Efficient and approximate evaluation of urban coverage is important for wireless network planning. While standard statistical propagation models are fast, they do not directly describe the physical environment of a specific urban scene and consequently often fail to accurately capture local blockage and site-specific propagation effects. Ray tracing can model these effects more directly, but becomes costly when testing many tiles, frequencies, and transmitter heights simultaneously. To address this problem, the present study investigates the use of an RT-supervised simulation-to-simulation tile-based learning framework for path loss prediction based on geographic environmental information. This methodology first builds realistic 3D city scenes from geographic data, then uses offline ray tracing to generate supervision labels across multiple carrier frequencies and base-station heights. Each city region is divided into 500 m by 500 m tiles, which are then further discretized into 125 by 125 grids. For each tile, raster priors, such as occupancy, normalized height, and a valid-ground mask, are prepared. During training and inference, the model input is organized as an 8-channel raster tensor together with a 2D condition vector for frequency and transmitter height. The raster tensor combines three stored environment priors and five online-generated transmitter-related feature maps. By utilizing masked supervision, the network learns the excess loss residual exclusively on valid outdoor pixels, and the final path loss map is reconstructed by combining the residual prediction with the FSPL prior. The final model in this work was trained on 134,317 samples and validated on 33,589 samples. In the in-city setting, used as a preliminary verification before subsequent cross-city experiments, it achieved an MAE of 5.0116 dB and an RMSE of 9.3182 dB. On the formal cross-city test with a completely unseen target city, it achieved an MAE of 4.8536 dB and an RMSE of 9.3504 dB. These results demonstrate that the proposed framework can provide a stable tile-level approximation of RT-generated path loss maps under multiple conditions. Because both training labels and evaluation references are generated by RT rather than drive-test measurements, the present study should be understood as a simulation-to-simulation surrogate framework rather than a direct validation of real-world propagation accuracy.
Huai et al. (Tue,) studied this question.