Mapping dead trees in Rhode Island using semi-automated classification of summertime aerial imagery

Abstract Mapping individual dead trees from aerial imagery is useful for assessing habitat quality, monitoring forest health, identifying risks to infrastructure, and guiding forest management strategies. Dead trees can be identified in summertime aerial imagery but computer-assisted methods are needed for efficient mapping across large areas. This study evaluated pixel- and object-based unsupervised classification and Deep Learning for mapping individual dead trees using summertime true-color aerial imagery. The study area was located in Rhode Island forests which had high rates of deciduous tree mortality due to a Spongy moth outbreak in 2015–2017. The unsupervised approaches included spectral and morphological filters to reduce commission error. The Deep Learning approach used a training/validation dataset with 22,504-point features representing dead trees. The pixel-based method had the best performance (F1 = 0.84), the object-based method had slightly lower performance (F1 = 0.79), and Deep Learning had the worst performance (F1 = 0.67). The pixel-based method had the added benefits of being automatable and not requiring training data whereas the object-based method could not be automated and Deep Learning required a large training dataset. The study showed that dead deciduous trees can be mapped from true-color aerial imagery and that an automated classification can yield high accuracy.