Integración y comparación de técnicas de aprendizaje automático y modelos de regresión por bosques aleatorios en el proceso de predicción de parámetros de ecuaciones de perfil fustal para la especie Eucalyptus pellita

Abstract

This study evaluates and compares the effectiveness of traditional taper models optimized using neural networks (NN) and random forests (RF) to accurately estimate the volume of individual Eucalyptus pellita trees. Nine traditional models were fitted using the Levenberg-Marquardt algorithm, with the three best models (Riemer, Bigging, and Rentería) selected, achieving coefficients of determination (R²) of 99.54%, 99.2%, and 98.57%, and RMSE values of 0.17, 0.22, and 0.29, respectively. The models were then optimized using NN (backpropagation and recurrent networks) and RF, demonstrating superior performance compared to the base models. Recurrent NNs showed lower training and testing losses, and the Riemer model optimized with backpropagation (RMSE: 0.183, R²: 0.944) stood out for its balance between accuracy and complexity, making it suitable for practical forestry applications due to its robust R² and fewer parameters. This study highlights the potential of machine learning for simultaneous prediction of stem profile parameters and volume, enhancing efficiency in forest volume estimation with significant implications for forestry planning and management.