Global Framework for Chlorophyll-a Monitoring in Inland Lakes: Integrating Remote Sensing, Machine Learning, and Databases - Achievements and Challenges

Eutrophication is a significant environmental concern, which is often monitored through Chlorophyll-a (Chla) concentrations in inland and coastal waters. While traditional in-situ measurement methods are accurate, these are time-intensive, labor-demanding, and limited in spatial and temporal resolution. In recent years, remote sensing and machine learning approaches have emerged as promising alternatives for environmental monitoring, although their effectiveness is limited by challenges such as constrained in-situ data availability, the variability of water characteristics, and difficulties in transferring models across regions. Existing global models prioritize data quantity over quality, often lacking in comprehensive analysis of relationships between water quality parameters and remote sensing bands and indices. This study aimed to enhance global Chla prediction accuracy by improving data quality and identifying key predictive features using Earth Observation (EO) data. Two feature groups were examined: Group 1 (reflectance values from single bands and band ratio indices) and Group 2 (reflectance values from single bands combined with mathematical transformations of multiple bands). Machine learning models, including Random Forest (RF), Least Squares Boosting (LSBoost), Support Vector Regression (SVR), and Gaussian Process Regression (GPR), were assessed for overall performance, cross-validation accuracy, and transferability to external datasets. Among tested models with their own dataset, GPR achieved the highest overall accuracy (R² = 0.95, RMSE = 2.82 µg/L with Group 2 features), while SVR exhibited the weakest performance. For transfer validation using data from external lakes, RF (R² = 0.73, RMSE = 12.39 µg/L) and LSBoost demonstrated the greatest transferability. Spatial-temporal predictions of Chla over 2023–2024 successfully captured seasonal trends by revealing reliable and consistent patterns of Chla distribution. The present study highlights the potential of the proposed framework for global Chla monitoring in inland waters, also, emphasizing the potential in areas outside the training dataset.

Keywords: global chla monitoring, transferability, remote sensing, machine learning