Optimizing Satellite Turbidity Retrieval with Advanced Deep Learning Approaches

Abstract

In this study, we developed and evaluated both empirical and AI-based models to estimate turbidity in Gwangyang Bay, located on the southern coast of the Korean Peninsula. A dataset was constructed using Sentinel-2 satellite imagery and in situ turbidity measurements obtained from the automatic marine water quality monitoring network. The empirical model was developed based on an exponential regression using selected spectral bands. The AI models were constructed by applying a tree-based boosting algorithm and a deep learning approach using artificial neural networks. Notably, under strong tidal current conditions, incorporating tidal-related variables as training features improved the performance of the deep learning model. These findings suggest that the complementary use of empirical and AI models can enhance the accuracy of turbidity prediction and contribute to more efficient marine environmental monitoring.