Abstract: Aiming at the accuracy limitations of traditional statistical models and single machine learning algorithms in water quality remote sensing inversion, this study proposes an ensemble learning optimization method based on multi-source remote sensing imagery. The research employs multiple linear regression and three machine learning models (K-Nearest Neighbors, Support Vector Regression, and Multilayer Perceptron), along with three ensemble strategies (Bagging, Boosting, and Stacking), to invert water temperature, dissolved oxygen, and pH values in Poyang Lake. The results show that the Stacking strategy based on ridge regression performs best in water temperature inversion, and the R² value is 0.829, which is 8.2% higher than the k nearest neighbor which performs best in the single model; the Random Forest model within the Bagging ensemble demonstrates the best performance in pH remote sensing inversion, with an R² value of 0.868, representing a 12.9% improvement over the optimal single model (KNN); The Random Forest model also excels in dissolved oxygen inversion, achieving an equally high R² value of 0.868, which is 11.7% higher than the best single model (KNN). By optimizing the combination of spectral band features in the imagery, the study successfully validates the effectiveness of using limited spectral bands within the ensemble learning framework. The results demonstrate that ensemble learning methods significantly enhance the capability of satellite data in water quality inversion compared to single-model approaches, providing an efficient solution for lake environmental monitoring.