Improving Boreal Summer Precipitation Predictions From the Global NMME Through Res34‐Unet
Xuan Tong, Wen Zhou, Jiangjiang Xia
Abstract
Abstract Global warming and climate change have increased the frequency and intensity of floods and droughts, limiting economic development and threatening human survival. Therefore, accurate global forecasts well in advance of precipitation are essential to facilitate timely adaptation. Current seasonal forecasts are based mainly on numerical models, but raw forecasts suffer from systematic bias and under/overdispersion problems and cannot be directly used in applications. In addition, bias correction methods for global forecasts need to be further developed. Based on a fusion of ResNet34 and Unet, called Res34‐Unet, deep learning post‐processing is proposed to correct global precipitation forecasts of the North American Multi‐Model Ensemble (NMME). Compared with raw global NMME predictions, post‐processed precipitation predictions can be improved by up to 45%, which is significant at different latitudes. Feature importance analysis shows that precipitation itself, meridional wind, and sea surface temperature are key factors.