PubAg

Main content area

Reliability of marine faunal detections in drone-based monitoring

Author:
Colefax, Andrew P., Butcher, Paul A., Pagendam, Daniel E., Kelaher, Brendan P.
Source:
Ocean & coastal management 2019 v.174 pp. 108-115
ISSN:
0964-5691
Subject:
algorithms, artificial intelligence, attitudes and opinions, baitfish, beaches, coastal zone management, computer software, dolphins, environmental factors, fauna, monitoring, sharks, turtles, weather
Abstract:
An increase in shark bites, declining shark populations, and changing social attitudes, has driven an urgent need for non-destructive shark monitoring. While drones may be a useful tool for marine aerial surveillance, their reliability in detecting fauna along coastal beaches has not been established. We developed a drone-based shark surveillance procedure and tested the reliability of field-based fauna detections and classifications against rigorous post-analysis. Perception error rates were examined across faunal groups and environmental parameters. Over 316 shark surveillance flights were conducted over 12 weeks, out of a possible 360, with adverse weather preventing most flights. There were 386 separate sightings made in post-analysis, including 17 sightings of shark, 125 of dolphin, 192 of ray, 19 of turtle, 15 of baitfish school, and a further 18 sightings of other fauna. When examining error rates of field-based detections, there were large differences found between fauna groups, with sharks, dolphins, and baitfish schools having higher probabilities of detection. Some fauna, such as turtles, were also more difficult to classify following a detection than other groups. The number of individuals in a sighting, was found to have significant but relatively subtle effects, whilst no environmental covariates were found to influence the perception error rate of field-based sightings. We conclude that drones are an effective monitoring tool for large marine fauna off coastal beaches, particularly if the seabed can be distinguished and post-analysis is performed on the drone-collected imagery. Where live field-based detections are relied upon, such as for drone-based shark surveillance, the perception error rate might be reduced by machine-learning software assistance, such as neural network algorithms, or by utilising a dedicated ‘observer’ watching a high-resolution glare-free screen.
Agid:
6341147