Exploring the contrasting lake CO fluxes and influencing variables in four large shallow subtropical lakes with different hydrological connectivity.

Lake carbon dioxide (CO) evasion is a crucial component of global carbon cycle, yet the influence of environmental factors on CO emissions within different hydrological connectivity remains uncertain. Based on multiple machine learning methods, we investigated seasonal and annual CO fluxes, and the discrepancy in contributions and threshold behaviors of key variables to CO flux variations, in two river-connected lakes (Hongze and Poyang) and two non-river-connected lakes (Chaohu and Taihu). Random Forest performed well in CO flux simulations, with R values ranging from 0.91 to 0.95. Our results indicated that all four lakes acted as CO sources. Hongze Lake had higher CO fluxes in the wet season than in the dry season, while the other three lakes showed the opposite pattern. Solar radiation and pH were the dominant factors influencing CO flux variations in Hongze Lake and Taihu Lake, respectively. CO flux variations in river-connected lakes were more sensitive to total phosphorus (TP) concentration than in non-river-connected lakes, due to shorter hydrologic residence time, causing a faster response of chlorophyll-a to TP, with the lag times of chlorophyll-a responses to TP being 0-1 months for river-connected lakes and 1-4 months for non-river-connected lakes. Conversely, CO emissions in non-river-connected lakes were strongly influenced by pH, with the pH peak occurring 1 month earlier than the peak of CO fluxes during the wet season. Threshold behaviors were revealed for pH, solar radiation, wind speed, TP and total nitrogen (TN) concentrations, and lake level/volume, with non-linear responses affecting CO flux dynamics in each lake. These mechanistic findings enhance the scientific understanding on lake carbon emissions.