Estimation of lime (Tilia begonifolia Stev.) trees height using nonlinear models

Abstract

The diameter and height of trees are essential variables for biomass prediction, carbon storage, and forest stand development. Compared with height, measuring the diameter of trees more convenient and is associated with lower cost and error. In this study, nonlinear models were used to estimate height of lime trees in the Shafaroud forests of Guilan. Lime (Tilia begonifolia Stev.) trees are distributed from low to high altitude of 1800 m in Shafaroud forests and have an important role in preserving its natural composition and stand structure. A systematic random sampling method within a 200 × 200-meter network was applied for data collection. Data were collected from 48 circular sample plots with 1000 m2 at altitudes from 500 to 950 m (parcels no. 29 and 30 in 16th compartment) as well as from 50-500 m (parcels no. 14 and 18 in 17th compartment). Modeling was performed with 12 commonly used nonlinear models and multilayer perceptron neural networks with the Levenberg-Marquardt algorithm, which has the advantage to accommodate the complex nonlinear relationships between input and output data. Performance criteria including root mean square error (RMSE), adjusted R2, AIC, and MAD were used to compare the results. Results showed the highest performances of Burkhart-Strub (1974) in mid-altitude and Stoffels-Van Soeset (1953) models in low-altitude forests, while artificial neural network (ANN) returned the highest accuracy and performance in both sites. It decreased the RMSE by 5.54% in sub-mountain and 7.35% in low-land forests compared to the best applied nonlinear models. Although the suggested nonlinear models were accurate enough for the study site, the ANN method is preferred for its higher accuracy.