Classifying individual dead trees using high-resolution multispectral imagery from an unmanned aerial vehicle.

Quantifying tree mortality using conventional field surveys is very labour-intensive, limiting the number of trees that can be monitored. High-resolution imagery from Unmanned Aerial Vehicles (UAVs) overcomes this limitation, allowing for the assessment of greater numbers of trees at larger spatial extents. However, uncertainty arises in classifying individual dead trees from UAV imagery due to variation in imaging conditions and species composition across sites. In this study, we describe the performance of dead tree classification models based on high-resolution UAV imagery. We used a DJI Matrice 200 v2 equipped with a Sentera AGX-710 sensor to collect 5-cm resolution multispectral imagery across 58 boreal forest sites in western Canada. We identified individual canopy trees within these sites from local maxima of a Canopy Height Model, calculated a set of spectral indices for them, and recorded their status as alive or dead. We performed a 60-40 split into training and testing data and then fit logistic regression models to predict mortality status from crown-level spectral indices. The mean validation accuracy of our models was 99.51% for live trees and 83.08% for dead trees. Classification accuracy was influenced by species composition and lighting conditions. Classification success was higher in conifer-dominated stands (e.g., spruce and pine) than in deciduous stands (aspen). Images acquired under the overcast conditions consistently produced higher classification accuracy than those acquired under sunny and partly sunny conditions. Out of twelve vegetation indices that we tested, we found that brightness, percent greenness, GLI (Green Leaf Index), NDI (Normalized Difference Index), NDVI (Normalized Difference Vegetation Index), and NIR (Near-Infrared Reflectance) tended to perform well across different sites, lighting conditions, and models. Overall, this study demonstrates the strong potential of UAV-based multispectral imagery for classifying dead trees and predicting tree mortality.