Modeling the sea-surface $p$CO$_2$ of the central Bay of Bengal region using machine learning algorithms

Abstract

The present study explores the capabilities of advanced machine learning algorithms in predicting the sea-surface $p$CO$_2$ in the open oceans of the Bay of Bengal (BoB). We collect the available observations (outside EEZ) from the cruise tracks and the mooring stations. Due to the paucity of data in the BoB, we attempt to predict $p$CO$_2$ based on the Sea Surface Temperature (SST) and the Sea Surface Salinity (SSS). Comparing the MLR, the ANN, and the XGBoost algorithm against a common dataset reveals that the XGBoost performs the best for predicting the sea-surface $p$CO$_2$ in the BoB. Using the satellite-derived SST and SSS, we predict the sea-surface $p$CO$_2$ using the XGBoost model and compare the same with the in-situ observations from RAMA buoy. The model performs satisfactorily, having a correlation of 0.75 and the RMSE of $\pm$ 12.23 $\mu$atm. Further using this model, we emulate the monthly variations in the sea-surface $p$CO$_2$ for the central BoB between 2010-2019. Using the satellite data, we show that the central BoB is warming at a rate of 0.0175 per year, whereas the SSS decreases with a rate of -0.0088 per year. The modeled $p$CO$_2$ shows a declination at a rate of -0.4852 $\mu$atm per year. We perform sensitivity experiments to find that the variations in SST and SSS contribute $\approx$ 41$\%$ and $\approx$ 37$\%$ to the declining trends of the $p$CO$_2$ for the last decade. Seasonal analysis shows that the pre-monsoon season has the highest rate of decrease of the sea-surface $p$CO$_2$.