Novel Insights in Deep Learning for Predicting Climate Phenomena

Breakthroughs in Artificial Intelligence (AI) have demonstrated its potential to uncover novel insights and make predictions of unprecedented accuracy. Nevertheless, thorough assessment of their performance is crucial to ensure reliability in its results. One of the state-of-the-art machine learning methods for predicting the El Niño-Southern Oscillation (ENSO) phenomenon two years in advance is the CNN method presented by Ham et al. [22]. In this study, a comprehensive evaluation of the final forecasts is conducted to gain deeper insight into the deep learning model. We assessed the method's performance using k-old cross validation method as well as different evaluation methods, which led us to some surprising results. K-old cross validation shows that CNN method is not sensitive to training period. To ensure a holistic assessment, we also evaluated the time complexity of the method. compared to other ML methods, CNNs strike a balance between computational efficiency and predictive power, particularly for tasks involving spatial data like climate predictions. In addition, the divergence between pre and post-2000 accuracy proves CNN's disproportionate skill over different periods. The correlation criterion targets the ability to detect El Niño or La Nina's trend without considering its magnitude and strength, whereas value-based criteria target the strength and trueness of the results. Thus, to reach a fair judgment about any method, these two kinds of criteria and a holistic assessment from different points of view have to be considered.