Automatic Scaling of Vertical Ionograms Based on Generative Adversarial Network

Abstract Ionospheric vertical sounding is a well‐established and widely used ground‐based technique for ionospheric detection. Efficient and accurate automatic scaling of ionograms is crucial for real‐time applications of vertical sounding. However, current scaling methods face challenges in achieving precise trace segmentation due to noise, interference, and various ionospheric disturbances. This paper proposes a deep learning model for ionogram automatic scaling based on generative adversarial network (GAN), which is named IASGAN. The model integrates a 50‐layer residual network (ResNet50) and a feature pyramid network (FPN) as the generator, with a multi‐layer convolutional neural network (CNN) as the discriminator. Given a vertical ionogram, the generator produces segmentation result that closely resemble the corresponding label, while the discriminator provides feedback loss to the generator for adversarial training, thereby enhancing the segmentation performance of the generator. Experimental results demonstrate that the IASGAN model can precisely and effectively autoscale E, F1, and F2 layer traces. Compared to existing scaling methods, the IASGAN model can produce finer trace extraction from ionograms, with a mean maximum critical frequency absolute deviation (D‐MCF) of 0.0803 MHz and a mean minimum virtual height absolute deviation (D‐MEH) of 5.8205 km. This capability can provide technical support for the extraction of characteristic parameters and ionospheric inversion, which is significant for real‐time acquisition of ionospheric characteristics and structure information.