Carbon stock assessment using machine learning based modeling in Pinus massoniana plantations

• A model was developed to characterize the relationship between above- and below-ground carbon stocks; • NET best simulated individual soil layers, RF performed better for the entire profile; • The K enables more accurate regional assessment of forest ecosystem carbon stocks; • Stand, topography, climate, and soil strongly drive K. Accurate estimation of forest ecosystem carbon stocks is vital for sustainable forest management and climate change mitigation strategies. Yet, belowground carbon quantification remains one of the greatest challenges. Drawing on literature and field data from Pinus massoniana plantations in China, this study quantified above- and belowground carbon stocks at the stand level. Using four machine learning algorithms, we constructed the relationship index ( K ) of the above- and below-ground carbon stocks model, aiming to validate a feasible method for regional-scale forest ecosystem carbon stock assessment. We quantified the effects of potential predictor variables on the K , finding that stand factors, topography, climate, and soil properties jointly explained 33%–60% of the variations in the K across soil layers ( P < 0.001). Above- and below-ground carbon stocks exhibited a positive correlation ( P < 0.001), while K decreased significantly with both increasing soil depth and arbor-layer carbon stocks ( P < 0.05). Among the models, the Artificial Neural Network achieved the best performance for individual soil layers ( R ² ≥ 0.64, P < 0.001), whereas Random Forest was more effective for the entire soil profile ( R ² = 0.50, P < 0.001). These findings underscore the feasibility of inferring soil carbon stocks from aboveground carbon through machine learning algorithms in forest ecosystems, providing a robust framework for large-scale carbon stock assessments using forest inventories and remote sensing data.