Objective Plantation-grown Pinus koraiensis is an important timber and ecological tree species in the forest regions of Northeast China. The longitudinal variation in its stem wood density is crucial for wood quality sorting, and for estimating biomass and carbon stocks. However, there is currently a lack of longitudinal models for Pinus koraiensis stem wood density, and it is difficult for traditional models to accurately capture its longitudinal variation characteristics. This study aims to analyze the longitudinal variation pattern of Pinus koraiensis wood density and develop a high-accuracy predictive model to provide a basis for the precise assessment of wood properties and biomass.

Method Based on 402 wood density samples collected from 30 Pinus koraiensis trees in Mengjiagang forest farm, this study first selected the optimal basic model from nine candidate wood density models. On this basis, traditional and easily accessible stand variables (number of trees per hectare, basal area per hectare, average stand height, average stand diameter at breast height) were introduced to construct a generalized linear model, and the continuous autoregressive error structure CAR(1) was adopted to reduce the autocorrelation of model residuals. Meanwhile, a wood density model was constructed based on the theory of generalized additive models (GAM), and the fitting effects of the six smoothing functions of GAM were compared. Finally, the predictive performance of different modeling approaches was compared using ten-fold cross-validation method.

Result (1) The wood density of Pinus koraiensis exhibits a distinct nonlinear pattern along the stem longitudinal gradient, showing an overall trend of initially decreasing and then increasing with relative height. (2) Among the base models, our self-developed Model 9 achieved the best fitting performance. The generalized linear model developed based on Model 9 by introducing trees per hectare further improved the model fitting performance. Compared with Model 9, its R²_adj increased by 2.4%, MAE decreased by 0.7%. (3) The application of a continuous autoregressive error structure of order 1, CAR (1), effectively reduced residual autocorrelation in both the basic and generalized linear models. After correction, the MAE values of the two models decreased by 2.4% and 2.5%, respectively, compared with those before correction. (4) The generalized additive model (GAM) constructed using cubic regression splines (CR) with variables including DBH, mean annual increment of DBH, crown width, relative height, and number of trees per hectare exhibited the best predictive capability. The ten-fold cross-validation results further confirmed that the GAM has superior robustness and generalization ability. Compared with the generalized linear model, its R²_adj increased by 13.25% and the MAE decreased by 9.54%.

Conclusion This study systematically constructed a longitudinal wood density model for the stems of Pinus koraiensis, and revealed its longitudinal variation pattern of "first decreasing and then increasing". The generalized additive model is superior to the basic model and the generalized linear model in fitting performance, prediction accuracy, and stability, and can accurately estimate the wood density at any stem height of Pinus koraiensis. This model provides a reliable tool for segmented wood quality assessment, precise biomass and carbon stock estimation, and sustainable management of plantation-grown Pinus koraiensis in the region.