OPTICAL AND SAR VEGETATION INDICES FOR EUCALYPTUS VOLUME MODELING: A MACHINE LEARNING APPROACH

Stand Volume constitute an essential variable that integrates the various levels of forestry planning. The present studies study aim to assess the performance of optical and L-Band SAR vegetation indices from ALOS-AVNIR-2 and ALOS-PALSAR, respectively, for eucalyptus stand volume retrieval in eastern Brazil, using three different machine-learning algorithms: Artificial Neural Network, Random Forest and Support Vector Regression. We hypothesized that combination of spectral indices from optical, which interacts with upper parts of canopy, and SAR data, which can penetrates canopy surface, can may lead to a better accuracy of volume estimation. Five main indices contributed, in different levels, to volume predictions of eucalyptus stands using the different machine learning algorithms: NDVI and R (optical indices), and Pt, VSI, BMI (SAR indices), proving the complementarity of both sensors information. Random Forest algorithm were the most appropriate machine-learning algorithm for data analysis yielding an R² value of 0,778 and RMSE of 11,561 (4,578%), outperforming ANN and SVM.