Two decades of flood-pulse-driven chlorophyll-a dynamics in an Amazon floodplain lake

Amazonian floodplain lakes rank among the most dynamic aquatic ecosystems globally, yet their monitoring is hampered by extreme hydrological variability, high turbidity, and optical complexity. We investigated chlorophyll-a (Chl-a) dynamics in Lago Grande de Curuai, a large floodplain lake hydraulically connected to the Amazon River, by integrating multi-campaign in situ measurements with two decades of MODIS surface reflectance (2001-2024) within a machine learning and Explainable Artificial Intelligence (XAI) framework. Among tested regressors, an optimized Support Vector Regression (SVR_Optuna) consistently outperformed Random Forest, XGBoost, LightGBM, Partial Least Squares, and linear models, achieving R² = 0.855 in log1p space and RMSE = 12.5 mg m^-3 after back-transformation in holdout validation. Applied to MODIS monthly composites, the model reconstructed 24 years of Chl-a variability, revealing recurrent seasonal maxima during receding and low-water phases of the Amazon flood pulse. Spatial analyses highlighted enhanced concentrations in semi-isolated basins and stronger dilution near riverine connections. Pixel-wise uncertainty metrics-derived from temporal variability, bootstrap resampling, and climatological anomalies-indicated robust retrievals in pelagic zones but higher uncertainty along river-lake interfaces. SHAP-based XAI confirmed the dominant role of red and red-edge spectral features (665-709 nm) in Chl-a prediction, consistent with bio-optical theory and strengthening model interpretability. Beyond Curuai, this scalable approach enables systematic monitoring of floodplain lakes across the Amazon Basin, where hydrological connectivity, extreme droughts, and intensifying human pressures increasingly affect water quality.