1636 / 2019-08-23 20:07:04

Improving Particulate Organic Carbon Estimations in the Surface Ocean from Satellite-derived Variables using Machine Learning Techniques

7、地理与地理信息科学 > 专题6：海洋、湖泊与河流先进监测技术

The World Ocean plays a vital role in the carbon cycle by regulating the amount of the carbon dioxide in the atmosphere, because oceanic POC particles act as the biological pump that moves carbon to the deep ocean and enables its long-term storage. Therefore, analyzing spatiotemporal variations of oceanic POC reservoirs and fluxes are important for understanding the role of the marine biosphere in the global carbon cycle. Ship-based in situ measurements are costly, labor-intensive and lacking spatial explicitness, while ocean color remote sensing has greatly improved our understanding in spatiotemporal POC variations over the global ocean. In literature, some simple empirical algorithms have been developed for POC estimations from remote sensing reflectance (Rrs), particulate backscattering coefficient or particulate attenuation coefficient. However, these algorithms may produce biased estimations for some scenarios, especially for productive coastal waters.
Therefore, this study tried to improve POC estimations from remote sensing data using the machine learning technique. In this study, in-situ POC concentration data collected around the World Ocean were obtained from PANGAEA and SeaBASS. As for the satellite ocean color data, the OC-CCI version 3.1 data, which is a merged product from SeaWiFS, MERIS and MODIS, were used. In addition, satellite-derived sea surface temperature and wind speed are also used as input for POC estimation. To extract matchups, in-situ POC measurements were matched to daily satellite data, and a 3×3 window was used to extract pixel values around each sampling station. As for the machine learning techniques, Recursive Feature Elimination (RFE) was used to selected sensitive features for POC estimation, and Gradient boosting decision tree (GBDT) was used to calibrate POC retrieval model.
A total of 3769 matchups were obtained, 2518 matchups were used to calibrate POC retrieval model, and the left were used as the test dataset. The performance of the calibrated GBDT model was compared with the POC model using the band ratio of 443 and 555 nm. Results showed that the GBDT model obtained high calibration and validation accuracy than the traditional band ratio model (Figure 1). Therefore, machine learning techniques should be promising in improving POC estimations in the surface ocean from satellite data.