A data-driven framework for assessing the impact of environmental factors on phytoplankton biomass in a eutrophic lake

Phytoplankton dynamics in eutrophic lakes demonstrate nonlinear environmental responses, complicating algal bloom forecasting. A novel hybrid LightGBM-GAM framework that combines the machine learning-driven feature selection of Light Gradient Boosting Machine (LightGBM) with ecological interpretability strengths of Generalized Additive Models (GAMs) is developed to predict phytoplankton biomass (PB) in Lake Taihu. Using decadal (2009–2018) monthly data from five lake stations, the framework outperformed traditional multivariate GAMs. It achieved adjusted R2 values of 0.36–0.56 and explaining 41.8–64.2% of PB variance, with lower AIC values and superior capture of extreme bloom events. LightGBM identified water quality parameters (chemical oxygen demand, total phosphorus, total nitrogen) and meteorological factors (air temperature, air pressure, solar radiation, wind speed) as dominant predictors, contributing over 85% cumulative feature importance. GAMs further elucidated their nonlinear and site-specific effects, revealing higher PB in northern coastal regions (e.g., Meiliang Bay) compared to central zones. This dual-phase methodology bridges machine learning efficiency with ecological mechanism elucidation, enabling precise identification of bloom triggers and spatially differentiated management thresholds. Our findings advance predictive frameworks for lake ecosystems while providing operational guidance for nutrient control in northern Taihu’s hypereutrophic zones.