A robust learning framework for spatial-temporal-spectral radio map prediction

Spectrum map prediction, as one of the key technologies for spectrum situation generation, has garnered increasing attention. Although significant efforts have been devoted to spectrum map prediction, most existing spectrum map prediction schemes focus on the spatial-temporal domain (single frequency spectrum map prediction) with ideal spectrum measurements, neglecting inherent spectral correlations, missing data, and outliers. This work aims to achieve accurate spatial-temporal-spectral spectrum map prediction with incomplete and corrupted observations. A two-stage learning framework integrating deep learning and tensor completion is designed to address the above challenges. Specifically, we first model the spectrum map as a fourth-order spectrum tensor to fully exploit the inherent spatial-temporal-spectral structures of the spectrum data. Second, a developed Transformer with forget-sparse interpolation attention mechanism is employed to fill in partial values of the future spectrum map. Finally, we propose a novel online spectrum map prediction algorithm that integrates the Alternating Direction Method of Multipliers (ADMM) and Recursive Least Squares (RLS). Validated on real-world spectrum measurements, our proposed framework has a significant advantage over the state-of-the-art baselines.