Stabilize adversarial training for GAN-based enhancements in satellite precipitation nowcasting

Develop training strategies that stabilize generative adversarial networks used to enhance satellite-derived precipitation nowcasts (specifically the GAN-TUPANN variant applied to GOES-16 RRQPE data within the TUPANN framework), ensuring robust adversarial learning and reliable performance across lead times without the instabilities observed in current experiments.

Background

The paper proposes TUPANN, a physically aligned, satellite-only precipitation nowcasting model and evaluates a GAN-augmented variant (GAN–TUPANN). While GAN–TUPANN produces visually sharper predictions, quantitative metrics (CSI and HSS) show mixed or degraded performance compared to the base model, indicating training instability and trade-offs between visual realism and forecast skill.

In the Limitations and future work section, the authors explicitly state that stabilizing adversarial training remains an open challenge, motivated by observed inconsistencies in GAN-based enhancements during ablation studies.