Spatiotemporal MUF Inference for HF Communications: A Personalized Federated Learning Framework

Abstract

High-frequency (HF) communications are vital in both military and commercial fields due to their ability to propagate across the horizon. Accurately predicting the maximum usable frequency (MUF) is crucial for improving the efficiency of HF ionospheric propagation. However, when there are only a limited number of ground-based stations for vertical ionospheric sounding, it becomes challenging to infer and predict MUFs in unfamiliar territories. In this article, we present a personalized federated learning framework that utilizes mutual information to make spatiotemporal inferences for MUFs across multiple stations. To begin with, we extract the trend component using time-series decomposition theory and then verify its significant spatial correlation using dynamic time warping (DTW). This enables us to identify the spatial relationships present in the local models of the federated learning system. Next, by monitoring changes in mutual information among different local models during the training phase of the trend component, we verify the existence of these spatial relationships. Furthermore, we introduce and validate the effectiveness of the personalization strategy employed in the proposed framework. Finally, experimental results demonstrate that the proposed framework significantly reduces the root mean square error (RMSE) and mean absolute error (MAE) in MUF predictions for ten cities with varying geographic latitudes.