Automated marine turtle photograph identification using artificial neural networks, with application to green turtle
Conclusions
Our motivation for developing an automated marine turtle photo ID method was to support conventional turtle survey programmes, which usually use flipper tag to identify individual animals. A turtle is likely to outlive a flipper tags t by decades, and a photo ID system can be used in the field to identify a turtle that has lost its tag. The photo ID system can also be applied to turtle in the open water, including juvenile and male turtle, since it is easy to take suitable photos of turtle in the open water. Automation is achieved using an ordinary pattern recognition exercise, in which ANNs compare the pattern of interest in two photos, and conventional turtle survey can provide the large numbers of tagged turtle photo that are needed to bring certaintly to the process of training and testing the ANNs. We have described how to develop an automated marine turtly photo ID system, as illustrated by our preen turtle photo ID system, MYDAS. Green turtle have distinctive postocular scutes, so the required photo is the side of a turtle’s head, and the scutes pattern can be adequately defined by a small number of pixels, which facilitated our work.
Our photo ID system has three aspects of interest to artificial intelligence specialists. First, it uses 50 separate ANNs to compare the scute patterns in two photos, so the outcome of the comparison is the mean of 50 separate prediction. Second, the ANNs have a single output layer neuron with a linear vtransfer function, allowing us to assess the confidence of a given image match or no-match determination. Third, improvements to the system can be made without having to re-do the initial work to isolate the pattern of interest in theoriginal turtle photo, and we have discussed the expected nature of these improvements.
MYDAS has a success rate better than 95% in correctly determining whether a new photo matches a photo in a database. One research goal is to understand the extent to which scute patterns change or fade as a turtle ages, and we anticipate that the photo ID system will need to be futher trained, as photo become available, to tolerate such things. We are planning to develop a second photo ID system for application to loggerhead turtles.