Coal seam gas (CSG), an unconventional resource of energy, is gaining global interests due to its natural abundance and environmental benefits in comparison to more traditional energy sources. Its production is mainly controlled by the underlying fracture network of coal, called "cleats". The discrete fracture network (DFN) model is widely applied for characterising fracture networks due to its ability of accounting for physical geological parameters of cleats. However, common discrete fracture network (DFN) models fail to preserve the local heterogeneity by assuming planar and smooth cleat surfaces, while real cleats have rough walls and variable opening apertures. This paper aims to characterise the roughness of cleat surfaces at the core scale by developing a novel framework: rough-walled discrete fracture network (RW-DFN) model. The model integrates the pore-scale roughness obtained from micro-computed tomography (micro-CT) imaging of coals into the discrete representation of fracture networks. Analysis of the fracture surfaces obtained from micro-CT imaging demonstrates random, isotropic surfaces following a Gaussian distribution. RW-DFN gives lower permeability than that of the traditional DFN by up to 30%, and its permeability estimation is more accurate with significantly fewer errors (6.5%) than traditional ones (25.1%). This indicates that, to be able to characterise these reservoirs, traditional DFN models may over-estimate the production while our proposed RW-DFNs provide more deterministic results. Overall, the method applies micro-CT imaging to obtain the internal surfaces of coal fractures in a non-destructive manner and reconstruct representative RW-DFN models. The developed RW-DFN models are not restricted by the imaging resolution, so that they are favourable for direct numerical simulation of permeability. In addition, RW-DFN models can be constructed with extended domain size, so they can be incorporated into existing reservoir characterisation frameworks for the prediction of coal properties. (C) 2016 Elsevier Ltd. All rights reserved.