Multi‐Scale Attention‐Enhanced Deep Learning Model for Ionogram Automatic Scaling

Abstract

AbstractTo increase the accuracy of ionogram automatic scaling, a deep learning model—multi‐scale attention‐enhanced UNet is proposed. Correspondingly, a multi‐scale attention‐enhanced (MSAE) sub‐network is developed which involves a spatial attention nearest up‐sampling module and several residual channel attention modules with multi‐scale skip connections. They contribute to multi‐scale feature fusion and augmentation of the learning ability for enhancing faint and elongated profile traces of ionograms. The MSAE sub‐network input consists of multi‐scale feature maps which could be optimally employed to make the network effectively utilize useful information from the encoders and decoders. Incidentally, a dual channel spatial attention block is embedded between the encoder and the decoder for deeper detail extraction. When the proposed model is applied to scale different electron density profiles of ionograms based on an open data set, the experimental results show the segmentation performance evaluation indexes: the precision and the recall rate can be improved by 6.9% and 26.1%, respectively, compared to automatic real‐time ionogram scaling with true‐height routine. Another set of indexes: the mean intersection over union and the F‐score are superior to that of other several contrasted deep learning models, which can be improved by 3% and 1.6%, respectively, compared to the original UNet model.