[M-GI26] Data-driven approaches for weather and hydrological predictions

[MGI26-05] Forecasting monthly sea surface temperature using Koopman mode decomposition

*Daiya Shiojiri¹, Yoshihiko Susuki², Shunji Kotsuki¹ (1.Chiba University, 2.Kyoto University)

Keywords:Koopman Model Decomposition, Sea Surface Temperature, Data-driven prediction

In recent years, remarkable progresses have been made in technologies related to Earth observations, and large amounts of observation data are being accumulated every day. By applying machine learning methods to such data, various studies have been conducted to predict Earth systems from the data alone. In general, machine learning is an important research subject in the Earth science field because it enables state estimations with a much smaller computational cost than physical models. On the other hand, the time evolution process of many Earth systems is highly nonlinear, which often makes state estimations by machine learning difficult. Therefore, it is important to introduce machine learning methods that can be applied to systems with nonlinear time evolution. One of the machine learning methods suitable for nonlinear systems is deep neural networks (DNNs), but the lack of training data can often be a problem. Therefore, it is useful to explore relatively simple data-driven approaches that represent nonlinear time evolution without using a large amount of training data.
In this study, we investigate Koopman mode decomposition (KMD) as a method that can efficiently approximate Earth systems’ nonlinear time evolution by linear operators. The KMD is a method to represent a finite-dimensional nonlinear time-evolving system as a sum of infinite number of oscillating modes with a single frequency. However, since it is not realistic to reproduce an infinite number of modes on a computer, several methods have been proposed to approximate the system as a sum of a finite number of oscillating modes. In this study, we apply three KMD approaches for monthly sea surface temperature (SST) training data and validate whether we can predict SST during the test period using decomposed modes by KMD. As a result, we succeeded in predicting monthly SST accurately compared to KMD-based SST used in an existing study.