The soil's compression modulus (Es) is one of the most critical mechanical parameters for studying land subsidence in urban strata. Meanwhile, the vertical heterogeneity and lateral discontinuity in the distribution of Es is one of the leading causes of TBM (tunnel-boring-machine) entrapment in the field. In this study, we first obtain the minimum redundancy input variables for Es prediction by comparing 20 purely data-driven models. Based on this, we propose a static directed and weighted graph based on spatial correlation and sequence stratigraphy for mechanism-driven learning, the spatial-geological stratigraphic graph (SG-SPG), where edges encode essential spatial-geological information in direction and weights. Then we built a spatial-geological graph attention network (SGGAT) for node-level estimation and edge-level correction of Es prediction, including two attention layers to adequately capture the dependencies between spatial location and lithology. The results show that our method outperforms state-of-the-art models (R2 = 0.99, RMSE = 0.0593 MPa). Furthermore, through network analysis of SG-SPG, we find that increasing global transitivity significantly improves accuracy, while decreasing local connectivity and degree centrality of nodes slightly increases the prediction residuals. Our approach provides interpretability for graph neural network (GNN) and is advantageous for geophysical property prediction problems with Spatio-temporal heterogeneity.