Enhancing CNN Path Loss Estimation Performance using Satellite Image and Distance Feature

ABSTRACT We employ convolutional neural networks (CNNs) with distance feature and satellite image for path loss (PL) estimation at sub‐6 GHz and millimetre wave (mmWave) frequencies. In order to avoid complex preprocessing of embedding distance feature into the image, we append this feature at the earliest, after the convolutional blocks of a CNN‐based VGG‐16 architecture. This is intuitive since the following fully‐connected (FC) layer performs feature aggregation, thus, it combines the injected distance feature with the extracted features from the image. We propose three VGG‐16 structures which vary in how the distance information is included. Performance is then evaluated in terms of training and prediction times, root mean square error (RMSE) and correlation coefficient, while performance without appended distance serves as benchmark. We observe that the inclusion of distance parameter gives more accurate estimation in terms of RMSE and a very strong correlation between the predicted and estimated PL values. Moreover, the proposed structures typically converge more quickly. Among the proposed structures, the one aided by a logarithm‐of‐distance model, is the most computationally efficient, leading to and reduction of training time and prediction time, respectively. Additionally, the VGG‐16‐based PL predictors yield lower RMSE by up to 2.4 dB and higher correlation compared to the 3GPP 38.901 urban macro cell (UMa) empirical model.