Uncertainty-Aware Radio Map Reconstruction Using Diffusion Models With Sparse Noisy Samples

Abstract

In large-scale wireless systems, reconstructing radio environment maps (REMs) from sparse measurements is critical for resource management and spectrum allocation. However, practical deployments are often confronted with the challenges of coarse environmental information and noise-contaminated sparse measurements. To address these issues, this letter proposes a diffusion-model-based probabilistic inference framework to enhance the robustness of REM reconstruction under noisy and sparsely sampled conditions. Specifically, a multimodal feature fusion network is designed to incorporate coarse building and vegetation information during the denoising process, thereby enabling more accurate noise estimation and improving reconstruction performance. Furthermore, to mitigate the impact of measurement noise, a posterior-likelihood-based sampling strategy is introduced, which explicitly combines the observation term with the prior during the diffusion sampling process. This allows the model to generate higher-quality samples and improves robustness against measurement noise. Experimental results demonstrate that the proposed method significantly outperforms baseline approaches in terms of REM reconstruction accuracy, environmental awareness capability, and overall robustness.