Precipitation nowcasting of satellite data using physically-aligned neural networks
Abstract
Accurate short-term precipitation forecasts predominantly rely on dense weather- radar networks, limiting operational value in places most exposed to climate ex- tremes. We present TUPANN, a satellite-only model that decomposes the fore- cast into physically meaningful components: a variational encoder-decoder infers motion and intensity fields from recent imagery under optical-flow supervision, a lead-time-conditioned MaxViT evolves the latent state, and a differentiable advec- tion operator reconstructs future frames. We evaluate TUPANN on both GOES-16 and IMERG data, in up to four distinct climates at 10–180-min lead times using the CSI metric. Comparisons against optical-flow, deep learning and hybrid base- lines show that TUPANN achieves the best or second-best skill in most settings, with pronounced gains at higher thresholds. The model produces smooth, inter- pretable motion fields aligned with numerical optical flow and runs in near real time. These results indicate that physically aligned learning can provide nowcasts that are skillful, transferable and global.