Spatial transcriptomics enables gene expression profiling while preserving cellular spatial information. However, current technological constraints lead to significant data loss and noise in spatial transcriptomic datasets, presenting substantial challenges for critical analytical tasks including cell clustering, differential gene expression analysis, and spatial domain identification. To overcome these limitations, we developed stMGRL, an innovative deep multi-task masked graph representation learning framework specifically designed for spatial transcriptomics data analysis. stMGRL employs an integrated architecture combining a masked graph autoencoder, a zero-inflated negative binomial (ZINB) distribution module, and a graph contrastive learning framework to address key challenges in spatial transcriptomics data analysis. This unified approach simultaneously improves data imputation accuracy and clustering robustness by maintaining topological consistency while correcting for technical artifacts. Comprehensive benchmarking across multiple spatial transcriptomic datasets demonstrates that stMGRL significantly outperforms existing methods in spatial feature extraction and domain identification accuracy. Our framework represents an effective solution for multi-task integration in spatial transcriptomics, with demonstrated robustness and superior performance across diverse experimental conditions.