cGAN-ViT for Dynamic EM Map Representation and Reliable Localization of Distortion Sources

Localization based on electromagnetic (EM) maps in Global Navigation Satellite Systems (GNSS)-denied environments is often undermined by dynamic EM distortions that violate static map assumptions and induce systematic localization bias. This paper addresses this challenge by formulating dynamic EM mapping as the coupled estimation of distortion-region masks and the classification and localization of underlying sources. A physics-driven simulator synthesizes dipole, line-current, and quadrupole sources with sinusoidal amplitude modulation to enable controlled learning across diverse noise and interfer ence conditions. The proposed pipeline integrates a conditional generative adversarial network (cGAN) utilizing a 3D U-Net generator for spatiotemporal mask prediction with a Vision Transformer (ViT) that performs instance-wise source typing and 2-D localization in physical units. On a 128×128 grid covering a 12.7×12.7 m2 area, the end to-end system attains an overall localization root mean squared error (RMSE) of 0.447 m and a 90th-percentile error of 0.797 m, achieving a detection F1-score of 0.8 under the nominal noise configuration. Temporal analysis indicates that expanding the observation window from one to five frames yields the most significant performance gains, while two-source separation tests establish an empirical spatial resolution limit of 0.87 m. These results demonstrate that a data-driven yet physics-grounded dynamic EM-map representation can support distortion-aware localization and subsequent navigation in electromagnetically cluttered environments.