Predicting land use carbon budgets and unveiling nonlinear driving mechanisms using interpretable machine learning: a case study of Zhejiang province

Abstract

Terrestrial carbon budgets are a critical factor influencing global climate and ecological environments. Investigating the influencing factors and prediction of carbon emissions and carbon storage is of great significance for regional low-carbon development. However, traditional static linear models fail to capture the nonlinear relationship and threshold effects between carbon budgets and their drivers. In this study, carbon emissions coefficients and the InVEST model were employed to estimate carbon emissions and carbon storage in Zhejiang Province, respectively. Furthermore, a novel coupled prediction framework was introduced, integrating the PLUS model with the gradient-boosted XGBoost machine learning algorithm to project land-use-related carbon emissions and storage in Zhejiang Province for 2030. Subsequently, the SHAP (Shapley Additive exPlanations) model was applied to further identify the driving factors behind the spatial differentiation of the carbon budget. The case study results from Zhejiang Province demonstrate: (1) Carbon Emissions rose by 2,661.53 × 10 ⁴  tons during 2010–2020 and may increase by another 5,734.35 × 10 ⁴  tons by 2030, while Carbon Storage continues to decline, causing the provincial carbon storage-to-emission ratio (CO) to decline from 4.5 to 3.6, Most cities are facing a carbon budget imbalance. (2) Spatially, carbon emissions exhibit significant agglomeration and heterogeneity, presenting a distinct distribution pattern characterized by “high in the center, low in the periphery” within urban cores, coupled with a broader trend of “higher in the north and lower in the south” across the province. Offering deeper insights for formulating low-carbon development strategies tailored to different urban contexts and development scenarios.