A machine learning algorithm to retrieve the red peak of phytoplankton absorption spectra from ocean-colour remote sensing

Abstract

Light absorption by microscopic phytoplankton in marine ecosystems is a crucial process underpinning biological production and global biogeochemical cycles. Accurate estimation of phytoplankton absorption coefficients, an inherent optical property of ocean water, can improve remote sensing applications including spectral photosynthesis models and assessments of ocean health, biodiversity, and climate change impacts. However, considerable uncertainty exists in current satellite retrievals of phytoplankton absorption coefficients, particularly for ɑ_ph(676) - the phytoplankton absorption peak at red wavelengths near 676 nm - which is an input to several novel and advanced satellite algorithms. This uncertainty hinders operational use of algorithms for assessing phytoplankton physiology, size structure and oceanic carbon pools from space. We aimed to improve satellite-based estimation of ɑ_ph (676) using advanced machine learning (ML) techniques. We compiled a comprehensive in situ dataset (n = 1576) of ɑ_ph(676) from published databases and matched with remote-sensing reflectance Rrs at six wavelengths (412, 443, 490, 510, 560, and 665 nm) from the Ocean Colour Climate Change Initiative. We extensively evaluated multiple base ML algorithms: Random Forest (RF), Gradient Boosting Machines, and Linear Regression; and implemented ensemble ML models: RF with Grid Search Cross-Validation, eXtreme Gradient Boosting Ensembled Model, Ensemble Forecast, Stacked Voting, Optimised Ensemble and Meta Stacking, integrating the base models through cross-validated hyperparameter tuning. Meta Stacking outperformed individual ML models in predictive accuracy across temporal resolutions, showing best results with daily composites. Our study addresses key limitations of previous models, including small training datasets, inconsistent performances, and lack of ensemble comparisons. We present a robust, extensively trained and validated ensemble ML model that significantly improves ɑ_ph(676) estimation and opens the possibility of routinely using in ocean colour remote sensing.