Evaluating the impact of tunnelling on above ground structures in urban areas highly relies on prediction of the settlement trough at surface level. Uncertainties in soil properties and other system characteristics for soft ground (e.g. Alluvium soil) can significantly affect the prediction. Classically, empirical formulas are used for prediction of the settlement using the experience from previous tunnelling projects. Empirical formulas do not consider the soil-lining interaction or the method of construction and lack a theoretical background for ground movement in continuum mechanics. Analytical and numerical approaches have developed to address some of these deficits, but have not to date, taken account of the many uncertainties involved in the process. This thesis introduces a stochastic method in the context of discrete random finite element theory to probabilistically predict the tunnelling-induced surface settlement. The method proposes a single mechanism (i.e. an uncertainty and sensitivity analysis framework) in which multiple sources of uncertainty can be considered within a single model (e.g. heterogeneity of the soil profile, variability in surcharge loads, and material properties). The power of the method is then examined through application to case studies involving two large-scale, shallow tunnelling projects excavated in alluvium soil. The results are compared with monitoring data, with estimations from deterministic finite element (FE) models and empirical formulas. Compared to the results of classical FEs and empirical approaches, application of the new probabilistic approach provides better understanding of the development of the settlement trough at surface level in both case studies. The output parameters for the tunnelling-induced settlement trough (volume loss and maximum settlement) are in good agreement with actual monitoring data in both case studies. Notably, the prediction result of the empirical formula in the first case study is conservative and for the second one is quite non-conservative. The new approach, on the other hand, provides more accurate predictions and insights into the effect of different sources of uncertainty.