Estimating spatial and temporal variation in ocean surface pCO in the Gulf of Mexico using remote sensing and machine learning techniques.

Research on the carbon cycle of coastal marine systems has been of wide concern recently. Accurate knowledge of the temporal and spatial distributions of sea-surface partial pressure (pCO) can reflect the seasonal and spatial heterogeneity of CO flux and is, therefore, essential for quantifying the ocean's role in carbon cycling. However, it is difficult to use one model to estimate pCO and determine its controlling variables for an entire region due to the prominent spatiotemporal heterogeneity of pCO in coastal areas. Cubist is a commonly-used model for zoning; thus, it can be applied to the estimation and regional analysis of pCO in the Gulf of Mexico (GOM). A cubist model integrated with satellite images was used here to estimate pCO in the GOM, a river-dominated coastal area, using satellite products, including chlorophyll-a concentration (Chl-a), sea-surface temperature (SST) and salinity (SSS), and the diffuse attenuation coefficient at 490 nm (Kd-490). The model was based on a semi-mechanistic model and integrated the high-accuracy advantages of machine learning methods. The overall performance showed a root mean square error (RMSE) of 8.42 μatm with a coefficient of determination (R) of 0.87. Based on the heterogeneity of environmental factors, the GOM area was divided into 6 sub-regions, consisting estuaries, near-shores, and open seas, reflecting a gradient distribution of pCO. Factor importance and correlation analyses showed that salinity, chlorophyll-a, and temperature are the main controlling environmental variables of pCO, corresponding to both biological and physical effects. Seasonal changes in the GOM region were also analyzed and explained by changes in the environmental variables. Therefore, considering both high accuracy and interpretability, the cubist-based model was an ideal method for pCO estimation and spatiotemporal heterogeneity analysis.