Reconstruction of Sea Surface Salinity Fields Over the Bay of Bengal Using Machine Learning Approach

This study introduces a novel method for reconstructing historical sea surface salinity (SSS) fields over the Bay of Bengal (BoB) prior to satellite data. This approach utilizes the light gradient boosting machine (LightGBM) learning technique to reconstruct monthly SSS fields at a 1° spatial resolution using exclusively in situ observations. The method utilizes monthly averaged satellite-based SSS fields and in situ observations from 2011 to 2021. The accuracy of the reconstructed SSS is assessed with independent buoy observations, satellite measurements, model output, and climatology for the years 2010 and 2022. Reconstructed SSS successfully captured the complex spatial and temporal variability of the BoB basin and consistently outperforms both climatology and the model. Root mean square error (RMSE) of 0.72 psu (0.74 psu), 0.85 psu (0.83 psu), and 0.75 psu (0.76 psu) is obtained when reconstructed SSS, climatology SSS, and numerical model SSS are compared with satellite measurement during 2010 (2022). Reconstructed SSS also displays lower RMSE (0.48 psu) than climatology (0.50 psu) during the validation with independent buoy observations. Results promisingly suggest the robustness of the developed method and reliability of reconstructed SSS over climatology and numerical model output over BoB.