Comparison of Three Techniques to Identify and Count Individual Animals in Aerial Imagery

Abstract

Whether a species is rare and requires protection or is overabundant and needs control, an accurate estimate of population size is essential for the development of conservation plans and management goals. Current wildlife surveys are logistically difficult, frequently biased, and time consuming. Therefore, there is a need to provide additional techniques to improve survey methods for censusing wildlife species. We examined three methods to enumerate animals in remotely sensed aerial imagery: manual photo interpretation, an unsupervised classification, and multi-image, multi-step technique. We compared the performance of the three techniques based on the probability of correctly detecting animals, the probability of under-counting animals (false positives), and the probability of over-counting animals (false negatives). Manual photo-interpretation had a high probability of detecting an animal (81% ± 24%), the lowest probability of over-counting an animal (8% ± 16%), and a relatively low probability of under-counting an animal (19% ± 24%). An unsupervised, ISODATA classification with subtraction of a background image had the highest probability of detecting an animal (82% ± 10%), a high probability of over-counting an animal (69% ± 27%) but a low probability of under-counting an animal (18% ± 18%). The multi-image, multi-step procedure incorporated more information, but had the lowest probability of detecting an animal (50% ± 26%), the highest probability of over-counting an animal (72% ± 26%), and the highest probability of under-counting an animal (50% ± 26%). Manual interpreters better discriminated between animal and non-animal features and had fewer over-counting errors (i.e., false positives) than either the unsupervised classification or the multi-image, multi-step techniques indicating that benefits of automation need to be weighed against potential losses in accuracy.