Quantile Remap Calibration for Precipitation Nowcasting

Abstract

Precipitation nowcasting is critical for severe weather warnings, yet deterministic deep learning models often underestimate precipitation intensity, leading to poor performance on threshold-based metrics like the Critical Success Index (CSI). We propose Quantile Remap Calibration (QRC), a training-free post-hoc method that maps predicted intensity quantiles to observed quantiles, correcting systematic marginal miscalibration without modifying spatial structure. Our near-miss analysis reveals that 38.6% of false negatives at heavy rain thresholds are intensity near-misses, validating the hypothesis that intensity underestimation drives CSI gaps. On the SEVIR benchmark, QRC improves CSI-M-POOL16 from 0.4660 to 0.5249 (+12.6%) and CSI-219-POOL16 from 0.2083 to 0.2692 (+29.2%), closing 104% of the gap to CasCast using only post-hoc calibration without model retraining. QRC provides practitioners with a simple, effective baseline for improving extreme-event detection before investing in expensive generative post-processing.