Natural fractures conduct fluids in subsurface reservoirs. Quick and realistic predictions of the fracture network organization and its fluid flow efficiency from limited amount of data is critical to optimize resources productivity. We recently developed a method based on multiple point statistics (MPS) technique to produce geologically-constrained fracture network simulations. The method allows to account for the intrinsic non-stationarity of these networks by considering a multivariate input data instead of averaged distribution of fracture parameters. In addition, the method considers probability maps reflecting the influence of fracture drivers in the network variability. Consequently, the simulated fracture networks derived from the innovative MPS approach are geologically better constrained than in classical discrete fracture network modelling approaches. This paper proposes to apply this method in subsurface conditions where available data are sparsely distributed. We developed a workflow where data are gathered from wellbore and from additional sources (outcrops). These data are used to extrapolate a network around the borehole as training images and themselves are extrapolated at the reservoir scale following a geological probability map.This work also presents innovations on the way how training images and probability maps that may integrate more geology constrain than relying almost entirely on available data.