Does Multispectral RPAS Imagery Improve Individual Tree Crown Segmentation Compared with RGB Imagery in Young Eucalyptus Stands?

Abstract This study assessed whether the additional spectral information provided by RPAS-based multispectral imagery improves semi-automatic individual tree crown segmentation relative to RGB imagery in young Eucalyptus spp. stands. RGB and multispectral datasets were acquired 180 days after planting. Representative variables derived from each sensor were selected using Principal Component Analysis and used as input for k-means segmentation, followed by vector-based post-processing and comparison with reference polygons delineated through visual photointerpretation. Multispectral segmentations achieved higher Precision, shorter centroid distances, greater Coincidence Range, and lower modified MSEI, indicating more accurate segment positioning, stronger boundary agreement with the reference delineations, and lower overall geometric discrepancy. Although RGB imagery produced a higher mean Recall, this was associated with lower spatial precision and a greater visual occurrence of merged adjacent crowns, thereby limiting the reliable individualisation of trees. These findings indicate that, in young Eucalyptus spp. stands, multispectral information improves key geometric aspects of individual tree crown segmentation, especially when the aim is to derive tree-level attributes.