Deep Learning-Based Path Loss Prediction Using Side-View Images in an UMa Environment

Several models using deep neural network (DNN) have been proposed to estimate the path loss characteristics in the urban macrocell (UMa) environment, most of which use the height of buildings around Tx and Rx as input images to the convolutional neural network (CNN) part. However, buildings between Tx and Rx affect the path loss characteristics more than the buildings surrounding Tx and Rx, since the over-rooftop propagation becomes dominant when a Tx antenna is installed higher than the surrounding buildings. In other words, the DNN in these models does not consider the propagation paths over rooftops. If such paths were considered, it would be possible to improve the estimation accuracy. In this work, we propose a new model that uses the side-view image of a building on the straight line between Tx and Rx as input, in addition to the conventional top-view image. We then evaluate the effect using measurement data, the RMS error of the proposed model was 5.63 dB, while that of the conventional model was 9.46 dB.