Attribution of changes in vegetation structure and carbon–water fluxes in the Ji-shaped bend of the Yellow River based on process-guided machine learning

Arid and semi-arid ecosystems play an important role in the global terrestrial carbon cycle, but the relative contributions of climate change, elevated CO₂, and land-use change (LUCC) to their carbon–water processes remain uncertain. In this study, we developed a process-guided machine learning (PGML) framework integrating Biome-BGC and LightGBM, which achieves synergistic simulation of mechanistic constraints and data-driven modeling through residual learning. PGML significantly improved the pixel-scale simulation accuracy of LAI, GPP, and ET in the Ji-shaped Bend of the Yellow River, a typical arid and semi-arid region of China, with R 2 values of 0.70, 0.95, and 0.91, respectively, and reduced RMSE by 61%–83% compared with the calibrated process model. Full-factorial attribution showed that LAI, GPP, ET, and WUE all increased significantly during the study period, with the growth rate of GPP exceeding that of ET, thereby promoting an increase in WUE. Under the process representations and driving factors considered in the PGML framework, elevated CO₂ was identified as the largest simulated contributor to the increases in LAI, GPP, and ET, with contribution rates of 71.8%, 78.7%, and 75.2%, respectively; climate change contributed approximately 15%, while the direct contribution of LUCC was less than 0.25%. Meanwhile, the regional mean climate–CO₂ interaction effect was negative, indicating that local climatic conditions generally weakened the CO₂ fertilization effect. These results reveal that elevated CO₂ made the largest contribution to the enhancement of vegetation structure and carbon–water fluxes in the Ji-shaped Bend of the Yellow River over the past two decades.