The partial pressure of carbon dioxide in water (pCO₂) is an important indicator of CO₂ accumulation and potential air–water exchange in lakes. However, remote-sensing retrieval of pCO₂ remains challenging in large, shallow eutrophic lakes due to the interacting biological, physicochemical, and optical processes. This study combined in situ pCO₂ measurements, Landsat imagery, and aquatic environmental variables to develop a heterogeneous stacking ensemble comprising random forest (RF), gradient boosting decision tree (GBDT), and extreme gradient boosting (XGBoost) models for Lake Chaohu. Model comparison showed that the stacking model outperformed the individual models, with an R² of 0.80, RMSE of 104.22 μatm, RPD of 2.24 and SMAPE of 17.50%. During the four investigated years, lake-wide pCO₂ increased from 2013 to 2017 and was lower in 2021 and 2025; the maximum value in 2017 exceeded that in 2025 by 209.74 μatm. Autumn generally exhibited higher pCO₂ than winter. Persistent hotspots occurred in the western basin, nearshore waters in the northwest, and the central connecting zone, while lower values characterized the eastern basin and open waters in the southern part. SHAP analysis identified chlorophyll-a (Chl-a) as the leading predictor. However, its moderate Pearson correlation with pCO₂ (r = 0.34) indicates that its predictive importance may reflect nonlinear multivariate associations rather than a direct causal relationship. Water temperature (WT), dissolved oxygen (DO), pH, salinity (Sal), total dissolved solids (TDS), and electrical conductivity (EC) also contributed to the variability of pCO₂. These findings provide a remote sensing approach to monitor lake pCO₂ and characterize carbon dynamics in shallow eutrophic lakes.