In our last piece, we raised the topic of conservation research photography, and highlighted some potential uses: monitoring individuals, studying population dynamics, and researching behavioral patterns, to name a few. We concluded with a ‘Pandora’s box’ of questions and issues facing associated with photography as a tool: photo database management, comparing and matching photos, assessing error rates, and figuring out how to apply it all to research and management questions. What follows is a case study describing how Jerod has addressed some of these questions.
I am currently involved with a long-term study directed by Daniel Fortin, a professor at Quebec City’s Université Laval. My work focuses on behavioral factors associated with range expansion of plains bison. Estimating population size, and thus animal density, is a key factor in my study. Due to the costs and logistical constraints of tracking bison with more traditional methods, I have been looking for photography-related solutions.
I searched for software that could deal with the research photography issues I outlined in the previous post. I did find some software, but unfortunately there were few programs, and most of them were very specific to the species in question. Essentially, to work with existing software, bison were going to have to quickly develop stripes or scales. In addition, I found that the photo matching methodologies, i.e., how the software chooses a match, were quite disparate and generally difficult to repeat.
I found myself wondering, “What do you do when you need software that no one has written?” Well, first, you search through literature one more time. You try different search engines, Google Scholar and Web of Knowledge (Science), hoping for a hit. But in my case, there was no choice but to make something up.
I had a little bit of coding knowledge in the statistical and data management program R. I made contact with another researcher who was estimating deer ages by measuring horns from game camera photos using Buck Score. I also developed a list of all the things that I wanted to be able to do. Then, it was a question of whether or not I could get R to do it.
After some serious research into R’s capabilities, I figured out many of my data management “wants.” But, figuring out how to calculate a ‘similarity’ score between two photos was a new world for me. I had many discussions with colleagues possessing much more statistical experience than me. We ultimately concluded that the likelihood principle was the best way to compare a suite of data from two sources.
In a nutshell, what I came up with is what I call L-PIC, or Likelihood-based Photo Identification Code. The algorithm imbedded in the software calculates a “similarity score” between two photos. For each trait being compared between two photos, the software asks, “What is the probability that the measurements from one photo are the same as the measurements of a potential match, given what we know about the distribution of possible measurements for the trait in question?” For each potential match, L-PIC calculates the sum of likelihoods, where a higher number equals a higher probability of a match.
I have tested L-PIC using photographs of plains bison (Bison bison) taken during my field seasons in Prince Albert National Park, Saskatchewan. I focus on variation in bison horns, which I measure using facial photographs of bison. I take eight horn measurements (e.g., maximum width, outer length, diameter, etc.) and then compare each photo’s measurements with L-PIC. So far, L-PIC’s error rate is less than 6% for my bison photos.
To continue testing L-PIC, I have estimated the population size of this bison herd in 2011 using capture-mark-recapture models, photos, and other field data such as cow-calf and cow-juvenile ratios. The results revealed a population size of around 200 animals (95% CI = 170-229). To verify these results, I compared them with the park’s minimum count aerial surveys (conducted each winter since 1995). These data have provided me confidence that the population estimates generated by L-PIC are plausible. The 2011 aerial survey estimate, as expected, fell just inside the lower 95% confidence interval of L-PIC’s estimate.
The population in Prince Albert National Park is one of the few free ranging herds left in North America, and thus its persistence is important for the conservation of bison. By referring to these results, park biologists will be able to state the population size with more certainty, as they prepare their next status report for the Committee on the Status of Endangered Wildlife in Canada.
I am attracted to photo identification because of the potential to reduce stress on individual animals, to ‘ground truth’ aerial surveys, and to gain insights without the expense of GPS collars. Based on these factors, I anticipate image-recognition software may prove to be a key tool in our global research and management tool box.
Jerod is a Ph.D. student in the Département de Biologie and Centre d’Étude de la Forêt at Université Laval. This research is supported by Université Laval, Parks Canada, Natural Sciences and Engineering Research Council of Canada, and Canadian Foundation for Innovation. For more information about Jerod and his research, visit firstname.lastname@example.org.
Bethann is a freelance communications consultant, author, artist, and educator specializing in ecology, food systems, and nonprofit topics. For more information about Bethann, visit http://www.fruitrootleaf.com/home.
All photo credits: Jerod A. Merkle, Université Laval