Optimized Ensemble Machine Learning Models for Predicting Phytoplankton Absorption Coefficients

The machine learning (ML) model provides an alternative method for estimating inherent optical properties (IOPs) in clear and coastal waters. This study introduces an effective approach by employing ensemble machine learning techniques, such as random forest, gradient boosting, extra tree, adaboost, bagging, and voting model, to predict phytoplankton absorption coefficient ( $\text{a}_{\mathrm {ph}}(\lambda)$ , $\text{m}^{-1}$ ) at selected key wavebands of 443, 489, 510, 555, and 670 nm in clear and coastal waters. The optimization of the hyperparameters of these models through Bayesian techniques ensured high predictive accuracy. Furthermore, this research highlights the critical importance of wavelengths 670, 489, and 510 nm through feature importance analysis. The models exhibit excellent performance in terms of the coefficient of determination (R2) value when predicting phytoplankton at various wavelengths (e.g., 443, 489, 510, 555, and 670 nm). The R2 value of around 0.9033 is obtained for the absorption coefficient of phytoplankton aph at the wavelength 510 nm. The lowest mean squared error (MSE) of 0.0001 was achieved at the green waveband (i.e., 555 nm). Other statistical matrices, such as mean absolute percentage error (MAPE) and mean absolute error (MAE), have shown a low error across the selected wavelengths. It is found that the predicted phytoplankton absorption coefficients are in close agreement with actual values. This study shows the success of optimized ensemble models for both global and selected regional datasets that can accurately derive $\text{a}_{\mathrm {ph}}(\lambda)$ , which will contribute to the improvement of ocean primary productivity modelling and understanding the distribution of phytoplankton blooms