Interpretable convolutional neural network for analyzing precipitation in the pre-rainy season of South China

Abstract

Abstract The research and application of convolutional neural networks (CNNs) on statistical downscaling have been hampered by the fact that deep learning is highly dependent on sample size and is considered to be a black-box model. Therefore, a CNN model with transfer learning (CNN-TL) is proposed to study the pre-rainy season precipitation of South China. First, an augmented monthly dataset is created by sliding a fixed-length window over the daily circulation field and precipitation data for the entire year. Next, a base CNN network is pretrained on the augmented dataset, and then the network parameters are tuned on the actual monthly dataset from South China. Finally, guided backpropagation is conducted to obtain the distribution regions of the key features and explain the net. The coefficient of determination (R2) and root mean square error (RMSE) show that the CNN-TL model has higher explanatory power and better fitting performance than the feature extraction-based random forest (RF). Compared with the base CNN, the transfer learning approach can improve the explanatory power of the model by 10.29% and reduce the average RMSE by 6.82%. In addition, the interpretation results of the model show that the critical regions are primarily South China and its surrounding areas, including the Indochinese Peninsula, the Bay of Bengal, and the South China Sea. Furthermore, the ablation experiments and composite analysis illustrate that these regions are very important.