ADH-CLSNet: Cross-Latitude Spatiotemporal Network for Atmospheric Duct Heights Diagnosis

Timely estimation of atmospheric duct height is of great importance for maritime electromagnetic propagation analysis, collision avoidance, and navigation safety management. However, existing atmospheric duct height methods mainly focus on single-station observations, making large-scale global estimation difficult due to the sparse distribution of atmospheric duct samples and the complexity of data collection. In addition, existing deep learning-based methods, such as Long Short-Term Memory (LSTM) and Particle Swarm Optimization (PSO)-LSTM, have limited capability in modeling the spatiotemporal correlations of atmospheric duct variations. To address these issues, a novel cross-latitude spatiotemporal framework, termed ADH-CLSNet, is proposed. First, considering the high computational complexity of deriving atmospheric refractivity and duct height from ERA5 temperature, humidity, and pressure data, an end-to-end lightweight mapping model is designed to directly estimate atmospheric duct height from raw meteorological variables. Furthermore, to characterize the spatial heterogeneity of atmospheric ducts across different latitudinal regions, the global atmospheric duct data are divided into low-, mid-, and high-latitude zones for region-specific feature extraction. Dedicated cross-latitude spatiotemporal modules are then constructed to capture temporal dependencies and spatial correlations within different latitudinal regions. Finally, a spatiotemporal fusion module is introduced to reconstruct global atmospheric duct height distributions from the outputs of the three regional branches. Experimental results on global ERA5 datasets demonstrate that the proposed ADH-CLSNet achieves superior performance compared with conventional and deep learning-based baseline methods, effectively capturing the spatiotemporal characteristics of atmospheric duct variations while significantly improving global atmospheric duct height estimation accuracy.