Predicting Aboveground Carbon Storage in Cold-Region Urban Forests Using Structural and Species Diversity

Urban forests are important nature-based solutions for increasing carbon storage in cities and supporting carbon neutrality goals. However, accurate prediction of aboveground carbon storage (AGC) remains challenging in urban forests. This is mainly due to their complex canopy structure, diverse species composition, and fragmented habitat conditions. Traditional approaches based only on field surveys or spectral information often fail to capture three-dimensional forest structure and related ecological processes. In this study, we focused on a typical urban forest in a cold-region city. We integrated three-dimensional canopy metrics derived from unmanned aerial vehicle LiDAR (UAV LiDAR) with field-based species survey data. Based on these data, we developed a set of structural diversity and species diversity indicators. We then used multiple linear regression, random forest, support vector regression, XGBoost, and CatBoost models to evaluate the predictive performance of different variable combinations for AGC. To improve model interpretability, we applied SHapley Additive exPlanations (SHAP) to quantify the contribution and response direction of key structural and species variables. The results showed that three-dimensional structural diversity was the main driver of AGC prediction in urban forests. Vertical complexity, canopy cover, and spatial heterogeneity played especially important roles in explaining variation in carbon storage. Species diversity showed limited predictive power when used alone. However, it provided complementary information on community composition and ecological processes when combined with structural variables. The combined use of structural and species indicators improved both model performance and ecological interpretability. These findings suggest that AGC in urban forests is mainly controlled by structural attributes, while species diversity provides additional ecological information. The proposed explainable machine learning framework improves AGC prediction and helps identify the ecological drivers of urban forest carbon storage. This study provides methodological support for urban forest carbon assessment, green space structure optimization, and fine-scale urban forest management.