Abstract
Particulate organic carbon (POC) is an essential component of the carbon pump within marine organisms. Exploring estimation methods for POC holds substantial significance for understanding the marine carbon cycle. In this study, we investigated the spatial heterogeneity of 30 factors and POC concentrations using geodetector to account for nonlinearity, diversity, and complexity. Ultimately, 20 factors including sea surface temperature, sea surface salinity, and chlorophyll-a were selected as modeling variables. Six machine learning models—backpropagation neural network, convolutional neural network, attention-based neural network, random forest (RF), adaptive boosting, and extreme gradient boosting were used to compare their performance. The results indicate that among the six machine learning algorithms, RF exhibits the strongest performance, with a root mean square error of 0.11 [log(mg/m 3 )] and an average percentage deviation of 2.73%. Global annual average sea surface POC concentrations were estimated for 2007 and compared to NASA’s POC product. The outcomes indicate that the RF model-based estimation method displays enhanced accuracy in estimating POC concentrations within intricate coastal environments, while the backpropagation neural network performed better in estimating POC concentrations in open ocean areas. Leveraging the RF model, global sea surface POC concentrations were estimated for the years 2007 through 2016, enabling a spatiotemporal analysis. The analysis unveils heightened POC concentrations in coastal regions and lower levels in open ocean areas. Furthermore, POC concentrations were greater in high-latitude regions compared to mid and low latitude counterparts. In conclusion, the global sea surface POC product in this study exhibits heightened spatial resolution and improved data completeness in contrast to other products. It enhances the accuracy of conventional POC estimation methods, particularly within coastal regions.