Data Reconstruction for Remotely Sensed Chlorophyll-a Concentration in the Ross Sea Using Ensemble-Based Machine Learning

Abstract

Polar regions are too harsh to be continuously observed using ocean color (OC) sensors because of various limitations due to low solar elevations, ice effects, peculiar phytoplankton photosynthetic parameters, optical complexity of seawater and persistence of clouds and fog. Therefore, the OC data undergo a quality-control process, eventually accompanied by considerable data loss. We attempted to reconstruct these missing values for chlorophyll-a concentration (CHL) data using a machine-learning technique based on multiple datasets (satellite and reanalysis datasets) in the Ross Sea, Antarctica. This technique—based on an ensemble tree called random forest (RF)—was used for the reconstruction. The performance of the RF model was robust, and the reconstructed CHL data were consistent with satellite measurements. The reconstructed CHL data allowed a high intrinsic resolution of OC to be used without specific techniques (e.g., spatial average). Therefore, we believe that it is possible to study multiple characteristics of phytoplankton dynamics more quantitatively, such as bloom initiation/termination timings and peaks, as well as the variability in time scales of phytoplankton growth. In addition, because the reconstructed CHL showed relatively higher accuracy than satellite observations compared with the in situ data, our product may enable more accurate planktonic research.