Ad-SACNet: Spatial Attention Compression Network for Atmospheric Duct Inversion With Radar Sea Clutter Data

As an abnormal atmospheric phenomenon, atmospheric duct makes electromagnetic waves deviate from the original orbital direction and cause radar detection blind spots. Therefore, accurate atmospheric duct prediction can take corresponding measures to correct and compensate the duct structure in time and provide radar air traffic control warning. However, when accounting for the inhomogeneous variations of atmospheric ducts, high-dimensional atmospheric duct profile modeling is required, and such high-dimensional inversion involves enormous computational complexity and consequently impairs the practical inversion speed. Additionally, deep learning-based atmospheric duct inversion models, such as long short-term memory (LSTM) and convolutional neural network (CNN), fail to fully leverage information regarding the spatial inhomogeneous variations of atmospheric ducts. Thus, this study proposes a framework named Ad-SACNet, which inverts the variations of the atmospheric duct modified refractive index profile (M-profile) from the perspective of accounting for the high-dimensional and complex variations of atmospheric ducts. Specifically, in our method, the high-dimensional atmospheric duct M profile is first spatially compressed to reduce the computational complexity of the inversion model. Then, three multi-scale modules are operated to mine spatial inhomogeneous features. Finally, the output of the three modules is integrated by the fusion module to fully exploit the spatial inhomogeneity of the atmospheric duct, which are exploited to attain the final atmospheric duct inversion results. Based on radar simulation and measured data, this paper comprehensively evaluates and compares the proposed Ad-SACNet. The accuracy of the inversion obtained for simulated and measured radar sea clutter by Ad-SACNet is better than the most advanced baseline.