Estimating surface pCO2 in the northern Gulf of Mexico: Which remote sensing model to use?. (1st December 2017)

• Record Type: Journal Article
• Authors: Chen, Shuangling; Hu, Chuanmin; Cai, Wei-Jun; Yang, Bo
• Is Part Of: Continental shelf research. Volume 151(2017)
• Page Start: 94
• Page End: 110
• Publication Date: 2017-12-01
• Subjects: Surface pCO2 -- TA -- DIC -- Northern GOM -- MODIS -- Remote sensing; Continental shelf -- Periodicals; Submarine geology -- Periodicals; 551.41
• DOI: 10.1016/j.csr.2017.10.013
• Languages: English
• ISSNs: 0278-4343

Abstract

Abstract: Various approaches and models have been proposed to remotely estimate surface p CO2 in the ocean, with variable performance as they were designed for different environments. Among these, a recently developed mechanistic semi-analytical approach (MeSAA) has shown its advantage for its explicit inclusion of physical and biological forcing in the model, yet its general applicability is unknown. Here, with extensive in situ measurements of surface p CO2, the MeSAA, originally developed for the summertime East China Sea, was tested in the northern Gulf of Mexico (GOM) where river plumes dominate water's biogeochemical properties during summer. Specifically, the MeSAA-predicted surface p CO2 was estimated by combining the dominating effects of thermodynamics, river-ocean mixing and biological activities on surface p CO2 . Firstly, effects of thermodynamics and river-ocean mixing ( p CO2@Hmixing ) were estimated with a two-endmember mixing model, assuming conservative mixing. Secondly, p CO2 variations caused by biological activities (Δ p CO2@bio ) was determined through an empirical relationship between sea surface temperature (SST)-normalized p CO2 and MODIS (Moderate Resolution Imaging Spectroradiometer) 8-day composite chlorophyll concentration (CHL). The MeSAA-modeled p CO2 (sum of p CO2@Hmixing and Δ p CO2@bio ) was compared with the field-measured p CO2 . The Root Mean Square Error (RMSE) was 22.94 µatm (5.91%), with coefficient of determination (R 2 ) of 0.25, mean bias (MB) of − 0.23 µatm and mean ratio (MR) of 1.001, for p CO2 ranging between 316 and 452 µatm. To improve the model performance, a locally tuned MeSAA was developed through the use of a locally tuned Δ p CO2@bio term. A multi-variate empirical regression model was also developed using the same dataset. Both the locally tuned MeSAA and the regression models showed improved performance comparing to the original MeSAA, with R 2 of 0.78 and 0.84, RMSE of 12.36 µatm (3.14%) and 10.66 µatm (2.68%), MB of 0.00 µatm and − 0.10 µatm, MR of 1.001 and 1.000, respectively. A sensitivity analysis was conducted to study the uncertainties in the predicted p CO2 as a result of the uncertainties in the input variables of each model. Although the MeSAA was more sensitive to variations in SST and CHL than in sea surface salinity (SSS), and the locally tuned MeSAA and the empirical regression models were more sensitive to changes in SST and SSS than in CHL, generally for these three models the bias induced by the uncertainties in the empirically derived parameters (river endmember total alkalinity (TA) and dissolved inorganic carbon (DIC), biological coefficient of the MeSAA and locally tuned MeSAA models) and environmental variables (SST, SSS, CHL) was within or close to the uncertainty of each model. While all these three models showed that surface p CO2 was positively correlated to SST, the MeSAA showed negative correlation between surface p CO2 and SSS and CHL but the locally tuned MeSAA and the empirical regression showed the opposite. These results suggest that the locally tuned MeSAA worked better in the river-dominated northern GOM than the original MeSAA, with slightly worse statistics but more meaningful physical and biogeochemical interpretations than the empirical regression model. Because data from abnormal upwelling were not used to train the models, they are not applicable for waters with strong upwelling, yet the empirical regression approach showed ability to be further tuned to adapt to such cases. Highlights: Both mechanistic and empirical pCO2 models are developed for northern GOM in summer. Sensitivities of each model to the input variables (SST, SSS, CHL) are analyzed. Strength, weakness, and applicability of each model are compared and discussed.