Accurate Radio Map Estimation from Sparse Measurements with U-Nets

Abstract

This paper presents a deep learning approach for predicting cellular signal strength maps in urban environments using a U-Net architecture. The network processes building and antenna configuration data as images and optionally integrates sparse measurements through an additional calibration channel. Using only 5% of available ground-truth data reduces the root-mean-square error (RMSE) by 1.2 dB compared to predictions without measurements. The model remains robust under measurement noise up to 12.9 dB and consistently outperforms interpolation and k-nearest-neighbor regression. The proposed approach offers a scalable, accurate, and computationally efficient solution for radio map estimation and network planning.