Abstract: The Radio Map (RM) describes the spatial field distribution and coverage of radio signals in a specific area, which helps optimize the layout of wireless networks and improve the utilization of the spectrum. Since radio wave propagation is complex in densely built urban environments, building high-precision radio maps is a challenge. A convolutional neural network (CNN) model based on the open-source complex urban environment radio propagation simulation dataset is proposed. It combines the use of vanilla convolution and differential convolution to construct an accurate radio maps from sparse spatial field strength of radio signals. The differential convolution is used to extract high-frequency information of radio propagation, while the vanilla convolution provides a global feature representation. Furthermore, a content-guided attention mechanism is introduced to enable the model to focus on the key areas of each channel, thereby improving the accuracy of the map estimation. Experimental results show that the proposed method has a higher accuracy and stability in radio map construction than the existing baseline methods, with an average RMSE reduction of 13%. It also shows stronger robustness in noisy environments. This method has built high-accuracy and stable radio maps in urban environments with dense buildings, which has potential application value for wireless network optimization and radio regulation.