A two-stage machine learning framework for Chl-a reconstruction and short-term prediction in eutrophic lakes using GOCI observations

Harmful algal blooms (HABs) in eutrophic lakes threaten aquatic ecosystems and public health, but short-term chlorophyll-a (Chl-a) prediction remains challenging because satellite observations are often discontinuous and bloom dynamics are regulated by interacting environmental and ecological factors. This study developed a two-stage Extreme Gradient Boosting (XGBoost) framework for satellite-derived Chl-a reconstruction and short-term prediction in Lake Taihu. The first-stage model (XGBoost_Interp) reconstructed Chl-a on dates without valid GOCI/GOCI-II retrievals and generated a continuous daily Chl-a series for 2011–2024, achieving good performance (R 2 = 0.83, RMSE = 5.72 μg/L, MAPE = 9.16%). Using the reconstructed series and environmental variables, the second-stage model (XGBoost_Pred) predicted subsequent Chl-a with comparable accuracy (R 2 = 0.83, RMSE = 5.62 μg/L, MAPE = 8.97%). However, independent validation from January to March 2025 showed lower accuracy (R 2 = 0.69, RMSE = 9.45 μg/L) and some overestimation, suggesting that further calibration and evaluation are still needed when valid satellite observations are unavailable for extended periods. SHapley Additive exPlanations (SHAP) identified wind speed, temperature, atmospheric pressure, day of year, and antecedent Chl-a as important predictors. After lake-specific recalibration, the framework also performed well in Lake Chaohu, indicating its practical potential for other large eutrophic lakes. Nevertheless, broader application requires further validation with nutrient-related variables, multi-sensor data, high-frequency field observations, and more lakes with different environmental conditions.