Implementation of artificial intelligence for analysis of long-term climate variability

Abstract

A simulation workflow is formulated to facilitate an application of modern data-driven approaches to climate problems by considering both theoretical aspects and modern tools of program implementation. A real supervised classification problem is considered as a use case. A convolutional neural network based model is developed to make the supervised classification of large-scale atmospheric circulation regimes. It has been found that a key prerequisite to ensure the success of the thus developed deep learning model is the proper matching of the data preprocessing and model architecture with the nature of the problem being considered. Particularly, applying of the seasonal anomalies transformation, tuning of the convolutional filters dimension and batch size adjustment have been found to be crucial to ensure a satisfactory accuracy of the regime recognition and a stable operation of the model. The generation of feature maps is demonstrated to be useful both as an interpretability tool and a heuristic approach to tuning of hyperparameters.