Abstract: To address the challenges of accurately reconstructing stratigraphic structures and quantifying uncertainty under sparse borehole data conditions in tunnelling engineering, a stratigraphic modeling method coupling multi-source feature-driven learning with a Markov random field was proposed. Multi-source features were constructed based on water content, plasticity index, and liquidity index, and a data-driven model was employed to obtain the probability distribution of stratigraphic classes. Anisotropic spatial interpolation was then applied to reconstruct continuous feature fields. On this basis, a Markov random field was introduced to optimize the stratigraphic structure with spatial constraints, and information entropy and classification margin were adopted to quantify the uncertainty of prediction results. A sparse-borehole sampling case based on engineering geological data from an open-cut section of Shenzhen Metro Line 20 was constructed for validation. The results show that the proposed method effectively reconstructs complex stratigraphic structures under sparse borehole sampling conditions, achieving an overall classification accuracy of 90.97%. Misclassifications were mainly concentrated near stratigraphic interfaces and could be effectively identified using the uncertainty indicators. It is concluded that the method improves both spatial continuity and classification accuracy, while enabling effective characterization of spatial uncertainty, providing reliable support for geological analysis and construction risk assessment in tunnelling engineering.