Identify Patterns of Flash Heavy Rainfall in Limited Area by FConvNeXt

The properties and distributions of precipitation are often determined by specific synoptic patterns. Hence, the objective identification of corresponding impact patterns is an important field of research for improving rain forecasting. However, the identification of the weather patterns producing intense rainfall is much more challenging. Since they are violent and local, impact patterns tend to be meso- or smaller-scale systems and are often incompletely presented or only presented in limited regions. In this paper, a deep learning network with a feature cross-fusion module, FConvNeXt, was proposed to address this difficulty and showed great potential. Four major patterns corresponding to intense rainfall in the Beijing–Tianjing–Hebei Region were studied. Statistical testing showed that FConvNeXt performed better than ConvNeXt and ResNet and that the model could identify the weak synoptic forcing type, the subtropical high-pressure type, and the low-vortex pattern with high accuracy. Furthermore, a strictly independent 2021 dataset was tested, and FConvNeXt maintained an equal if not even slightly better performance in spite of a decrease in the subtropical high-pressure type. Meanwhile, the study showed that the accuracy in identifying the upper-level trough type is the lowest for the three deep learning methods, which maybe because the northeast vortex was intercepted in the limited region, making  it difficult to distinguish from the shallow upper-level trough type. This study is useful for improving the fine objective of forecasting intense rainfall.

In summary, in contrast to previous objective classifications on large-scale weather systems in large regions, this study explored the objective classification of meso- and small-scale weather patterns that correspond to heavy rainfall and flash flooding within a limited region. Advanced deep learning models were employed that showed significant potential for this application. Furthermore, a new cross-fusion feature extraction module was proposed that improved the accuracy of the LVT classification within a limited region. Moreover, the study introduced a pre-training model to improve the training speed, which improved the accuracy and significantly shortened the training time.