U-Net for Indoor Pathloss Prediction from Sparse Measurements with Physics-Informed Features

Khoren Petrosyan, Hrant Khachatrian, Rafayel Mkrtchyan, Theofanis Raptis

Keywords: Pathloss radio map prediction, deep learning, sparse measurements, ASPP, U-Net, physics-informed features, electromagnetic propagation, indoor wireless networks

TL;DR: U-Net with physics-informed features achieves 5.17 dB RMSE for indoor pathloss prediction from 0.02-0.5% sparse measurements in MLSP 2025 competition.

Abstract: This work was conducted in the context of the MLSP 2025 Sampling-Assisted Pathloss Radio Map Prediction Data Competition. We propose a physics-aware feature engineering approach combined with a U-Net architecture featuring ResNet-34 encoder and Atrous Spatial Pyramid Pooling (ASPP) module to reconstruct indoor pathloss maps from extremely sparse ground-truth samples (0.02\% and 0.5\% sampling rates). Our method transforms the three-channel input into eight physics-informed channels incorporating free-space pathloss, cumulative transmittance losses, log-distance from the antenna, and a binary padding mask. Through a combination of geometric augmentations and multi-scale feature extraction, we achieve competitive performance across both uniform and strategic sampling scenarios. The model attains a weighted RMSE of 5.17 dB, while maintaining inference times of approximately 100ms per map including preprocessing—orders of magnitude faster than ray-tracing baselines.

Submission Number: 8