Improving sub-seasonal forecasting of East Asian monsoon precipitation with deep learning

Accurate subseasonal forecast of East Asian summer monsoon precipitation (EASM) is pivotal, impacting the livelihoods of billions. However, the proficiency of state-of-the-art subseasonal-to-seasonal (S2S) models in forecasting precipitation remains constrained. We developed a convolutional neural network regression model, harnessing the more reliably predicted atmospheric variables from dynamic models to enhance their forecast skills for precipitation. The outcomes of the CNN model are promising: a 12% increase in accuracy and a 10% reduction in RMSE for precipitation forecast at the lead time of one week. The predictive skill of dynamic models for atmospheric variables shows a significant correlation with the performance of the CNN model. Ablation experiments on various predictors reveal that xx is the most influential factor affecting the CNN model's performance.