Explainable AI for Spectral Analysis of Electromagnetic Fields

This study explores the application of Explainable Artificial Intelligence (XAI) techniques to the spectral analysis of electromagnetic field (EMF) measurements conducted across various frequency bands in Cyprus. Using an extensive data set obtained from mobile telephony, radio broadcasting, and terrestrial digital television sources, this work aims to capture real-world EMF exposure levels spanning frequencies from 30 MHz to 6 GHz. The frequency bands analyzed include FM radio (87.5-108 MHz), VHF, UHF television, and various mobile telephony zones (700 MHz, 800 MHz, 900 MHz, 1800 MHz, 2100 MHz, 2600 MHz, and 3600 MHz), representing both traditional and emerging wireless technologies. A comparative evaluation of six machine learning algorithms was conducted: XGBoost, LightGBM, Random Forests, k-Nearest Neighbors, Neural Networks and Decision Trees to assess prediction performance across each frequency band. Furthermore, SHAP (SHapley Additive exPlanations) was employed to elucidate the contribution of spatial and demographic characteristics to the intensity of the EMF. The results show that ensemble tree-based methods, particularly Random Forests and LightGBM, consistently outperformed simpler models in accuracy and interpretability. The integration of SHAP enabled transparent feature attribution, revealing distinctive exposure patterns linked to urban topology, population density, and built environment metrics. This approach enables data-driven EMF exposure mapping, aiding urban infrastructure planning, compliance monitoring, and public health risk evaluation.