Potential Improvement of Forest Growth and Yield Modeling Incorporating Remote Sensing and Machine Learning

Forest growth and yield projection is essential for sustainable forest management planning. In Brazil, the Clutter model is widely applied for yield forecasting and commonly achieves prediction errors below 10%. However, its traditional formulation relies primarily on dendrometric variables, which may limit its ability to represent physiological growth processes. This study evaluated whether the integration of remotely sensed variables improves the predictive performance of the Clutter model and compared its accuracy with that of machine learning approaches. The dataset consisted of continuous forest inventory measurements from clonal Eucalyptus plantations in southeastern Brazil. Landsat 8 imagery was processed to derive eight spectral indices and eight texture metrics based on gray-level co-occurrence matrices. Model performance was evaluated for the traditional Clutter model, modified versions incorporating remote sensing variables, and five machine learning algorithms: Random Forest, Support Vector Machine, Extreme Gradient Boosting, Artificial Neural Network, and K-Nearest Neighbor. The traditional Clutter model showed robust performance (RMSE = 8.51%, <inline-formula> <tex-math notation="LaTeX">$R^{2} \ge 0.93$ </tex-math></inline-formula>, r = 0.91). The inclusion of remote sensing variables in the Clutter framework resulted in marginal improvements, reducing RMSE by approximately 0.05%. Machine learning models achieved greater, though still modest, gains, with Support Vector Machine presenting the best performance, improving accuracy by up to 0.7%. Among the predictors, SR, NBRI, and Haralick correlation were the most influential. These results indicate that remote-sensing-based ecophysiological variables can enhance forest growth and yield estimation, particularly when combined with nonlinear machine learning techniques capable of modeling complex interactions.