Spotting Cheetahs: Identifying Individuals by Their Footprints

The overall goal of this technique is to develop a footprint identification algorithm to provide conservation biologists with a noninvasive and cost effective tool for identifying individual cheetah from digital images of their footprints. This method can help address key challenges in wildlife conservation. For example, by providing reliable data on the number and distribution of individual animals.

Thus helping to mitigate human wildlife conflict and to prevent illegal poaching of endangered species. The main advantage of this technique is that it provides a robust, accurate, cost effective and noninvasive approach for conservation monitoring. It also draws on the tracking skills of indigenous communities, consequently helping to revitalize their traditional ecological knowledge.

In the early morning or late afternoon, for maximum light contrast, lay in approximately one centimeter deep path of builder sand, taking care that there is a two to three meter wide run between three to 15 meters along a habitual movement path. Using standard gardening tools, wet, and smooth the substrate to improve the print quality and definition of the path, manually removing any leaves and pebbles. When the path is ready, lure a cheetah habituated to handling across the path with a food reward, leading the animal away from the path once the footprints have been made.

Next, draw a circle around each left hind footprint to highlight the position of the individual prints along the trail. Here, I can tell that these are right prints. If you look at your own hand, you will see that from the index finger you go up and then you got a row of three tapering down.

And you can see that the cheetah's footprint actually does a very similar thing. So, this angle is a right print. Then, without touching the print place a metric scale about one centimeter below and to the left of the footprint with a photo ID slip on which the animal ID, the date, the footprint series and the photographer name are written.

Then, make sure the camera is level and point the lens directly overhead to avoid any parallax error in the image with relation to the scale or photo ID slip, taking care that the footprint, scale and photo ID slip completely fill the frame. After imaging each footprint trail, brush the tracks away and prepare the surface for recording the next trail. To extract the image features, open the footprint identification technique add-in, in the data visualization software.

Next, select Image Feature Extraction in the footprint identification technique application window, and use the mouse to drag and drop the first footprint image into the image feature extraction window. Using the Resize button, adjust the footprint image so that it fits inside the graphics window. Then, click to place markers on the lowest points of toes five and two, and select Rotate.

Rotate the image horizontally on the line joining the points to standardize the orientation. A set of cross hairs will automatically appear. Make sure the scale factor box is set to 10 centimeters and then place two points 10 centimeters apart on the ruler.

Then, use the template to place 25 sequential, defined anatomical landmark points on the print. A prompt will appear to guide and track the sequence of the points. The accurate and consistent placement of the landmark points is critical to the resulting robustness of the classifier algorithm.

Users must be familiar with the anatomy, orientation and laterality of the footprint to be able to distinguish it from any other noise in the image. When all of the landmarks have been placed, select Derived Points to generate an additional 15 landmark points and complete all of the data fields for the footprint image. Then, click Append Row to send the 136 scripted distance, angle and area variables to a row in the database.

When all of the footprint features have been cataloged, copy all of the rows in the database and paste them below the database to generate the reference centroid value. For pairwise, robust, cross-validated, discriminant analysis from the main menu of the image analysis software select and open the robust, cross-validated Pairwise Analysis window. Next, for the Input X Model Category, select Cheetah.

And for the input trails, select Trails. The Y column footprint measurements, as continuous variables, are automatically populated. Click Run.

A progress bar illustrating the analysis progress will appear, as well as a data table displaying the pairwise comparisons of the trails. As the footprint features from the database are analyzed, two outputs, an assigned self, non self table to describe the classification distance between each validation pair and the classification matrices of the different trails selected for comparison, and the contour probability will be shown. Confirm that the Show Model button giving the variables used for each comparison and the Distance Threshold box giving the distance between centroids are also shown.

Then, select the Clusters button at the base of the assigned self, non self table to display the two tables. The first table displays the distance between any two trails and the second table is used to generate a cluster dendrogram. To visualize the classification clusters, click on any branch of the dendrogram to color code it.

Finally, select the algorithm that consistently gives the highest accuracy. Adjusting the the threshold value to set the algorithm to produce the outcome that best approximates to the known number of animals in the training database. The accuracy of the algorithm is determined by the identity and the number of the variables, the contour probability and the threshold value that determines the number of clusters.

In these graphs, the outcomes of a pairwise comparisons of trails from the same individual and two different individuals based on the footprint identification technique customized model are shown. The classifier incorporated into the model is based on the presence or absence of overlap between the ellipses. This same algorithm can then be used to identify unknown cheetahs.

For example, these dendrograms of a sample of trails from several cheetahs illustrate the correct prediction when the algorithm is optimized and when the algorithm is sub optimal. Holdback trials validate the algorithm derived from the training set of known individuals. The varying test size is plotted against itself, the predicted value for each test size iteration, and the mean predicted value for each test size.

Indeed, even when the test set size is increased considerably compared with the training set size, the mean predicted value is similar to the expected value. The accuracy of the individual identification is consistently greater than 90%for both the predicted number of individuals and the trails from the same individual versus those from different individuals that were classified correctly, as illustrated by this cluster dendrogram representing all 38 individual cheetahs analyzed. From 110 trails and 5, 886 pairwise comparisons, there were 46 misclassifications, giving an accuracy of 99%This technique can be stand alone or it can be a complement to other noninvasive monitoring techniques, such as camera trapping and DNA sampling.

It can also be used to identify the age, class and sex of the individual animals. One major challenge in conservation is getting enough data on the range of big cats in the wild. This technique has a lot of potential application in that regard because it can greatly augment the amount of data we're able to collect for these populations.

Since its development, this technique has paved the way for researchers in the fields of law enforcement and medical and scientific imaging to explore reliable, objective and accurate methods for classifying objects and images.