A new back-analysis method of ground stress is proposed with comprehensive consideration of influence of to- pography, geology and nonlinear physical mechanical properties of rock on ground stress. This method based on non- uniform rational B-spline （NURBS） technology provides the means to build a refined three-dimensional finite element model with more accurate meshing under complex terrain and geological conditions. Meanwhile, this method is a back- analysis of ground stress with combination of multivariable linear regression model and neural network （ANN） model. Firstly, the regression model is used to fit approximately boundary loads. Regarding the regressed loads as mean value, some sets of boundary loads with the same interval are constructed according to the principle of orthogonal design, to calculate the corresponding ground stress at the observation positions using finite element method. The results （ boundary loads and the corresponding ground stress） are added to the samples for ANN training. And on this basis, an ANN model is established to implement higher precise back-analysis of initial ground stress. A practical application case shows that the relative error between the inversed ground stress and observed value is mostly less than 10 % , which can meet the need of engineering design and construction requirements.