A Hybrid Machine Learning Model to Estimate Chlorophyll-a in Clear and Coastal Waters

The spatial and temporal variations of Chloropyll-a (Chl-a) in clear and coastal waters are critical for understanding the health of marine environment. To estimate Chl-a from oceanic and coastal waters, several empirical and semi-empirical models based on reflectance are available. However, Machine learning models have been used to simulate multiple environmental parameters over the past few decades, and their applications to water quality parameters have recently gained traction. This study proposes a novel approach to modeling Chl-a by using Fuzzy C-Means clustering based Neural Network (NN) which is optimized through Bayesian Optimization (BO) based on the band configuration of the Moderate Resolution Spectroradiometer Aqua (MODISA) with a wide range of variations. The training data were initially grouped into three clusters based on remote sensing reflectance values, and separate NN models were created for each cluster. Subsequently the hyperparameters of the NN models were optimized. The dataset includes (i) global in-situ measurements of NASA bio-Optical Marine Algorithm Dataset, (ii) SeaWiFS satellite matchups, and (iii) simulated dataset for the Red Sea. It exhibited significant variations in Chl-a levels under both oligotrophic and coastal conditions. Accuracy assessment of the present study is performed by comparing the modeled and observed values of the Chl-a. The performance matrices computed of the developed model were promising. Therefore, this study provides a potential approach for the retrieval of Chl-a in clear and coastal waters where the performance of existing algorithms is deteriorated to estimate precise values of Chl-a. The findings of this research could advance our understanding of biogeochemical cycles and processes in marine and open ocean waters.