Machine learning and regression models to predict multiple tree stem volumes for teak

Abstract

The quantification of a stand’s wood stock is one of the most important procedures for Tectona spp. (teak) management. An optimal method for estimating tree volume must accommodate the variation of the data collected in the inventory. This study evaluated alternative methods to estimate the volume on stems of teak trees. We cut and measured the outside bark, inside bark, and heartwood diameters at heights of 0.1, 0.5, 1.5 m and every meter thereafter until the minimum outside bark diameter reached 3 cm, using 180 trees of ages 3 to 12 years. We tested two approaches (A1 and A2) to estimate stem volumes of the outside bark, inside bark, and heartwood (v ob, v ib and v hw, respectively): modelling the tree volume in A1, and the taper model in A2, with the techniques of regression, artificial neural network (ANN) and support vector regression (SVR). In addition, we obtained the percentages of heartwood, sapwood and bark in the stem. The accuracies of the estimates were evaluated using bias, correlation and root-mean-square error (RMSE). Validation was made for the outside bark, inside bark and heartwood volumes, and the outside bark, inside bark and heartwood diameters, from the taper models. In A1, the ANN technique more accurately predicted v ob and v ib, with RMSE% values of 18.88% and 18.54%, respectively; for v hw, the regression technique was more accurate, with RMSE% equal to 26.87%. In A2, the regression technique obtained the highest precision in the prediction of v ob, v ib and v hw, with RMSE% values of 13.99%, 13.31% and 26.50%, respectively. Approach A2 showed more accurate results compared with A1 for predicting the multiple volumes of teak trees with the three tested techniques.