Integrating Machine Learning and Stand Variables to Model Tree Volume in High-Elevation Pinus patula Plantations

Benavides-Huamán F.E.; Chávez-Juanito Y.A.; Alva Mendoza D.M.; Villena-Velásquez J.J.; Elera-Gonzales D.G. · International Journal of Forestry Research (2026) · DOI: 10.1155/ijfr/8870162

Resumen

Accurate individual-tree volume estimation is essential for the sustainable management of Pinus patula plantations in the Peruvian Andes, where forest inventory data and locally validated growth models remain limited. This study evaluated machine learning algorithms for predicting individual-tree volume with bark and compared them with the Schumacher–Hall allometric model under two predictor sets: a tree-level set based on diameter at breast height (DBH) and total height (H), and a stand-level integrated set including DBH, H, basal area (G), dominant height (HD), stand density (N), and age (t). A dataset of 125 felled P. patula trees from 21 plots in Chota Province was used. Random forest (RF), support vector machine (SVM), artificial neural network (ANN), and Schumacher–Hall models were evaluated using repeated grouped 3-fold external cross-validation, repeated 30 times, to account for plot-level structure and avoid information leakage. ANN with stand-level predictors achieved the best overall external validation performance (R2 = 0.9699, RMSE = 0.0371 m3, MAE = 0.0262 m3), whereas the tree-level Schumacher–Hall model remained highly competitive and produced the lowest MAPE (12.5%). Permutation importance showed that DBH and H were the dominant predictors, while basal area provided the strongest stand-level contribution. These results indicate that ANN models integrating stand descriptors can improve volume prediction stability and provide a robust tool for forest inventory and management in high-Andean P. patula plantations. Copyright

Instituciones: Universidad Nacional Autónoma de Chota