Spatial downscaling of MODIS Chlorophyll-a with machine learning techniques over the west coast of the Yellow Sea in South Korea

Effective water quality monitoring of coastal areas through the measurement of Chlorophyll-a (Chl-a) has remarkably progressed by ocean color remote sensing. Among different sensors, Moderate Resolution Imaging Spectroradiometer (MODIS) Level 3 products provide reliable global representations of the Chl-a concentration. On the other hand, due to the coarse spatial resolution of MODIS data, its applicability is limited for spatially complex coastal regions. To overcome this limitation, a few downscaling techniques have been suggested based on the polynomial regression method. However, this type of regression has some restrictions, such as sensitivity to outliers, and nonlinear types of machine learning algorithms have not been tested in downscaling Chl-a datasets. Therefore, three machine learning (ML) techniques, support vector regression (SVR), random forest regression (RFR), and long short-term memory (LSTM), were developed using the Sentinel-2A/MSI bands as predictors and MODIS Chl-a as a predictand and compared their results with the results of multiple polynomial regression (MPR), to find the most suitable model for downscaling MODIS Chl-a in coastal area of South Korea. The obtained results showed that the 2nd degree MPR and SVR-Radial Basis Function (RBF) illustrate the best performance in the winter and summer days, respectively. In addition, LSTM is less sensitive to the changes in all variables (sensitivity index range from 0.31 to 0.48). Overall, we conclude that the downscaling approach based on ML models, especially SVR-RBF, can serve as a suitable alternative in some cases to produce high-resolution Chl-a maps, especially for coastal marine water quality monitoring.

Automated summary

This study explored the downsclaing of MODIS Chl-a data using machine learning models. Results showed that support vector regression and 2nd degree polynomial regression performed best in different seasons, while long short-term memory model exhibited the lowest sensitivity to changes in variables.