AI and the World's Hard Problems · Earth & climate science · arXiv · 2026-10-20
Wet-Biased Data Pipelines Mislead the Training and Evaluation of Generative Precipitation Nowcasting
Generative precipitation nowcasting models are usually trained and evaluated on wet-biased datasets, and test cases selected based on target observations can shift the distribution the model learns. The authors show analytically that this sampling bias can conceal training bias. Using a fixed satellite-to-radar diffusion model, they show that the calibration assessment can reverse depending on the test set: the model looks under-predictive on a wet-biased test set and over-predictive on a deployment-representative one.