A new wavelet transform and merging generative adversarial network (WTM-GAN) model for TEC spatial inpainting

Abstract

Due to the uneven distribution of ground observatories, the effective data coverage of global ionospheric TEC is below 50%. The International GNSS Service provides a global ionosphere map based on a single shell assumption, derived from the ground-based observations. This serves as the main reference for global ionosphere morphology study. In this work, a new GAN model, wavelet transform and merging generative adversarial network (WTM-GAN) is proposed, designed for spatial completion of ionospheric TEC data with observation coverage deficiency. WTM-GAN is designed with an encoder–decoder architecture, using a Haar wavelet filter and a multilayer decoder employing segmentation and merging techniques. The performance is rigorously tested, achieving root-mean-square errors of 2.117 TECu and 0.908 TECu during both high and low solar activity years, respectively, and it obtains improvement of 0.945 TECu and 0.739 TECu over the comparison models. It also attained a peak signal-to-noise ratio over 32 dB, outperforming all comparisons. During geomagnetic storms, WTM-GAN effectively captures features in the equatorial ionization anomaly region, demonstrating enhanced spatial observation augmentation accuracy and stability. This framework offers a robust solution for TEC data completion, improving the reliability of ionospheric studies.