Approach for estimating the vertical distribution of the diffuse attenuation coefficient in the South China Sea.

The vertical distribution of the diffuse attenuation coefficient K(z, λ) is critical for studies in bio-optics, ocean color remote sensing, underwater photovoltaic power, etc. It is a key apparent optical property (AOP) and is sensitive to the volume scattering function β(ψ, z, λ). Here, using three machine learning algorithms (MLAs) (categorical boosting (CatBoost), light gradient boosting machine (LightGBM), and random forest (RF)), we developed a new approach for estimating the vertical distribution of K(z, 650), K(z, 650), and K(z, 650) and applied it to the South China Sea (SCS). In this approach, based on in situ β(ψ, z, 650), the absorption coefficient a(z, 650), the profile depths z, and K(z, 650), K(z, 650), and K(z, 650) calculated by Hydrolight 6.0 (HL6.0), three machine learning models (MLMs) without or with boundary conditions for estimating K(z, 650), K(z, 650), and K(z, 650) were established, evaluated, compared, and applied. It was found that (1) CatBoost models have superior performance with R ≥ 0.92, RMSE≤ 0.021 m, and MAPE≤ 4.3% and most significantly agree with HL6.0 simulations; (2) there is a more satisfactory consistency between HL6.0 simulations and MLMs estimations while incorporating the boundary conditions; (3) the estimations of K(z, 650), K(z, 650), and K(z, 650) derived from CatBoost models with and without boundary conditions have a good agreement with R ≥0.992, RMSE ≤0.007 m, and MAPE≤0.8%, respectively; (4) there is an overall decreasing trend with increasing depth and increasing offshore distance of K(z, 650), K(z, 650), and K(z, 650) in the SCS. The MLMs for estimating K(z, λ) could provide more accurate information for the study of underwater light field distribution, water quality assessment and the validation of remote sensing data products.