A Hybrid Combination of a Convolutional Neural Network with a Regression Model for Path Loss Prediction Using Tiles of 2D Satellite Images

Wireless communications networks require very accurate design, especially for the 5G networks that is an aggregation of various network types in its Heterogeneous Ultra Dense Network (H-UDN) architecture. The limitation is that path loss models with large prediction scope and high accuracy are not available. In this work we developed a novel architecture for a multiple environment and multiple parameter path loss prediction model using a combination of a Convolutional Neural Network (CNN) and a regressor. The CNN extracts features from 2D satellite images and together with some numerical features are trained to a regressor model. Various machine learning algorithms were used as the regressor and their performances evaluated. A least decrease of 1.0262dB in Root Mean Squared Error (RMSE) was achieved by our model, in comparison to a deep learning architecture in which Multiple Layer Perceptron is used in place of the regressor. We also demonstrated that using an image composed of tiles of satellite images of the receiver and transmitter locations, and other points along the path from transmitter to receiver improves results over using the image at the receiver location only.