Machine Learning‐Based Radio Environment Map Estimation for Indoor Visible Light Communication

ABSTRACT This paper presents a machine learning‐based methodology for constructing optical radio environment maps (REM) for indoor visible light communication (VLC) systems, explicitly positioned as an empirical benchmarking and evaluation framework. First, the REM estimation is evaluated on configurable simulated datasets generated by a physics‐based indoor VLC simulator, and then the selected method is benchmarked against alternative learning and interpolation‐based methods on experimental measurements. The evaluation of decision trees and multilayer perceptron (MLP) models on simulator RSS datasets for different room geometries and transmitter configurations shows that tuning sample size, epochs and batch size allows MLPs to achieve superior accuracy‐performance trade‐offs. Subsequently, a comprehensive comparison of MLP against kernel‐based regressors, interpolation methods and neural architectures (CNN and U‐Net) reveals that although spatial extrapolation degrades performance, multiple models achieve comparable interpolation accuracy. Specifically, MLP optimally balance accuracy and computational efficiency, training roughly three times faster than convolutional architectures. Furthermore, data augmentation using SMOGN and Gaussian process regression (GPR) was investigated, demonstrating that GPR‐based synthetic data can reduce required real‐world measurements by up to 83% without compromising prediction accuracy. These results confirm the practical viability of data‐driven REMs for indoor VLC applications such as network planning, localisation and digital twins.