Neural Fitting for Sparse Radio Map Construction in LEO Satellite Network

Abstract

Recently, Low Earth Orbit (LEO) satellite networks (i.e., non-terrestrial network (NTN)), such as Starlink, have been successfully deployed to provide broader coverage than terrestrial networks (TN). Due to limited spectrum resources, TN and NTN may soon share the same spectrum, calling for fine-grained spectrum monitoring to facilitate cooperative frequency usage and interference mitigation. To this end, constructing a 4D radio map (RM)—encompassing three spatial dimensions and signal spectra—is crucial. However, such construction normally requires a massive sensor deployment and high-speed analog-to-digital converters for extensive spatial signal collection and wide power spectrum acquisition, posing significant practical challenges. In this work, we present DeepRM, a deep unsupervised learning framework that requires no ground-truth labels and leverages both neural compressive sensing (CS) and tensor decomposition (TD). First, we recast the CS process as a neural network training problem and devise a sparsity–performance balancing mechanism to reconstruct wideband spectra from sub-Nyquist samples, thereby obtaining the spectral information for each frequency. Second, a neural TD framework, jointly equipped with a lightweight Gaussian smoothing scheme, is employed to construct 3D RMs across the spatial domain for all frequencies, even with highly sparse sensor deployment. Extensive evaluations demonstrate that DeepRM achieves notably lower reconstruction error than corresponding state-of-the-art baselines, especially under limited sampling conditions. Beyond 4D RM construction, we further validated that the resulting 4D RM further supports advanced function for dynamic spectrum sharing, making it well suited for emerging LEO satellite networks.