Automatic classification of Eucalyptus saligna, weeds, bare soil and shadow using multispectral RPAS imagery, OBIA and Random Forest

The automatic mapping of recently established eucalyptus plantations is challenging because eucalyptus trees, weeds, bare soil and shadow coexist at a fine spatial scale and may exhibit overlapping spectral characteristics. This study developed and evaluated an object-based workflow for the automatic classification of Eucalyptus saligna, weeds, bare soil and shadow in a six-month-old commercial plantation in southern Brazil. Multispectral imagery was acquired over a 4.45 ha area using a remotely piloted aircraft system equipped with a Parrot Sequoia sensor. Radiometrically calibrated orthomosaics were used to derive vegetation indices, and Object-Based Image Analysis was applied to generate spatially coherent objects. Four hundred reference objects, equally distributed among the four classes, were used to derive the spectral samples for Random Forest classification. Mean spectral attributes were extracted, and six predictors—GNDVI, NDVI, TVI, CVI, NIR and RedEdge—were retained for modelling. Classifier performance was internally assessed using 10-fold cross-validation, in which nine subsets were used for training and the remaining subset was withheld for testing in each fold; no independent test set was used. Cross-validation yielded accuracies of approximately 86–92%, with a mean close to 90%. The internal classification assessment yielded an overall accuracy of 95.01% and a Cohen’s kappa of 0.93. Class-level F1-scores were 95.19% for E. saligna, 88.79% for weeds, 94.86% for shadow and 97.72% for bare soil. GNDVI was the most important predictor, followed by NDVI and TVI, whereas the main classification error involved confusion between E. saligna and weeds. The proposed approach successfully differentiated the main components of the six-month-old plantation and generated a detailed thematic map. This spatial information may support site-specific weed management, although independent validation across different sites, plantation ages and weed communities is required.