Advances in spatial transcriptomics enlarge the use of single cell technologies to unveil the expression landscape of the tissues with valuable spatial context. However, computational tools developed for single-cell transcriptomics have great limits in dealing with spatial transcriptomic data with high noise on detected transcript signals. Here we propose an unsupervised and manifold learning-based algorithm, STEEL, which identifies different cell types from spatial transcriptome by clustering cells/beads exhibiting both highly similar gene expression profiles and close spatial distance in the manner of graphs. Comprehensive evaluation of STEEL on various spatial transcriptomic datasets from 10X Visium platform demonstrates that it not only achieves a high resolution to characterize fine structures of mouse brain, but also enables the integration of multiple tissue slides individually analyzed into a larger one. STEEL outperforms previous methods to effectively distinguish different cell types of various tissues on Slide-seq datasets, featuring in higher bead density but lower transcript detection efficiency. Application of STEEL on spatial transcriptomes of early-stage mouse embryos (E9.5 to E12.5) successfully delineates a progressive development landscape of tissues from ectoderm, mesoderm and endoderm layers, and futher profiles dynamic changes on cell differentiation in heart and other organs. With the advancement of spatial transcriptome technologies, our method will have great applicability in high-resolution cell type identification and unbiased spatiotemporal data integration.