Single-cell RNA sequencing (scRNA-seq) data analysis faces significant challenges due to high dimensionality, sparsity, noise, and batch effects, all of which complicate accurate cell clustering and downstream tasks. To address these issues, we introduce scCMA, a novel method that integrates contrastive learning with a masked autoencoder framework to generate robust, high-quality cell embeddings. Through contrastive learning, scCMA enhances the discriminative power of feature representations by maximizing similarity within cell types and minimizing it across types, implicitly reducing batch effects without prior dataset knowledge. Simultaneously, the masked autoencoder randomly masks and reconstructs gene expression data, enabling the model to capture global dependencies and rare features while mitigating the impact of noise and sparsity. Evaluations across diverse datasets demonstrate that scCMA achieves superior accuracy in cell clustering, effectively mitigates batch effects without compromising biological heterogeneity, and sensitively identifies rare cell populations. Furthermore, the learned embeddings preserve developmental dynamics, enabling precise reconstruction of differentiation trajectories. The implementation code for scCMA is available at the following link: https://github.com/chenxofhit/scCMA .