Global Digital Elevation Models (GDEMs) are one of the most important sources of elevation data. In recent years, GDEMs have become increasingly popular with researchers due to their global coverage and free accessibility. The most commonly used GDEMs are AW3D, ASTER, and SRTM. Each of these models is produced by different technologies and have different strengths and weaknesses. This issue indicates that these data are not necessarily consistent with another, and their accuracy is dependent on the local topography of the earth. The main objective of this research is to fuse global digital elevation models to produce a model with higher vertical accuracy. In this regard, in this study, a two-step approach is proposed for fusing GDEMs. In the first step, a Geographically Weighted Regression (GWR) model is used to determine areas of the Earthchr('39')s surface that have similar properties. In other words, using the GWR model, regions of the study areas with similar behaviors are classified into the same classes. At this step, each of these study areas is classified into three, five, and seven classes. Among these modes, for both study areas, the best results are for five Class mode. In the second step, to fuse GDEMs, the optimum weight of each class defined for each of AW3D, ASTER, and SRTM models are estimated using the particle swarm optimization (PSO) algorithm. In order to evaluate the accuracy of the proposed method, it has been used to produce the fused DEM for two study areas of BumeHen and TazehAbad. In the first case study (BumeHen), the amount of Root Mean Square Error (RMSE) on test points in five class mode for AW3D, ASTER and SRTM are 4.58, 8.69 and  4.70 meters respectively, while it’s 3.97 meters for fused DEM. In the second case study (TazehAbad), the amount of RMSE on test points in five class mode for AW3D, ASTER, and SRTM are 3.33, 7.31, and 3.17 meter respectively and it’s 2.74 meters for fused DEM. The results show that the proposed method is capable of producing a higher accuracy model than any of the initial models by utilizing the potential of each of these input models in the fusion process.