Spatio-Temporal Reconstruction for Dynamic Radio Maps with Sparse Sampling

Abstract

Dynamic spectrum access and network optimization require accurate real-time radio maps in urban environments. However, high-density measurements are impractical, and extremely sparse sampling leads to poor reconstruction, particularly when mobile transmitters cause hotspot drift. This paper proposes a spatio-temporal reconstruction framework for dynamic radio maps under sparse sampling that couples dictionary-driven information efficiency maximization (DIEM) sampling, sparse Bayesian learning (SBL), and convolutional long short-term memory network (ConvLSTM).DIEM sampling maximizes the information efficiency of limited observations via random projection and Sherman-Morrison updates. SBL with Gaussian process (GP) residual correction recovers the spatial structure from sparse samples. Crucially, a ConvLSTM network learns the spatio-temporal evolution patterns from historical reconstructed frames to generate a temporal prior for the current frame, which is then fused with real-time sparse measurements to refine reconstruction accuracy. Experiments in the Munich urban scenario demonstrate that the proposed method effectively tracks dynamic hotspot drift at sampling rates as low as 0.1%, substantially outperforming pure spatial reconstruction baselines.