Water quality monitoring and management: integration of machine learning algorithms and Sentinel-2 images for the estimation of Chlorophyll-a

Abstract Accurate monitoring of chlorophyll-a (Chl-a) is critical for assessing eutrophication and water quality in inland aquatic ecosystems. While integrating Sentinel-2 MultiSpectral Instrument (MSI) imagery with machine learning (ML) is widely used for water quality assessment, challenges remain in optimizing Chl-a retrieval in optically complex small-scale reservoirs, particularly in semi-arid environments where traditional models often fail. This study introduces a machine learning (ML)-based approach to enhance Sentinel-2 MSI Chl-a retrieval, systematically evaluating six ML models: Artificial Neural Networks (ANN), CatBoost (CB), random forest (RF), ridge regression (RR), support vector regression (SVR), and extreme gradient boosting (XGB). By integrating Sentinel-2 MSI reflectance data with in-situ measurements from reservoirs exhibiting diverse trophic conditions, this study not only compares model performance but also applies feature importance analysis to refine spectral band selection for improved Chl-a retrieval. Among the models, RF demonstrated the highest predictive accuracy (R² = 0.93 calibration, R² = 0.91 validation), outperforming CB and XGB, whereas linear models such as RR and SVR were inadequate for capturing the nonlinear spectral relationships of Chl-a. The results highlight the advantages of ensemble-based ML models in handling complex water quality datasets, particularly in semi-arid and data-scarce regions. The proposed framework provides a robust, data-driven approach for integrating Sentinel-2 MSI imagery with AI techniques, supporting real-time water quality assessment and environmental management in diverse aquatic ecosystems. This research advances the application of ML for remote sensing-based water quality monitoring by systematically assessing model reliability, feature selection, and efficient application in semi-arid small-scale reservoirs. The proposed approach enhances the scalability of ML-driven remote sensing, offering a transferable framework for Chl-a estimation in data-scarce and environmentally vulnerable regions.