A machine learning approach for deriving spectral absorption coefficients of optically active oceanic constituents

Partitioning the total non-water absorption coefficient (anw(lambda), m-1) into subcomponent Inherent Optical Properties or absorption subcomponents (IOPs), i.e., absorption due to various optically active substances like phytoplankton, (aph(lambda),m-1), detrital material, (adm(lambda),m-1) and colored dissolved organic matter (CDOM), (ag(lambda), m-1) provide information about light interaction with various materials, understanding aquatic ecology and biogeochemical cycles. Some of the existing partitioning methods either assume shapes for the constituent's absorption or require additional ancillary variables and are computationally intensive. Hence, seventeen Machine Learning (ML) based models that neither assume shapes for constituent's absorption nor require ancillary inputs are trained and evaluated. The ML models are trained to provide one single output at a time that corresponds to a subcomponent IOP at a wavelength as output. To validate the performance of the trained ML models, anw(lambda) from two publicly available bio-optical datasets are used as input to provide the subcomponent IOPs as outputs. A quantitative statistical methodology used to rank various ML model's performance indicated that the two ensemble type models, Extremely Randomized Trees (ERT) and bagging method (EBG), obtained higher scores. In terms of computational time requirement, the EBG (time - 1.51 s) and ERT (11.5 s) models outperformed two existing models, Zhang's model (Zhang et al., 2015) - 32 s and Lin's model (Lin et al., 2013) - 1488 s with similar accuracy in deriving the subcomponent IOPs. The trained ML models can derive spectral subcomponent IOPs from satellite imagery and continuous profiling systems.

Automated summary

This study utilized machine learning models to partition the absorption coefficients of different optically active substances in water, providing insights into light interaction with various materials. In model performance evaluation, the Extremely Randomized Trees and bagging method models showed higher scores and shorter computational time compared to existing models.