[A-TT35] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT35-02] Construction of a estimation model for the partial pressure of carbon dioxide in seawater by machine learning using on-site water sampling data.

*Mitsuru Hayashi¹, Sou Toyonaga², Eiji Yamashita⁴, Masahiro Fujita³, Soichi Hirokawa³ (1.Research Center for Inland Seas, Kobe University, 2.Faculty of Ocean Science and Technology, Kobe University, 3.Graduate School of Maritime Sciences, Kobe University, 4.Okayama University of Science)

Keywords:carbon dioxide, Random Forest, Sensitivity Analysis, pCO2

Partial pressure of carbon dioxide in seawater (pCO₂), Water Temperature (T), Salinity, pH and Dissolved Oxygen (DO) were measured at Ushimado, Okayama prefecture, Japan from 1993 to 2010. A pCO₂ estimation model using T, S, pH and DO was constructed by using Random Forest, a machine learning method. A pCO₂ estimation model using T, S, pH, DO and the sampling date and time was constructed using a machine learning method, Random Forest. The accuracy of the estimate obtained by the holdout method was 3% relative error. Among the estimated parameters, pH and T were more important, while salinity and time were less important. Each parameter was varied one by one over the range of on-site data, and the sensitivity of the estimated model to the parameters was analyzed. An example of the response of pCO₂ estimates to variations in pH for each water temperature is shown in Figure 1. Estimates in the absence of training data were constant, as were estimates at the minimum and maximum values of the parameters. The response of pCO₂ to the parameters was qualitatively consistent with theory.