Comparative Study on Gap-Filling of GOCI-I Chlorophyll-a Product using Kriging and Random Forest

Abstract

The enormous missing pixels of ocean color imagery due to clouds and other reasons have made it difficult to monitor the spatio-temporal ecological distribution of the sea. Thus the interpolation method has been devised and enhanced for a few recent decades, and the machine learning method was applied as trials in recent years. In this study, we use the GOCI-I chlorophyll-a product(Level2), having 500 m and hourly resolution for the East Asian waters (111.32° to 148.67 in longitude, 21.54 to 48.22° in latitude) centered on the Korean Peninsula. Kriging, a traditional spatial interpolation technique, and Random Forest, an ensemble machine learning method, are adopted to fill the gaps in the GOCI-I chlorophyll-a product. The input variables to the random forest are wind vector, current vector, sea surface temperature(SST), and sea level anomaly(SLA). A mask with about a 30% missing ratio(MR) on the ocean area is prepared in advance. The hourly products have high missing rates of over 70%. Thus 5-day mean products(DM5) are generated to reduce the missing rate. New products with MR over 30% are generated and designated as test data after applying the mask in advance to DM5. Product without applying the mask is designated as validation data. Performances of Kriging and RF are evaluated through RMSE and MAD.