Modelagem do crescimento e produção de florestas plantadas usando indicadores de capacidade fotossintética derivados de sensoriamento remoto e aprendizado de máquinas

Abstract

Forest growth and yield modeling is fundamental for sustainable planning and management of planted forests. The Clutter model, widely used in Brazil, relies exclusively on conventional dendrometric variables (age, basal area, and site index), without considering physiological factors such as photosynthetic capacity. This study aimed to evaluate the incorporation of proxy variables for photosynthetic capacity, derived from remote sensing, into forest growth models, comparing the performance of the traditional Clutter model, its modified variant with spectral and textural indices, and machine learning algorithms. The research was conducted in commercial eucalyptus plantations in the Vale do Rio Doce region, Minas Gerais, using continuous forest inventory data (2013-2019) provided by CENIBRA. Reflectance values were extracted from cloud-free Landsat 8 images, calculating eight spectral indices (NDVI, GNDVI, SAVI, SR, NBRI, RVI, DVI, EVI) and eight texture indices based on the Gray Level Co-occurrence Matrix (Energy, Entropy, Correlation, IDM, Inertia, CS, CP, HC). Five machine learning algorithms were tested: Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Artificial Neural Networks (ANN), and K-Nearest Neighbors (KNN). Validation was performed through 5-fold cross-validation, using RMSE, RMSE%, Pearson correlation (r), R², and BIAS% metrics. Results demonstrated that the traditional Clutter model exhibited high accuracy, with RMSE% below 10% for volume, R² greater than 0.93, and r = 0.97. The incorporation of spectral and textural indices into the Clutter model promoted modest improvements, reducing RMSE% from 8.51% to 8.46% with SR and NBRI indices. However, machine learning algorithms, particularly SVM, demonstrated superior capability in incorporating these variables, resulting in accuracy gains of up to 0.7% compared to the traditional model. The SR, NBRI, and HC indices stood out as the best predictors. It is concluded that the inclusion of remote sensing variables can discretely enhance the predictive capacity of traditional models, with machine learning algorithms being more efficient in incorporating these variables, indicating a promising path for improving forest growth and yield models.