Abstract: As an important component of inland waters, shallow lakes are hotspots for CO₂ emissions. Due to the influence of eutrophication and aquatic macrophyte, CO₂ fluxes at the water-air interface of shallow lakes exhibit complex variability, posing challenges for high-accuracy simulation. To compare the performance of different machine learning models in simulating CO₂ fluxes in shallow lakes, we focused on a floating-leaved vegetation zone in eastern Lake Taihu. Based on CO₂ flux observations from an eddy covariance system, combined with meteorological, water quality, and vegetation variables, we developed four machine learning models, random forest (RF), support vector machine (SVM), backpropagation neural network (BPNN), and long short-term memory network (LSTM). Then, we evaluated the performance under three modeling scenarios, including growing season, non-growing season, and whole-season. Among the three modeling scenarios, the whole-season modeling approach achieved the best overall performance, with test-set metrics consistently outperforming those of the seasonal models. The RF model exhibited the highest accuracy and robustness under all the three scenarios. In the whole-season mode-ling scenario, the RF model achieved a coefficient of determination (R²) of 0.72 and a root mean square error (RMSE) of 0.57 μmol·m^-2·s^-1. For the growing-season model, the RF performance yielded an R² of 0.64 and an RMSE of 0.88 μmol·m^-2·s^-1, while in the non-growing-season model, the R² and RMSE were 0.61 and 0.43 μmol·m^-2·s^-1, respectively. The SVM and BPNN models showed comparable but inferior performance, whereas the LSTM model performed relatively poorly. Furthermore, we used recursive feature elimination (RFE) to identify the optimal combination of driving factors for the RF model under the whole-season scenario. The selected feature set included: surface water temperature (T_{w_20}), sediment temperature (T_s), dissolved oxygen (DO), air tempera-ture (T_a), incoming shortwave radiation (R_{s_in}), wind speed (WS), total nitrogen (TN), water pH, friction velocity (u^*), and normalized difference vegetation index (NDVI). This feature set further improved simulation accuracy (R²=0.76, RMSE=0.55 μmol·m^-2·s^-1) and effectively reduced model complexity. The SHAP analysis showed the significant influences of water temperature, radiation, dissolved oxygen, and vegetation index on CO₂ fluxes. The results would provide a useful methodological reference for CO₂ flux modeling and carbon cycle studies in shallow lakes.

Key words: Lake Taihu, floating-leaf vegetation, CO₂ flux, machine learning, random forest (RF)