Deep Learning for Seasonal Prediction of Summer Precipitation Levels in Eastern China

Skilled seasonal forecasting will effectively reduce the economic losses caused by droughts and floods. Because of the powerful data mining capability of deep learning networks, it is increasingly applied in studies of seasonal rainfall prediction. However, there remain two prominent issues in the modeling process: the lack of enough training samples and the effect of a small number of extreme values on the model optimization. To tackle these deficiencies, we combine strategies such as principal component analysis, reduction of model hidden layers, and early-stopping with Attention U-Net to construct a rainfall classification forecasting model. These steps reduced the model outfitting and improved the model generalization. The results show that the prediction accuracy of this network with leads of 1-3 months is obviously better than that of the numerical model. Further analysis also supports that the spatial features of precipitation predicted by the network are very close to the observations. Accurate summer precipitation forecasts can effectively reduce economic losses and casualties caused by heavy precipitation and flooding in China. Nowadays, seasonal forecasting still mainly relies on numerical models based on physical knowledge. However, the predictors of rainfall are so broad and complex that some physical processes are not yet fully understood. As a result, the improvement of the model skills encountered a bottleneck. Recently, deep learning (DL) network can achieve complex function approximations of data by automatically extracting and learning data features. Yet the existing DL seasonal prediction models suffer from over-fitting and interference from noise and extreme values. Here, we use Attention U-Net to predict rainfall levels and combine strategies such as principal component analysis, reduced model depth, and early-stopping. This network not only solves the two drawbacks mentioned above. It also shows the superiority in the prediction results. First, the rainfall level prediction of this network is more accurate than the numerical model at 1-3 months ahead. Second, the network captures the spatial features with excellent performance. The predicted distributions are largely consistent with the true values. The Attention U-Net model outperforms the numerical model in predicting summer precipitation levels with a forecast lead of 1 monthConverting a regression task into a classification task can eliminate the bad effects of noises and extremesThe principal component analysis and the early-stopping strategy facilitate the model's generalization