Monthly Prediction on Summer Extreme Precipitation With a Deep Learning Approach: Experiments Over the Mid-To-Lower Reaches of the Yangtze River

Abstract

Accurate predictions of monthly extremes assume paramount importance in enabling proactive decision-making, which however are lacked in skills even for state-of-the-art dynamical models. Taking the extreme precipitation prediction over the mid-to-lower reaches of the Yangtze River, China, as an instance, a multi-predictor U-Net deep learning approach is designed to enhance the prediction over the European Center for Medium-Range Weather Forecasts (ECMWF) model, with the single-predictor U-Net parallelly examined as the benchmark. Focusing on the precipitation extremes, an extreme associated component is incorporated into the model loss function for optimization. Besides, predictions composed by daily outputs with multiple lead times are imported as a comprehensive set in the training phase to augment the deep learning sample size and to emphasize enhancements in predictions at the monthly timescale as a whole. Results indicate that the multi-predictor U-Net effectively improves predictions of extreme summer precipitation frequency, showing distinct superiority to the raw ECMWF and the single-predictor U-Net. Multiple evaluation metrics indicate that the model shows a significant positive improvement ratio ranging from 65.1% to 80.0% across all grids compared to the raw ECMWF prediction, which has also been validated through applications in the two extreme summer precipitation cases in 2016 and 2020. Besides, a ranking analysis of feature importance reveals that factors such as humidity and temperature play even more crucial roles than precipitation itself in the multi-predictor extreme precipitation prediction model at the monthly timescale. That is, in such a deep learning approach, the monthly prediction on extreme precipitation benefits significantly from the inclusion of multiple associated predictors.