Sparse-Guided RadioUNet with Adaptive Sampling for the MLSP 2025 Sampling-Assisted Pathloss Radio Map Prediction Data Competition

Ryoichi Kojima, Satoshi Ito, Tatsuya Nagao, Masato Taya

Keywords: path-loss prediction, sparse supervision, U-Net, adaptive sampling, frequency transfer learning

TL;DR: A 5-channel Sparse-Guided RadioUNet leverages reflection, transmission, distance, and ultra-sparse samples to predict indoor path-loss maps with high accuracy in milliseconds.

Abstract: The MLSP 2025 “The Sampling-Assisted Pathloss Radio Map Prediction Data Competition” challenges participants to construct accurate path-loss maps from building-scale ray-tracing data while leveraging sparse ground-truth PL samples along with physics-informed inputs such as reflection, transmission, and distance maps. Each submission is evaluated under four conditions—Task 1 and Task 2, each tested with 0.50% and 0.02% sampling budgets—and the final score is computed as a weighted average of the RMSEs across these four settings. We address this challenge with Sparse-Guided RadioUNet, a five-channel U-Net architecture that integrates physicsinformed inputs and sparse supervision. The model excludes sampled pixels from the loss computation, allowing it to focus on interpolating unknown regions. A two-stage frequency transfer learning routine—multi-frequency pretraining followed by fine-tuning on the target band—enables the network to generalize across propagation conditions while adapting to band-specific characteristics. To further regularize spatial consistency, we introduce a spread loss term that penalizes discontinuities, though it may occasionally oversmooth sharp transitions. For Task 2, we propose a lightweight edge–strata–boundary (ESB) sampling algorithm that combines saliency, distance strata, and LoS/NLoS boundary cues to place informative probes under a tight sampling budget. Our experiments demonstrate that this framework enables high-quality PL map prediction with as little as 0.5% or 0.02% supervision. The full pipeline runs in 2.5 ms per building—orders of magnitude faster than raytracing—and achieves strong validation accuracy without relying on dense ground truth. These findings highlight the potential of structured sparse learning and efficient sampling in scalable radio frequency environment modeling.

Submission Number: 3