Modelling wildlife species abundance using automated detections from drone surveillance

Abstract

Reliable estimates of abundance are critical to the conservation of threatened species. Aerial surveying is a sampling method that has been used to estimate wildlife abundance over large or inaccessible areas. An increasing trend in aerial surveying methodology is to use remotely piloted aircraft systems (RPAS), also known as drones, in lieu of traditional manned aircraft systems such as planes and helicopters. Studies which used RPAS instead of manned aircraft have recently attempted to analyse imagery using automated detection methods. While there are a number of advantages to this approach, there are also potential issues in abundance estimation using this data since the errors associated with using these approaches are largely unaccounted for in established models of abundance estimation. In this paper we applied a model developed by Terletzky and Koons (2016) for fixed wing survey, to data derived from RPAS surveys that has been processed by an automated detection method for koalas. The data collected enabled ground-truthing of detections which allowed both the probability of detection and the probability of duplicate detection to be accounted for in abundance estimates, as well as a comparison between the estimates and the true number of koalas present on site. Overall, it was found that the Terletzky & Koons (2016) method resulted in artificial inflation of abundance estimates when using data collected from RPAS surveys with automated detection. This is likely to have resulted from false positive detections, which can have a considerable impact on the accuracy of automated wildlife counts. Incorporating more sources of error than the probability of detection and duplicate detection appears to be essential to improving abundance estimation for these novel survey methods. An exploration of additional covariates that could affect detection in RPAS-derived thermal imaging due the unique constraints of these technologies should be considered in future model development.