Multi-Input Model-Based Path Loss Prediction with Terrain and Environment Patches

This paper proposes a multi-input convolutional neural network model for path loss prediction that integrates terrain and environmental information extracted from digital elevation and land-cover maps. To simultaneously account for the terrain and environmental factors that strongly influence radio propagation, two-dimensional spatial square patches of 64×64 pixels centered at both the transmitter and receiver were generated and rotated such that the propagation path corresponded to the 0° direction and concatenated along the channel dimension. The concatenated patches were then combined with system parameters such as distance, frequency, and antenna height to form the inputs to the model. The proposed model was trained using 3.4 GHz propagation measurement data collected in Hongseong, Chungcheongnam-do, South Korea, and achieved a root mean squared error of 5.8044 dB on the test dataset.