Seasonal prediction of summer extreme precipitation frequencies over Southwest China based on machine learning

Abstract

It is challenging to predict extreme precipitation over regions with complex terrain, such as Southwest China (SWC). This study develops a statistical seasonal prediction scheme for summer extreme precipitation frequencies (EPFs) over SWC. In this scheme, the year-to-year increment method is first applied to the SWC summer EPFs and boundary forcings in preceding seasons. The first six leading modes of SWC summer EPFs are extracted by empirical orthogonal function (EOF) decomposition. The relationships between the six modes and preceding boundary forcings are investigated. According to these relationships, physics-based predictors for the principal components (PCs) of the modes are selected. The first two modes are taken as examples to analyze the mechanisms by which the predictors influence SWC summer EPFs. Two machine learning (ML) models, random forest and extreme gradient boosting, are trained using the data during 1980–2011. The ML models that perform best for each PC on the training data are used to be independently tested on data during 2012–2021. The predicted EPF patterns can be reconstructed by predicted PCs and observed EOFs. Prediction results indicate that the optimal ML models can well capture the variations in the six modes and spatial patterns of SWC EPFs and outperform the traditional stepwise regression models. The prediction skills for SWC summer EPF anomaly patterns show year-to-year variations which depend on the prediction results of the PCs. This scheme may help the seasonal predictions of SWC summer EPFs.