When Does Test-Time Adaptation Help Precipitation Nowcasting?

Test-time adaptation (TTA) addresses distribution shift by updating models on unlabeled target data during deployment. While widely studied for image classification, its efficacy for spatiotemporal regression tasks remains unclear. This paper presents a systematic evaluation of TTA methods applied to precipitation nowcasting. Using a SimVP model trained on the SEVIR radar archive, we evaluate adaptation under two conditions: real inter-year temporal drift (2017→2019) and multiplicative intensity shift simulating radar calibration degradation. The effectiveness of TTA depends strongly on the structure of the shift. Under real temporal drift, most methods provide no clear benefit, and gradient-free statistics recalibration (TBN) actively degrades performance (−1.3% average CSI). Under structured intensity shift, TTA consistently improves performance. Gradient-free recalibration alone recovers 82% of the available gain with zero gradient steps, and Sharpness-Aware Robust adaptation (SAR) yields the best overall result (+6.5% average CSI, +17.5% CSI-74 at severity 5). These results argue against treating TTA as a universal upgrade, and for explicitly matching the adaptation mechanism to the structure of the observed shift.