Predicting Tree Mortality Using Spectral Indices Derived from Multispectral UAV Imagery

Past research has shown that remotely sensed spectral information can be used to predict tree health and vitality. Recent developments in unmanned aerial vehicles (UAVs) have now made it possible derive such at the stand scale from high-resolution imagery. We visible multispectral bands UAV imagery calculate a set of indices for 52,845 individual crowns within 38 forest stands western Canada. then those mortality these canopy trees over following year. evaluated whether including leads more accurate predictions than derived wavelengths alone how performance varies among three different species (Picea glauca, Pinus contorta, Populus tremuloides). Our results show effectively mortality, with random model producing mean area under receiver operating characteristic curve (AUC) 89.8% balanced accuracy 83.3%. The exclusion worsened performance, but only slightly (AUC = 87.9%, 81.8%). found variation species, higher broadleaf (balanced 85.2%) two conifer 73.3% 77.8%). However, all models overpredicted by major degree, which limits use on an level. Further improvements as long-term monitoring, hyperspectral data cost-sensitive learning algorithms, training larger are necessary. Nevertheless, our demonstrate UAVs strong potential predicting annual trees.