Decoding Multiscale Carbon Emission Dynamics: A Systematic Framework Leveraging Land Use Patterns and Explainable Artificial Intelligence

Carbon emissions are a key driver of global climate change, and their accurate characterization and prediction are essential for promoting regional sustainable development.This study develops a novel carbon emission modeling framework that integrates interpretable machine learning with land use data, without dependence on socio-economic indicators.This allows for emission prediction beyond the decoupling stage and captures spatial distribution patterns with high accuracy and interpretability.The Yangtze River Delta region is used as a case study.By combining Geographic Information System-based kernel density analysis for land-use refinement with an optimized Extra Trees Regression model, the approach achieves high predictive performance (R² = 0.86 for testing set).Model interpretability is enhanced using Shapley Additive exPlanations, which reveal nonlinear effects of various land types on emissions.Future land use scenarios are simulated using the Optimized Land Expansion Analysis Strategy and a Cellular Automaton based on Multiple Random Seeds, achieving overall accuracy above 85%.Projections estimate total carbon emissions in the Yangtze River Delta will reach 1,581 million tons by 2030, with Shanghai and Suzhou contributing 224 and 172 million tons, respectively.County-level emissions are further examined using spatial econometric models, which identify a significant clustering pattern (Moran's I = 0.61).As industrial land becomes more dispersed, emission hotspots shift toward regional centers.