Evaluating the Performance of Machine Learning Approaches in Estimating Reforested Populus (Populus Deltoids) Trunk Weights

Abstract

Measuring the weight of trees for the users of this industry is always possible after cutting the trees and measuring it directly through a scale. This method has many difficulties and problems. Hence, researchers have always looked for indirect methods of measuring the weight of tree trunk wood to overcome these limitations. The current research aimed at providing a numerical model to estimate the weight of Populus Deltoids tree trunk. It examined 400 trees in the afforested areas of the west of the Gilan province in northern Iran, in a humid-moderate climatic region. Then 11 variables of each tree were measured before cutting them. These variables are diameter at height 1.30 meters (D1.3), diameter at height 1 meter, diameter at height 2 meters, diameter at height 3 meters (D3), diameter at height 4 meters (D4), collar diameter (DC), stump diameter, crown height, trunk height (Htr) and stump height; these were independent variables (input) in modeling. Then, the weight of the trunk of the trees was got after cutting the trees by direct measurement through a scale. Pearson's correlation test showed that variables D1.3, D3, D4, DC and Htr are the most effective variables. The combination of these five variables gave the arrangement of the models’ input scenarios. After dividing the trees into calibration phases (including 300 trees [75%]) and validation (including 100 trees [25%]), the research tested three models including multiple linear regression (MLR), multilayer perceptron (MLP) and hybridized form MLP with Genetic Optimization Algorithm (MLP-GA). It reported in comparison the MLR linear model as superior to the two artificial intelligence models MLP and MLP-GA. Comparing the performance of MLP and MLP-GA models showed that the combination of GA with MLP model can increase in average the accuracy of MLP model in estimating tree trunk weight by 16.6%. This superiority showed that the linear relationship between the tree variables and the trunk weight variable is stronger than the non-linear relationship between them. The evaluation criteria of the best estimation of tree trunk weight are equal to root-mean-square error = 65.98 kg and R² = 91.93%, which belongs to the MLR model. The current approach can provide information on the amount of wood of tree planting to gardeners and users of wood industry before cutting trees. It is probably a valuable research source for other similar and different climatic regions and other species of economic trees.