IonoGAN: An Enhanced Model for Forecasting Quiet and Disturbed Ionospheric Features From Predicted Ionograms

Abstract Ionograms are radar echo graphs that depict vertical ionospheric density profiles, structures, fluctuations, and irregularities, with the F region represented by F‐trace and Spread‐F features in the graphs. In this paper, IonoGAN, an enhanced neural network based on the Generative Adversarial Network architecture, is proposed for direct prediction of ionograms and the variation of these ionospheric conditions. This estimation is based on the trends of density profiles and the waves/structures presented in the ionogram sequence. The IonoGAN extends the spatiotemporal information‐preserving and perception‐augmented (STIP) ability by incorporating a Local‐Global discriminator to focus on the F region in ionograms. In addition, two scientific characteristics of ionospheric natural phenomena are extracted and used as constraints in the modeling: Spread‐F Classification Accuracy (SFCA) and Absolute Value of the Correlation Coefficient for the F trace (AVCC‐F). For training, ionograms from Hainan Fuke station (19.5°N, 109.1°E, magnetic 11°N) during 2002–2015 were processed into 36,435 sequences with Spread‐F phenomena and 147,147 sequences without. To strengthen their features, Spread‐F phenomena were further classified into types of frequency, range, mix, and strong range. After the parameter training, the IonoGAN achieved SFCA and AVCC‐F converging to their optimal values: on the 2016 test set, SFCA = 90.92%, AVCC‐F = 0.6917. This modification enables the network to effectively capture the distinct features of the ionospheric F trace and the Spread‐F phenomenon during both quiet and disturbed periods.