Response of lake water level to catchment low flows: insights from a combined approach of machine learning and structural modeling

Climate change and intensified human activities are increasingly disrupting hydrological connectivity between catchments and lakes, leading to more frequent and prolonged low–flow events. These conditions destabilize lake levels and amplify regional water insecurity. Poyang Lake, China’s largest freshwater lake with strong seasonal fluctuations, has experienced worsening dry seasons over the past two decades, partly due to altered tributary inflows. To investigate how the lake responds to the changed catchment inflows, we propose a hybrid attribution framework combining Random Forest (RF) and Partial Least Squares Structural Equation Modeling (PLS–SEM). RF captures nonlinear, lagged responses of lake levels to tributary flows, while PLS–SEM disentangles direct and indirect causal pathways among climate, vegetation, human activity, and runoff. Long-term records (1953–2022) revealed a structural shift around 2003, after which low–water–level frequency and severity intensified. Using runoff from all five major tributaries, RF results highlighted the Ganjiang River, the largest tributary, as the dominant driver and indicated strong cumulative and delayed effects. The lake levels responded to low flows with a lag of about one to two weeks. PLS–SEM further showed that precipitation was the primary driver of runoff, whereas rising temperature and vegetation recovery enhanced evapotranspiration and exacerbated drought. Human activities, including population growth and soil erosion, indirectly regulate runoff through changes in land water storage and vegetation. Poyang Lake thus exhibits a distinct “reduced inflow–rapid recession” mechanism. Integrating machine learning and structural modeling offers insights and a framework for lake–catchment drought management.