News
Marine megafauna populations are challenging to assess, thanks to their cryptic nature and patchy availability to many forms of remote sensing. The Duke University Marine Robotics and Remote Sensing lab (MaRRS) strives to advance marine wildlife assessment methodology by fusing unoccupied aerial vehicles (UAV), advanced sensor packages and computer vision algorithms. This combination promises to improve the efficiency, economy and safety for surveys that are often tedious and dangerous for those that conduct them in remote parts of the world.
In the spring of 2015, the MaRRS lab conducted surveys over two grey seal breeding colonies in Nova Scotia using a small fixed-wing UAV called an “ebee”, taking pictures of the colonies with standard RGB and thermal cameras mounted in the belly of the aircraft. In the thermal images, seal pups and adults showed up as hot “blips” on a frigid background of ice and frozen earth, presenting an ideal opportunity to compare how humans and automated machine learning approaches detect and count animals in remotely-sensed data. The MaRRS lab computer vision algorithm proved extremely accurate, yielding total seal counts only 2% different than manual counts by humans, even tackling a long-time hurdle in automated detection by consistently discriminating seals within closely packed “piles”.
The above case study is widely applicable to species that seasonally aggregate on land, particularly pinnipeds and colonial seabirds. UAVs, by their very nature, are capable of rapid deployment and can take advantage of temporal windows where weather is good and animals are visible on land. The MaRRS computer vision algorithm operates in the common program ArcMap (ESRI), and is designed for quick modification to apply to other pinnipeds and even entirely different genera. This type of flexible and easily-modifiable model design is critical for practical applications in wildlife management. Algorithm development is time consuming and if time must be taken to extensively retrain a model for each new dataset, many advantages in efficiency are lost over traditional, manual-counting methods.
As UAVs proliferate and more data is collected, analysis becomes a bottleneck for getting relevant information to resource managers and decision makers. Combining UAVs with computer vision is a way to stay ahead of the curve and ensure that big data is an advantage and not a stumbling-block for wildlife management.
In total, 3,355 grey seals were counted in this case study led by Alexander Seymour and his team at the Duke University Marine Laboratory, North Carolina, USA and Fisheries and Oceans Canada. The locations of the identified grey seals are available through the OBIS web site titled “Atlantic grey seal breeding colonies in Hay and Saddle Islands, Nova Scotia” at http://iobis.org/explore/#/dataset/4534. The more detailed information, georeferenced RGB pictures and thermal images are available through the OBIS-SEAMAP web site at http://seamap.env.duke.edu/dataset/1462.
Reference: Seymour, A., Dale, J., Hammill, M., Halpin, P and Johnston, D. 2017. Automated detection and enumeration of marine wildlife using unmanned aircraft systems (UAS) and thermal imagery. Scientific Reports. 7: 45127. https://www.nature.com/articles/srep45127.