Satellite-Based Machine Learning Models for Chlorophyll-a and TSS Retrieval in Abu Dhabi's Coastal Waters

Abstract

Effective estimation of water quality parameters is essential for the sustainability of the coastal ecosystem. This research examines the potential use of Sentinel-2 Satellite images with machine learning models to measure the concentrations of Cholorophyll-a (Chl-a) and Total Suspended Solids (TSS) in the coastal area of Abu Dhabi. Google Earth Engine (GEE) was utilized to obtain Sentinel-2 Level-2A surface reflectance values, which are collocated with the in situ data. Field measurements were obtained from various locations, with 365 and 196 available samples for Chl-a and TSS, respectively. The former had 165 collocated points, whereas the latter had only 77 points. For feature engineering, two strategies were compared: spectral indices from the literature and Principal Component Analysis (PCA) with raw bands. Four machine learning algorithms were examined to find the optimal model for each parameter by using 5-fold cross-validated hyperparameter tuning. The selected models are Random Forest Regression (RFR), Support Vector Regression (SVR), Extreme Gradient Boosting (XGB), and Partial Least Squares (PLS) Regression. For Chl-a, the analysis showed that a general model was limited by localized bloom events near coastal outlets. Creating a specialized "Ambient-Conditions" model by excluding these outliers greatly improved performance. The optimal Chl-a model (XGB with PCA on six bands) achieved the highest accuracy with Test R2 = 0.7 and Test RMSE of 1.62 µg/L, representing an 80% improvement in precision compared to the general model trained on the full dataset (Test R² = 0.65, RMSE = 8.21 µg/L). PCA + Random Forest (on 10 bands) was the optimal model for TSS, with R2 = 0.61, despite the small dataset size. The results demonstrated that merging machine learning and remote sensing is effective for retrieving Chl-a and TSS in challenging marine waters.