doppler shifts in high-mobility environments. As a prerequisite to demodulation and signal processing, automatic modulation classification (AMC) is essential for OTFS systems. However, a very limited number of works have focused on this issue. In this paper, we propose a novel AMC approach for OTFS systems. We build a dual-stream convolutional neural network (CNN) model to simultaneously capture multi-domain signal features, which substantially enhances recognition accuracy. Moreover, we propose a differentiated embedded pilot structure that incorporates information about distinct modulation schemes to further improve the separability of modulation types. The results of the extensive experiments carried out show that the proposed approach can achieve high classification accuracy even under low signal-to-noise ratio (SNR) conditions and outperform the state-of-the-art baselines.