Modeling forest growth and yield in Anatolian black pine stands located Ankara Forest District Directorate by artificial neural network

In this study, the inventory studies are carried out in a total of 180 sample plots of different age, stand density, and site index classes to develop growth models for pure Anatolian black pine stands distributed in Ankara Regional Forest Directory. In addition, a total of 300 trees were sampled to develop single and double entry tree volume equations and taper models. With the data obtained, (i) generalized height-diameter models, (ii) segmented taper models, (iii) single and double entry tree volume equations, (iv) stand growth models, (v) diameter distribution models and (vi) single tree diameter increment model has been developed. Both regression and artificial neural network techniques were used in the modeling phase. The suitability of the developed models to growth laws has been investigated. Adjusted coefficients of determination (Radj2) of the regression (RM) and Bayesian neural network models (BYSAM) developed to estimate tree heights are 0.82 and 0.81, respectively. Radj2 values of RM and BYSAM, which were developed to estimate the bole diameters at the heights, are 0.90 and 0.89, respectively. Radj2 values of single-entry RM and BYSAM, which were developed to estimate tree volumes, are 0.94 and 0.90. The double-entry RM and BYSAM's are 0.99 and 0.85, respectively. RM that was developed to predict stand parameters explained 87, 89, 96, 83 and 85% of the variation in quadratic mean diameter, mean height, basal area, number of trees and volume, respectively while BYSAM accounted for 83, 88, 97, 77 and 80%, respectively. RM and BYSAM that were developed to estimate the diameter increment have the same Radj2 (0.50). Weibull, Johnson SB and BYSAM, which were developed to estimate the diameter distributions, were found suitable in 139, 177 and 66 sample plots, respectively. Both RM and BYSAM have significant findings in terms of their growth laws. The suitably of BYSAM to the growth laws depends on the determination of hyperparameters such as learning rate and momentum term that control the learning process of the network. A properly designed artificial neural network model can have a good generalization capability and thus does not have the overfitting problem.