Abstract: To enable rapid and accurate path loss prediction in complex urban environments, this paper proposes a U-Net architecture guided by prior information from the geospatial electromagnetic characterization map (GEC-Map), termed GEC-UNet. The proposed model improves the basic U-Net architecture by introducing GEC-Map to effectively enhance its representation capability for complex environments, and embeds the convolutional block attention module (CBAM) to further strengthen the selective extraction of key geospatial electromagnetic features. Experimental results demonstrate that the root-mean-square error and mean absolute error of the proposed GEC-UNet on the test set are 4.52 dB and 3.29 dB, respectively. Compared with the basic U-Net, the prediction accuracy is improved by 23.37%. Compared with the conventional ray-tracing method, the prediction time is reduced by more than 90%, yielding significantly higher computational efficiency. The proposed model can effectively fuse electromagnetic environmental prior information, greatly improving path loss prediction accuracy while maintaining high computational efficiency. It provides a new technical approach for the application of GEC-Map in large-scale, high-precision radio wave propagation modeling.