Estimating Individual Leaf Area of Dendrocalamus brandisii: Allometric vs. Regression vs. Machine Learning

Accurate estimation of individual (culm-level) leaf area (LA) is essential to quantifying canopy structure, productivity, and carbon dynamics in bamboo plantations, yet it is often constrained by limited field data and high measurement costs. Using destructive sampling data from 52 Dendrocalamus brandisii culms in Sanya, China, we evaluated allometric models, regression approaches, and machine learning algorithms for estimating individual LA. Candidate models were built from diameter at breast height (DBH), culm height (H), and their combinations and were assessed via five-fold cross-validation, bootstrap uncertainty analysis, and the small-sample corrected Akaike Information Criterion (AICc). DBH was the dominant predictor of LA (r = 0.826), whereas adding H yielded no improvement due to strong collinearity (r = 0.922). The simple DBH-based power-law model (LA=a·DBHb, b ≈ 1.8) achieved the best balance between accuracy and parsimony, explaining ~61% of the cross-validated variance with stable bootstrap parameters. Although Random Forest showed high in-sample fit (R2 = 0.942), its cross-validated performance was low (R2 = 0.496), indicating overfitting when using small samples. These results demonstrate that DBH-based allometric models provide a robust, interpretable, and cost-effective tool for large-scale biomass and carbon stock assessment in D. brandisii plantations.