Conditional Benefits of Point-Level UAV LiDAR--Multispectral Fusion for 3D Vegetation Classification

Oscar Calva , Yves Plancherel

Multisensor remote-sensing fusion combines complementary measurements of landscape structure and surface reflectance, but adding a second modality does not necessarily produce uniform classification gains. This study evaluates point-level data fusion of unmanned aerial vehicle (UAV) LiDAR and
multispectral imagery within a three-dimensional (3D) point-cloud
representation. Tree--shrub classification in a heterogeneous tropical
landscape in Benguet, Philippines, was used as a diagnostic test case because
tall shrubs and lower tree canopies can occupy overlapping geometric spaces.
Green, Red, RedEdge and Near-Infrared bands, together with the Normalized
Difference Vegetation Index, were projected from separately acquired
orthomosaics onto individual LiDAR points to create a spectrally augmented
point cloud. Classification was performed using an extended implementation of
the open-source OpenPointClass framework. On the independent test set, fusion
increased overall accuracy from 83.8\% to 84.7\%, tree F1-score from 57.6\%
to 61.8\%, and tree Intersection over Union from 40.5\% to 44.7\%, while
shrub F1-score changed from 90.0\% to 90.4\%. A spatial analysis across five
tree--shrub composition groups showed that the contribution of the spectral
modality was conditional: the clearest tree gains occurred in shrub-dominated
areas where geometric ambiguity was greatest, whereas fusion produced a small decline in the balanced group. Residual displacement between the separately
acquired LiDAR and multispectral products was observed but could not be
quantitatively isolated. The results show that point-level multimodal fusion can
improve 3D vegetation classification, particularly for structurally ambiguous
classes, but that its benefit depends on landscape composition and spatial alignment between the sensor products rather than being uniformly positive.