The precision of slope stability assessment is highly affected by the randomness of soil parameters. Massive groups of soil parameters and slope geometry parameters are randomly generated by Latin hypercube sampling method according to their common distribution characteristics. For each group of parameters, safety factor is calculated by the strength reduction finite element method (SRFEM); and failure probability with consideration of the spatial variation of soil properties is investigated by combining Monte Carlo simulation and SRFEM under the two-dimensional random field. The sample data and corresponding safety factors and failure probabilities are then implemented in the training and testing processes of radial basis function (RBF) neural network to establish forecast models for slope stability analysis. The simulation results of an example show that the two-dimensional random field model can reasonably well reflect the spatial variation of soil properties; and the created RBF neural network-based forecast models not only has high prediction precision on safety factor and failure probability, but also can effectively save the computational time.