Counting Eucalyptus saligna at 45, 90 and 180 days after planting using multispectral RPAS imagery, Random Forest and U-Net

Accurate seedling counts are essential for assessing plantation establishment, identifying planting failures and supporting early silvicultural decisions. This study evaluated two complementary image-based workflows for counting Eucalyptus saligna seedlings at 45, 90 and 180 days after planting. Multispectral imagery was acquired over four commercial plantation areas using a Parrot Sequoia sensor, recording Green, Red, RedEdge and near-infrared bands, mounted on a Phantom 4 Pro remotely piloted aircraft system. Flights were conducted at 90 m above ground level, producing orthomosaics with a spatial resolution of approximately 10 cm. In the Random Forest workflow, Object-Based Image Analysis was used to classify image objects into E. saligna, weeds, bare soil and shadow. Seedling counts were subsequently obtained from the number of polygons classified as eucalyptus. At 180 days after planting, this workflow achieved an overall classification accuracy of 95% and a Cohen’s kappa of 0.92. However, its application at 45 and 90 days was limited by lower detection efficiency, extensive manual sampling and additional processing in geographic information system software. A U-Net-based model was therefore developed as the main counting approach. More than 1,000 field-referenced seedlings were used across independent training, validation and testing datasets, and predicted counts were validated against RTK field measurements. For seedlings at 90 and 180 days after planting, the U-Net workflow achieved a mean counting agreement of 97.2% across the four study areas. A separate model was developed for 45-day-old seedlings, but its performance was not satisfactory. These findings indicate that multispectral RPAS imagery combined with U-Net provides an effective and more operational workflow for counting E. saligna seedlings at 90 and 180 days after planting. Reliable counting at 45 days remains challenging at the evaluated spatial resolution and requires further methodological development.