Using machine learning to analyze the changes in extreme precipitation in southern China

Abstract

Convolutional neural network (CNN) is applied to identify the extreme precipitation events (EPEs) in southern China and the physical contributions are quantified for the changes in extreme precipitation in recent decades. The CNN correctly identifies about 96% of the observed EPEs based on the given large-scale atmospheric circulation. The discrimination made by the neural network is revealed by using the layer-wise relevance propagation method. The CNN is inclined to classify the circulation pattern as an extreme precipitation circulation pattern (EPCP), when a deep cyclonic anomaly and intense horizontal pressure gradient appear over southern China at the lower and middle troposphere. The extreme precipitation amount decreases during April and May (AM) and increases in June to August (JJA) after the early-1990s. Result of the quantitative partitioning analysis indicates that the dynamic and thermodynamic effects respectively contribute 297.6% and −234% to the reduction of extreme precipitation in AM. The decline in EPCP frequency dominates the decrease in EPEs. In JJA, the increased extreme precipitation after the early 1990s is attributed to both the interdecadal increase in the EPCP frequency and the increasing trend in the atmospheric moisture. The dynamic and thermodynamic changes play almost equal roles in the increased extreme precipitation in JJA.