A machine learning approach to estimate surface ocean pCO2 from satellite measurements.

Surface seawater partial pressure of CO 2 (p CO 2) is a critical parameter in the quantification of air-sea CO 2 flux, which further plays an important role in quantifying the global carbon budget and understanding ocean acidification. Yet, the remote estimation of p CO 2 in coastal waters (under influences of multiple processes) has been difficult due to complex relationships between environmental variables and surface p CO 2. To date there is no unified model to remotely estimate surface p CO 2 in oceanic regions that are dominated by different oceanic processes. In our study area, the Gulf of Mexico (GOM), this challenge is addressed through the evaluation of different approaches, including multi-linear regression (MLR), multi-nonlinear regression (MNR), principle component regression (PCR), decision tree, supporting vector machines (SVMs), multilayer perceptron neural network (MPNN), and random forest based regression ensemble (RFRE). After modeling, validation, and extensive tests using independent cruise datasets, the RFRE model proved to be the best approach. The RFRE model was trained using data comprised of extensive p CO 2 datasets (collected over 16 years by many groups) and MODIS (Moderate Resolution Imaging Spectroradiometer) estimated sea surface temperature (SST), sea surface salinity (SSS), surface chlorophyll concentration (Chl), and diffuse attenuation of downwelling irradiance (Kd). This RFRE-based p CO 2 model allows for the estimation of surface p CO 2 from satellites with a spatial resolution of ~1 km. It showed an overall performance of a root mean square difference (RMSD) of 9.1 μatm, with a coefficient of determination (R2) of 0.95, a mean bias (MB) of −0.03 μatm, a mean ratio (MR) of 1.00, an unbiased percentage difference (UPD) of 0.07%, and a mean ratio difference (MRD) of 0.12% for p CO 2 ranging between 145 and 550 μatm. The model, with its original parameterization, has been tested with independent datasets collected over the entire GOM, with satisfactory performance in each case (RMSD of ≤~10 μatm for open GOM waters and RMSD of ≤~25 μatm for coastal and river-dominated waters). The sensitivity of the RFRE-based p CO 2 model to uncertainties of each input environmental variable was also thoroughly examined. The results showed that all induced uncertainties were close to, or within, the uncertainty of the model itself with higher sensitivity to uncertainties in SST and SSS than to uncertainties in Chl and Kd. The extensive validation, evaluation, and sensitivity analysis indicate the robustness of the RFRE model in estimating surface p CO 2 for the range of 145–550 μatm in most GOM waters. The RFRE model approach was applied to the Gulf of Maine (a contrasting oceanic region to GOM), with local model training. The results showed significant improvement over other models suggesting that the RFRE may serve as a robust approach for other regions once sufficient field-measured p CO 2 data are available for model training. • Remote sensing of ocean surface p CO 2 in complex regions has large uncertainties. • A machine learning approach is developed for the Gulf of Mexico and Gulf Maine. • The approach shows significantly improved performance over other approaches. • Uncertainties in the estimated surface p CO 2 are within 10 μatm for a large range. [ABSTRACT FROM AUTHOR]