Research on Stacking Ensemble Learning-Based Remote Sensing Retrieval of Total Phosphorus Concentration in Poyang Lake and Its Multidimensional Driving Mechanisms

As a key ecological barrier in the Yangtze River Basin, Poyang Lake’s water quality is pivotal to regional environmental stability. Remote sensing retrieval of total phosphorus (TP) supports water-quality monitoring and eutrophication assessment. Yet sparse monitoring and limitations of conventional methods hinder accuracy in complex inland waters. To address this, we develop a TP inversion framework that integrates a two-stage, parameterized perturbation. Which based on aquatic optical characteristics, enhancing multispectral reflectance (B02–B12) and TP distribution patterns via spectral feature optimization. It expands the dataset to 14x. Built on stacking ensemble learning, the framework employs LightGBM and Extra Trees as base learners and a Random Forest meta-learner to achieve nonlinear fusion of TP estimates. The framework attained RMSE 0.0148 mg/L and R² 0.6121, a 6.02% improvement over the best single model (Extra Trees: RMSE 0.0159 mg/L, R² 0.5519). Spatial analysis showed higher TP (0.15--0.213 mg/L) along the northern and southern margins, influenced by agricultural runoff and urban wastewater. And lower values (0.062--0.12 mg/L) in central waters. These results align with field measurements and historical pollution patterns. Pearson correlation and PCA indicate that TP heterogeneity stems from combined exogenous inputs and endogenous releases, regulated by evapotranspiration, precipitation, and temperature through hydrometeorological coupling. This “spectral enhancement–integrated inversion” framework provides reliable TP monitoring and technical support for water-quality management in this ecologically sensitive region.