Atmospheric duct (AD) is a natural phenomenon affected by humidity or air pressure diversities, which impacts the propagation of electromagnetic signals. Recent studies have been focused on ADH prediction based on meteorological data which makes the task of real-time ADH prediction challenging and impacts the global representation of ADs leading to maritime communications security issues. We address these issues by discovering a novel way to abstract the spatial and temporal features from the historical data. Thus, we proposed a framework based on DNN models and attention mechanisms to predict the ADH in a real-time form from the abstracted features. Our experimental results demonstrate that our technique improves the performance of real-time prediction and reduces the acquisition cost, professional devices can receive it in real-time through IoT to avoid dangerous areas.