Spatial transcriptomics is essential for mapping the developmental spatial organization of tissues and organs. However, accurately integrating multiple datasets (slices) to capture detailed spatial-temporal dynamics and cell interactions remains challenging due to the neglect of slice-specific features in conventional methods. Here, we introduce SpaGA(Spatial Graph Alignment), a novel approach for the precise alignment of spatial transcriptomic slices. SpaGA harnesses the power of Singular Value Decomposition (SVD) matrix decomposition to distill informative low-dimensional features from the multi-slice datasets. Graph neural networks then process the features to learn disentangled representations, distinguishing common and private slice representations. The common representations are further aligned to mitigate batch effects across slices through adversarial learning, enhancing the accuracy of cell type and spatial region matching. Our comprehensive evaluation using 10x Visium, MERFISH, and Stereo-seq datasets demonstrates that SpaGA significantly outperforms current methods, highlighting its potential in the effective alignment of spatial transcriptomic data for a deeper understanding of tissue development and architecture. Implementations of SpaGA are available at https://www.dropbox.com/SpaGA.