While research on pCO2 (partial pressure of carbon dioxide in seawater) in surface seawater in coastal areas has advanced, it remains less extensive than studies in the open ocean. One reason is the difficulty of measuring pCO2 directly. In this study, we develop models to estimate pCO2 from commonly measured parameters using the random forest. For estimating pCO2 from seawater temperature, salinity, pH, and dissolved oxygen, random forest is considered to be valid compared to multiple linear regression. In terms of pCO2 characteristics, building estimation models separately for Osaka Bay and Bisan Seto and other areas improved the accuracy compared to the single model, which estimation errors were 32.0, 34.9, 20.9 µatm, respectively. Inputting the Seto Inland Sea Comprehensive Water Quality Survey data to the three models, we estimated pCO2 and the spatial distribution. Compared to measured spatial variation, estimated pCO2 showed a similar trend and the maximum relative error was about 17%. From estimating spatial distribution, we obtained similar characteristics to the in situ measurements, such as the lowest in the inner part of Osaka Bay, and the highest in the Bisan Seto.
Fujita M., Hayashi M., Yamashita E. & Hirokawa S., in press. Estimating pCO2 using random forest regression in the Seto Inland Sea, Japan. Journal of Oceanography. Article.
