杉木枝下高预测:非线性混合效应模型与机器学习的比较OA

To select efficient and low-bias prediction methods for clear bole height,measured data from 2 515 Cunninghamia lanceolata trees across 35 sample plots at Jiangle State-owned Forest Farm in Sanming City were used as research materi-als.This study systematically compared the predictive performance of nonlinear mixed-effects models and machine learning algorithms for clear bole height estimation in C.lanceolata plantations.A generalized nonlinear model was established via reparameterization,which was further extended into a mixed-effects model by incorporating random effects and variance functions.Meanwhile,random forest(RF)and extreme gradient boosting(XGBoost)models were separately developed.Five-fold cross-validation was adopted,and the Optuna algorithm was utilized to optimize model hyperparameters.The re-sults showed that the XGBoost model achieved the highest fitting accuracy on the training dataset,with a mean absolute er-ror(EMA)of 0.813 5 m,root mean square error(ERMS)of 1.109 2 m,and coefficient of determination(R2)of 0.888 4,demonstrating the superiority of machine learning in capturing complex nonlinear relationships.However,the nonlinear mixed-effects model with plot random effects yielded the optimal prediction accuracy on the independent test dataset,with EMA=0.995 0 m,ERMS=1.349 9 m,and R2=0.848 7,reflecting the generalization advantage of mixed-effects models when handling nested data structures.The nonlinear mixed-effects model integrating individual tree,stand,and competi-tion factors remains the mainstream approach for clear bole height prediction due to its strong interpretability,despite its complicated modeling workflow.In contrast,machine learning algorithms require no pre-specified functional forms and fea-ture high modeling efficiency.Although they deliver slightly lower accuracy on independent test sets,they possess great application potential.By optimizing key predictive variables and expanding sample sizes,machine learning is expected to achieve predictive performance comparable to or even superior to mixed-effects models with simplified input variables.