Emotion recognition based on electroencephalography (EEG) plays a significant role in brain-computer interface (BCI) applications. However, individual differences often hinder the generalization of emotion recognition methods to unknown subjects. To address this, we propose an unsupervised domain adaptive model with joint multi-source domain and two-step alignment strategy (JMS2A). The alignment strategy consists of two steps: (1) To capture the structured information from source domain, we combine multiple source domains into a mixed source domain. Simultaneously, a single source domain and the target domain are combined to form a pseudo-target domain, which is then indirectly aligned with the mixed source domain; (2) To extract latent class information from the target domain, we extend supervised contrastive learning to enable direct alignment between source and target domain. We evaluated JMS2A on the SEED and SEED-IV datasets, achieving accuracies of 95.30% and 86.55%, respectively. Experimental results demonstrate that our approach outperforms state-of-the-art methods. The source code is available at https://github.com/cccyangt/JMS2A.