Abstract:

Diffuse attenuation coefficient of downwelling irradiance K_d(z, λ) is an important parameter for estimating the profile distribution of underwater light filed and water constituents, and studying the photosynthesis of the phytoplankton and warning method of harmful algae bloom. K_d(z, λ) is a “quasi-inherent” optical property as a function of wavelength \lambda and depth z. Not only is it sensitive to absorption and scattering/backscattering coefficient, but also sensitive to the angular distribution of the normalized volume scattering function (i.e., scattering phase function). In this study, based on the volume scattering function [VSF, β(ψ, z)] in seven directions determined with a custom in situ device called VSAM (volume scattering and attenuation meter), the absorption coefficient a(z) determined with the WET Labs ac9 and ac-s, and the downwelling irradiance E_d(z) determined with the Satlantic Profiler Ⅱ OCI/R-200 and HyperPro Ⅱ in the north South China Sea (SCS) with a broad range, using LightGBM, Random Forest (RF) and CatBoost, three machine learning models for estimating the profile distribution of K_d(z, 650) were developed at first, and they were then evaluated by the key indicators including R²、RMSE、MAPE, as well as the comparison between in situ measured K_d(650) and estimated K_d(650). The evaluation indicated that the CatBoost model performed the best with R² of 0.8534, RMSE of 0.0472 m^-1, MAPE of 11.0585%, and the estimated K_d(650) was also closest to the measured K_d(650). Using the established CatBoost model, input inherent optical properties (IOPs) were the absorption coefficient, the volume scattering function (VSF), and their profile depth, the K_d(650) profile distribution among 15 m in the north SCS was estimated. The result shows that K_d(650) varies from 0.275 to 0.7 m^-1 at 5, 10 and 15 m underwater. At 5 m, K_d(650) is relatively stable while it varies greatly at 10 and 15 m. The contribution of volume scattering function distribution to K_d(z, λ) is considered in this study, which provides a new idea and method for accurate estimation and acquisition of K_d(z, λ) based on inherent optical properties (IOPs).

Key words: diffuse attenuation coefficient K_d(650), volume scattering function, absorption coefficient, machine learning