ST-CSFNet: Spatial-Temporal Cross Scale FNet for Radio Environment Map Forecasting

Abstract

Radio environment maps (REMs) forecasting is critical for wireless communication, yet existing methods have notable gaps, hindering efficient, reliable, and secure spectrum utilization. This letter introduces an innovative network, ST-CSFNet, that addresses these limitations, particularly their dependence on radiation source information and their failure to capture spatiotemporal dynamics. ST-CSFNet enhances REM forecasting accuracy through multiscale temporal analysis, capturing sequence dynamics across intervals, and by assessing spatial correlations, it accurately extracts spatial characteristics at various granularities. Specifically, it reduces MAE by 37.6%, RMSE by 29.5%, and increases $\mathrm {R}^{2}$ by 6.3%, showcasing significant advancements in REM forecasting accuracy.