Quantitative Turtle Analysis Project: Machine learning with turtles
The Quantitative Turtle Analysis Project (QTAP) was created in 2019 with the goal of investigating how machine learning can be used to study wildlife populations using capture-recapture methods. QTAP has specifically been researching how digital images of the eastern box turtle (Terrapene carolina carolina) can be used by automated programs to recognize unique individual turtles, in place of a physical marking systems like scute-notching, which is the current marking protocol on our turtles. Automated photo-identification of individuals can reduce the handling of individuals and also permit the collection of data by members of the community who are not trained to handle turtles.
Capture-recapture methods require the ability to uniquely identify individuals. Historically this has been done using physical marking systems such as bands or tags or, in the case of turtles, physical notching of their carapace. The process is physically invasive and time-consuming. We are using photo-based methods of identification of turtles using machine learning methods. Turtles provide an ideal case study for the use of machine learning-based identification because they are feature rich, they have unique coloration patterns and, in addition, morphometric measurements can be automatically extracted from standardized reference points on the carapace. For example, we use the points which define the scute line intersections.
The key challenges of this effort are:
- the development of computer algorithms which automatically extract individually identifiable information from photographs;
- identification of features or measurements that provide a reasonable degree of differentiation among individuals;
- integration of that information in statistical models for population size and demographic parameter estimation.
The QTAP has two analytical objectives:
First we are evaluating the use of morphometric measurements and other classification schemes to identify turtles. For example, morphometric measurements can be defined based on scute line intersection points to define a set of distance measurements that are informative about individual identity. Presently we use a Python-based GUI to manually identify these points, however we are evaluating automated methods for identifying the scute line intersection points using deep learning models.
Second, we use the individual encounter history data to fit spatial capture-recapture models to the turtle encounter history data.
The Quantitative Turtle Analysis Project (QTAP) was created in 2019 with the goal of investigating how machine learning can be used to study wildlife populations using capture-recapture methods. QTAP has specifically been researching how digital images of the eastern box turtle (Terrapene carolina carolina) can be used by automated programs to recognize unique individual turtles, in place of a physical marking systems like scute-notching, which is the current marking protocol on our turtles. Automated photo-identification of individuals can reduce the handling of individuals and also permit the collection of data by members of the community who are not trained to handle turtles.
Capture-recapture methods require the ability to uniquely identify individuals. Historically this has been done using physical marking systems such as bands or tags or, in the case of turtles, physical notching of their carapace. The process is physically invasive and time-consuming. We are using photo-based methods of identification of turtles using machine learning methods. Turtles provide an ideal case study for the use of machine learning-based identification because they are feature rich, they have unique coloration patterns and, in addition, morphometric measurements can be automatically extracted from standardized reference points on the carapace. For example, we use the points which define the scute line intersections.
The key challenges of this effort are:
- the development of computer algorithms which automatically extract individually identifiable information from photographs;
- identification of features or measurements that provide a reasonable degree of differentiation among individuals;
- integration of that information in statistical models for population size and demographic parameter estimation.
The QTAP has two analytical objectives:
First we are evaluating the use of morphometric measurements and other classification schemes to identify turtles. For example, morphometric measurements can be defined based on scute line intersection points to define a set of distance measurements that are informative about individual identity. Presently we use a Python-based GUI to manually identify these points, however we are evaluating automated methods for identifying the scute line intersection points using deep learning models.
Second, we use the individual encounter history data to fit spatial capture-recapture models to the turtle encounter history data.