Integrating environmental drivers and forest structure for regional prediction of aboveground carbon storage in Moso bamboo forests

Moso bamboo ( Phyllostachys pubescens ) forests are important components of global forest ecosystems and play a key role in regional carbon cycling. Understanding the ecological mechanisms underlying stand aboveground carbon storage (ACS) is essential for improving carbon assessment. However, the roles of stand structure and its mediating effects between environmental factors and carbon storage remain insufficiently clarified. In this study, data from 532 plots across eight provinces in southern China, including Jiangsu, Zhejiang, Jiangxi, Fujian, Sichuan, Guangxi, Hunan, and Hubei, were analyzed. Four machine learning algorithms, namely random forest (RF), boosted regression trees (BRT), artificial neural networks (ANN), and support vector machines (SVM), were applied to quantify the effects of stand, soil, climate, and topography variables on ACS. Structural equation modeling was further used to disentangle direct and indirect pathways. The results show that ACS is primarily regulated by stand structural attributes, with basal area (BA) as the dominant predictor and carbon content coefficient (CCC) as a secondary factor. However, rather than acting as an independent driver, the influence of stand structure emerges as a key mediating pathway through which topography and climatic variables, such as elevation and Hargreaves reference evaporation (Eref), indirectly shape carbon accumulation. This reveals that carbon storage is governed by a coupled system in which structural development and environmental conditions jointly determine ACS formation. Some soil and climatic variables exhibited method dependent importance, suggesting context-specific effects under environmental heterogeneity. Among the machine learning models, RF achieved the highest predictive accuracy (R 2 = 0.858, rRMSE = 18.256%), effectively capturing nonlinear relationships and interactions, followed by BRT, while SVM and ANN showed comparatively lower performance. When stand structural variables were combined with environmental factors, RF and BRT consistently outperformed the other models. Overall, these findings move beyond the conventional emphasis on dominant predictors by demonstrating that the role of stand structure is context-dependent and mediated by environmental conditions. This study highlights the ecological importance of structure–environment coupling mechanisms and supports the application of ensemble machine learning approaches for robust regional-scale estimation of ACS. • Integrated ML and SEM to identify direct and indirect drivers of ACS. • Identified BA and carbon content coefficient as dominant predictors of ACS. • Topography and climatic factors affected ACS mainly indirectly via BA. • Random forest achieved high accuracy (R 2 = 0.858; rRMSE =18.256%). • SHAP used to interpret the optimal model and quantify feature importance.