Effectiveness of using drones and convolutional neural networks to monitor aquatic megafauna

Abstract

AbstractAquatic megafauna are difficult to observe and count due to the inaccessibility and issues of detectability. Traditional transect and helicopter counts are useful for obtaining population estimates, but they often have logistical and cost limitations. The recent proliferation of drone technology offers an innovative way of surveying animal populations. However, data collected from drones are hindered by an analysis bottleneck that increases the time needed to process them. Convolutional Neural Networks (CNNs) are an emerging category of deep learning that can automate this data analysis process. Here, we compare traditional methods with drone surveys, by detecting and counting Nile crocodiles (Crocodylus niloticus) and common hippopotami (Hippopotamus amphibious). We evaluate the utility of CNNs for object detection and quantification in complex environments. Drone counts were more accurate than traditional methods; identifying 21% more crocodiles. Where vegetation was open, hippo counts with a drone showed a similar pattern (identifying 43% more). When vegetation was dense the drone produced less‐accurate population estimates than traditional methods. CNN accuracy was limited (85%) due to the reduced training dataset available for the CNN. However, with an expanded data set, object detection is likely to be more accurate, making it more applicable for expedited and automated data analysis.