Crown geometric descriptors versus deep convolutional networks for classifying plus trees of Neltuma pallida

Arcila-Diaz J.; Chinchay-Campos J.; Capuñay-Miñope J.; Atalaya-Urrutia W. · Trees, Forests and People (2026) · DOI: 10.1016/j.tfp.2026.101500

Resumen

Restoration programmes in the Peruvian equatorial dry forest depend on plus trees of Neltuma pallida, which are still selected tree by tree in the field. This study asked whether crown descriptors extracted from multispectral imagery acquired with an unmanned aerial vehicle reproduce the plus-tree labels of an open dataset as well as deep convolutional networks. In that dataset a plus tree had to meet minimum dendrometric criteria that include a crown diameter of 10 m, so the target is the assigned label rather than genetic quality. Fifty-one informative morpho-geometric, spectral and background descriptors were extracted from 999 images, and three classical algorithms were compared with three networks trained on the full scene and on the isolated crown under a common stratified five-fold cross-validation. Random forest reached 0.949 accuracy (95 % CI 0.933–0.961) and ResNet-50 on the isolated crown 0.950 (0.935–0.962); their paired difference, −0.1 percentage points (95 % CI −1.2 to 1.0), lay within a ± 2-point equivalence margin (TOST p < 0.001) in each of four training seeds. Crown size dominated: the 10 m crown-diameter criterion applied to the supplied mask reproduced 92.3 % of the labels without any fitting, and on 59 pairs matched for crown area the descriptor models and the isolated-crown networks fell to 0.58–0.65. Occlusion and background-swap tests showed that the full-scene networks relied mainly on the scene context. Because the 999 images derive from 688 photographs and carry no coordinates, all estimates are upper bounds of internal validation. Within these limits, descriptor-based models matched the best network at a fraction of the computational cost, and they support a screening prototype that reproduces an explicit size criterion rather than the identification of genetically superior trees.

Instituciones: Universidad Señor de Sipán