Radio environment map (REM) is an efficient enabler for practical cognitive radio networks by sensing the electromagnetic information within regions of interest dynamically. Most of works on Kriging-based method have proven that separate estimation for pathloss and shadowing can obtain more accurate REM construction. But these methods have some shortcomings that prior information is required for construction or disability for multiple transmitters scenario. In order to overcome the problems of urban REM construction mentioned above, this paper propose a residual Kriging algorithm based on generalized regression neural network (GRNN-RK) for that. The performance of proposed algorithm has been evaluated by the analysis of simulation results, and experiments show that GRNN is capable of improving Kriging in accuracy. Additionally, the influence of spread on REM construction is also experimented.