Applications of unsupervised machine learning in climate research

Abstract

The use of machine learning in climate science is expanding rapidly after its success in other fields. The more powerful data driven models of machine learning show promise to give us more accurate predictions of and potentially more insights into the physical climate system than conventional statistical models.

Many new applications for machine learning in climate science have been proposed in recent years, although it is uncertain which of them will prove to be fruitful lines of research.

In this thesis, we explore two well-established problems in statistical climate science and approach them using unsupervised machine learning techniques, namely climate mode extraction and the bias correction of simulations.

In the first part of this thesis, we develop a framework to test methods of mode extraction on climate-like data. We generate imitation global climate fields, which include as much of the expected complexity of real data as possible. We find that the newer mode extraction methods, which brand themselves as machine learning rather than conventional statistics, outperform the conventional approaches by more accurately extracting the known modes in the data. When applied to reanalysis and model surface temperature data, the newer methods extract a well-constrained ENSO signal and warming trend than the classical methods. However, no method can safeguard against generating false modes in all circumstances. We show the consequences of false mode extraction with examples and suggest how incorrectly extracted modes could lead to false proposed modal mechanisms.

In the second part, we explore the use of unsupervised deep neural networks for bias correcting large, simulated climate fields of multiple variables. We show that the structure of neural networks allows for more faithful bias correction of cross-variable and spatial correlations when used to map between simulations and reanalysis data. A particular advantage of these methods is that they can shift misplaced climate features. We also show the limitation of these networks and how the lack of constraints in optimising them can lead to unexpected bias correction results.

By exploring these uses of unsupervised machine learning in climate science we hope that can be developed further and be utilised more by researchers.