ABSTRACT Seasonal prediction of summer extreme droughts in the mid‐lower reaches of Yangtze River basin (MLYR) is of significant social‐economic value and has long been a big challenge in the meteorological community. Based on two machine learning models, the random forest (RF) and the Support Vector Machine (SVM), this study trained and validated a series of predictive models using observations from 1951 to 2021. The results show that the RF model built with the optimal climate index combination delivers the best seasonal prediction of summer extreme drought events in the MLYR, achieving a value of the area under the receiver operating characteristic curve of 0.85 and outperforming the SVM model based on its optimal climate index combination. A further interpretation of the decision trees in the optimal RF models reveals three significant decision paths and corresponding physical mechanisms leading to extreme droughts in the MLYR. The first one was caused by weakened or interrupted East Asian summer monsoon, which was further driven by persistently cold sea surface temperature (SST) in the Indian Ocean and western Pacific from the previous winter to the following spring. The second one was resulted from a joint influence of mid‐high latitude circulation anomalies and a westward extension of the western Pacific subtropical high with the driving forces from the North Atlantic Tripole mode. For the third one, the anomalous mid‐high latitude circulation possibly induced the extreme drought. However, no significant anomalies were found in the global SST and the driving forcing was suggested as the climatic noise. These findings deepen our understanding of nonlinear interactions among different preceding signals causing extreme droughts in MLRY and the optimal models have the potential for operational application, provided up‐to‐date climate indices are available.
Wei et al. (Thu,) studied this question.