Photo-identification is a valuable tool in the study of many large whale species especially for long-term studies of abundance and trends. Management and matching of larger catalogs, however, becomes increasingly time-consuming. Numerous automated and semi-automated systems have been developed to accelerate this process, but while some have been partially effective, their success rates especially with larger collections have not been ideal. We conducted a systematic test of two new automated matching algorithms. The first, HotSpotter, is driven by the pigmentation and scarring patterns on the flukes. The second, CurvRank, is driven by the digital measures of curvature along the trailing edge of the fluke. The test involved a sample of 2,777 cropped photographs of humpback whale flukes taken along the US West Coast by Cascadia and collaborators (primarily in 2014) as queries, representing the best image from an encounter. These were compared to a reference collection (Cascadia’s historical catalog through 2013) of over 4,041 photos of 3,235 individuals. Manual comparison required over 2,500 volunteer and 900 staff hours of matching effort. Both automated programs were very effective in finding the correct match as their top choice (74% for CurvRank and 69% for HotSpotter). The correct match was in the top two choices of one or both programs 90% of the time. This included at least 21 verified matches chosen as the top two results that were missed in the manual comparison. These were primarily found by CurvRank (20 of 21 were first rank) and mostly consisted of black flukes especially where there had been a change in fluke coloration or marks. The most common reasons for algorithm missed matches were a lack of distinct patterns (HotSpotter) or partially obscured or poorly visible trailing edge (CurvRank). These results promise to revolutionize the ability to maintain and compare large collections.
Citation:
Flynn, K., J. Calambokidis, H. Weideman, J. Crall, Z. Jablons, C. Stewart, C. Kingen, J. Van Oast, J. Holmberg. 2017. Testing of two new automated fluke identification algorithms and comparison to non-automated methods for humpback whales. Abstract (Proceedings) 22nd Biennial on the Biology of Marine Mammals, Halifax, Nova Scotia, October 22-27, 2017.