Unpacking the black box of deep learning for identifying El Niño-Southern oscillation

Abstract

By training a convolutional neural network (CNN) model, we successfully recognize different phases of El Niño-Southern Oscillation (ENSO). Our model achieves high recognition performance, with accuracy rates of 89.4\% for the training dataset and 86.4\% for the validation dataset. Through statistical analysis of the weight parameter distribution and activation output in the CNN, we find that most of the convolution kernels and hidden layer neurons remain inactive, while only two convolution kernels and two hidden layer neurons play active roles. By examining the connections of weight parameters between these active convolution kernels and hidden neurons, we can automatically differentiate various types of El Niño and La Niña, thereby identifying the specific functions of each part of the convolutional neural network. We anticipate this work progress is helpful for future studies on both climate prediction and a deep understanding of the artificial neural networks.