Multi-Task Learning for Real-Time Atmospheric Duct Characterization in Wireless Propagation Channels

Atmospheric ducting significantly influences radio wave propagation, especially in maritime and desert regions. Rapid refractive-index variations critically affect signal propagation, especially for next-generation systems operating in the sub-6 GHz and higher frequency bands. Existing atmospheric duct models are often limited in scope, typically focusing on single-output prediction tasks and relying on synthetic or surface-level inputs, which do not fully capture the vertical refractivity structure required for flight-level duct characterization. To address this gap, this paper introduces a multi-task learning (MTL) framework that combines convolutional neural networks (CNNs) with bidirectional long short-term memory (BiLSTM) networks to sequentially detect duct occurrence, classify duct types, and predict essential duct characteristics. The model is trained and evaluated using two years of vertical profile data for a coastal region, with stratified 5-fold cross-validation (CV) and independent hold-out test set. In addition to point-estimate performance, the framework is validated using uncertainty-aware Monte Carlo test-time stochasticity on the test data to quantify predictive uncertainty and assess robustness of duct characterization. Results demonstrate superior duct detection accuracy, reliable duct-type classification, and robust duct-parameter prediction compared with denser MTL-CNN and MTL-LSTM baselines. The proposed framework provides a practical basis for real-time atmospheric duct assessment and has the potential to enhance propagation forecasting for advanced radar and high-frequency communication systems.