Hybridised CNN approach for pathloss modeling: A comparative study with traditional and machine learning based pathloss models

Abstract

This study compares empirical, machine learning-based, and hybridised Convolutional Neural Network (CNN) based pathloss models. Field measurements of pathloss (dB), altitude (m), latitude, and longitude were measured from 9,217 data points along four major routes covering urban and suburban areas in Ilorin, Kwara State, Nigeria. The hybridized CNN-based pathloss model, integrating the architectural strengths of DenseNet and ResNet, was trained using 70% of the dataset, while 15% each was allocated for evaluation and testing. Training and simulations were performed on Google Colaboratory, leveraging GPU and TPU resources for computational efficiency. The results demonstrate that the hybridized CNN-based pathloss model outperforms empirical models and conventional machine learning approaches, achieving the lowest prediction of Mean Square Error (MSE) and Root Mean Square Error (RMSE) (MSE: 7.35, RMSE: 8.295) and the highest coefficient of determination (R^² = 0.80364). Random Forest and Extreme Gradient Boosting (XGBoost) models followed in performance, while the Transformer model exhibited the highest errors and lowest accuracy. These findings confirm the robustness and accuracy of the hybridized CNN-based pathloss model in enhancing path loss prediction, offering a more reliable approach for mobile network planning and optimization.