Weddell seal monitoring in McMurdo Sound, Antarctica: a novel approach

• Main Author: Dyer, Shanelle
• Format: Thesis
• Language: English
• Published: University of Canterbury 2021
• Subjects: Antarctic; The Antarctic; Ross Sea; McMurdo Sound; Weddell; Scott Base; Razorback; Turtle Rock; Antarc*; Antarctic Toothfish; Antarctica; Weddell Seal; Weddell Seals
• Online Access: https://hdl.handle.net/10092/102260; https://doi.org/10.26021/11308

The Ross Sea Region Marine Protected Area (MPA) was implemented to understand ecosystem dynamics and ensure that the Antarctic toothfish industry was not impacting key toothfish predators in the Ross Sea, Antarctica. Weddell seals, a named focal species of the MPA, are studied to help understand the impacts this fishing industry may have in the Ross Sea. This highlights the importance of understanding and quantifying Weddell seal population dynamics over space and time. Weddell seals are one of the most abundant seal species found in Antarctica, with McMurdo Sound colony populations among those most frequently studied since the 1950s. Seal tagging, ground counts, field camera and aerial surveys are among the most common methods employed to monitor long-term populations, with more recent approaches allowing for seal counts to be derived from satellite images and citizen science platforms. While these methods have provided a wealth of spatio-temporal data on the species, they can be particularly time- intensive, costly, and not always feasible due to logistical and accessibility constraints. Here, a novel method is proposed that utilises the spatio-temporal benefits field camera images offer, while greatly enhancing the seal counting process through the use of a semi- automatic image object detector. Satellite and citizen science-derived seal counts were also analysed to assess their efficacy in providing valuable population data. Field camera images were trialled using images collected from three well-known seal colonies during the summer months (Big Razorback in 2010, Turtle Rock in 2014, and Scott Base 2018-19 and 2019-20) and implemented with different camera setups. Results show that the object detection counts were not significantly different from the reference counts manually conducted by experts. For each different survey location, the R2 ranged from 0.889 to 0.997, with no absolute errors showing a strong correlation to the number of seals in each count. Detection accuracy typically dropped when the ...