Improving Path Loss Prediction Using Environmental Feature Extraction from Satellite Images: Hand-Crafted vs. Convolutional Neural Network

There is an increased exploration of the potential wireless communication networks in automation daily human tasks via Internet Things. Such implementations are only possible with proper design networks. Path loss prediction a key factor parameters such as cell radius, antenna heights, and number sites that can be set. As path affected by environment, satellite images network locations used developing models environmental effects captured. We developed model based on Extreme Gradient Boosting (XGBoost) algorithm, whose inputs numeric (non-image) features influence extracted from composed four tiled points along transmitter to receiver path. The predict for multiple frequencies, environments it incorporated into Radio Planning Tools. Various feature extraction methods included CNN hand-crafted their combinations were applied order determine best input features, which, when combined non-image will result XGBoost model. Although have advantage not requiring large volume data no training involved them, they failed this application use led reduction accuracy. However, was obtained image using GLCM resulting RMSE improvement 9.4272% against without images. performed better than Random Forest (RF), Learning Trees (ET), Boosting, K Nearest Neighbor (KNN) combination CNN, GLCM, features. Further analysis Shapley Additive Explanations (SHAP) revealed had highest contribution toward model’s output. variation values output presented SHAP summary plots. Interactions also observed between some dependence plots computed values. This information, further explored, could serve basis development explainable/glass box