Comparison of conventional and machine learning regression models for accurate prediction of selected optical active components - A case study: The Gulf of Izmit

Abstract

This study hypothesizes that advanced machine learning (ML) models can more accurately predict certain critical water quality parameters in marine environments compared to conventional regression techniques. We specifically evaluated the spatio-temporal distribution of Chlorophyll-a (Chl-a) and Secchi Disk Depth (SDD) in the Gulf of Izmit using in-situ measurements and Sentinel-2 satellite imagery from October 2021 and 2022. Among the models tested, the Support Vector Regression (SVR) model showed better predictive performance, achieving the lowest RMSE for SDD (1.11-1.70 m) and Chl-a (1.16-4.97 mg/m(3)) and the lowest MAE for SDD (0.86-1.43 m) and Chl-a (1.03-3.17 mg/m(3)). Additionally, the study observed a shift from hypertrophic to eutrophic Chl-a conditions and from mesotrophic-eutrophic to oligotrophic SDD conditions between 2021 and 2022, aligning with SVR model predictions and in-situ observations. These findings underscore the potential of ML models to enhance the accuracy of water quality monitoring and management in marine ecosystems.