Advanced Deep Learning Tool for Global Total Electron Contents Map Inpainting

Abstract:

Total electron content (TEC), as a measure of number of column electrons in the ionosphere, is routinely measured by Global Navigation Satellite System (GNSS) receivers. However, the limited coverage of GNSS causes approximately 52% missing data on the global TEC maps. The recovery of the missing TEC values can be treated as an image inpainting task, for which various deep learning algorithms have been developed. In this work, we adapt the spectral normalized patch generative adversarial network (SN-PatchGAN) with contextual attention for TEC map inpainting. Eighteen-year international GNSS Services (IGS)-TEC maps are used to train SN-PatchGAN. SN-PatchGAN shows much improved accuracy on recovering the missing TEC data, with more than 50% reduction on root mean squared errors for some important storm events, compared to our previous inpainting method based on deep convolutional GAN (DCGAN). SN-PatchGAN with the capability of the end-to-end inpainting significantly saves inpainting time by avoiding the iterative input mapping and Poisson blending that are necessary in the DCGAN inpainting method for good performance. This development shows the great potential of deep learning methods for automatic and accurate data completion for TEC maps.