While African leopards span the continent, inhabiting a variety of environments like grasslands, rainforests, deserts and savannas, their migration population changes can often be difficult to track — a difficulty that one Colorado State University student is working to change with the assistance of machine learning technology.
Cheng Guo, an international student from China and a Ph.D. candidate in the department of electrical and computer engineering, is working to develop machine learning algorithms to increase the identification of these leopards. Her work focuses on analyzing image databases created from strategically placed camera traps deployed by field researchers in Africa.
A task still done by hand, placing the tracking cameras is a labor-intensive activity for conservationists.
“They capture these images during several years because the scientists need to place (the cameras), and after (a) few months, they collect the images and change the battery or the technology,” Guo said. “And then after one year, they repeat this process.”
Guo’s machine learning program draws from a publicly available database published by Panthera, an organization dedicated to “creating a world where wild cats thrive in healthy, natural and developed landscapes that sustain people and biodiversity,” as outlined on its website.
Guo’s algorithm operates similarly to facial recognition technology, with slight variations.
“(Animal tracking software) is very limited because most machine learning (does) the person tracking,” Guo said. “So we have a lot of video cameras in the world. We can track any persons (and) know where they (are) going and which way. But the animals, because we have a limited dataset, … the small group is working on that, but we want to do that because we want the group better.”
Guo is advised by Professor Anthony Maciejewski and Associate Professor Agnieszka Miguel, chair of the department of electrical and computer engineering at Seattle University.
Their findings were recently published in IEEE Transactions on Automation Science and Engineering, where Guo served as the study’s first author. Entitled, “Automatic Identification of Individual African Leopards in Unlabeled Camera Trap Images,” the paper outlines their sampling methodology.
“The algorithm was evaluated using the Panthera dataset that consists of 677 individual leopards taken from 1,555 images,” the paper reads.
“I think (that) engineering is (a) very beautiful science because it’s, like, (a way) to help the science better, and using the current or developing the current algorithm (is a way) to do that,” -Cheng Guo, CSU Ph.D. candidate in electrical and computer engineering
Secretive in nature, the African leopard is often hard to capture on film, resulting in a limited number of images.
“(The algorithm has) a lot of training data, but currently for the leopard, you know, it’s a wild animal and elusive — very hard to capture,” Guo said. “So we have a very limited dataset.”
The dataset is fed through Guo’s machine learning program, which analyzes each image for physical characteristics unique to each African leopard pictured.
“We identify each unique spot pattern in the different leopards and use the algorithm automatically to estimate how many leopards (are) in the dataset,” Guo said.
The algorithm is designed to distinguish between the target animal species and other members of the local ecosystem and the natural environment.
“The cameras are automatically activated by motion or infrared sensors so that they may be triggered by moving animals, swaying vegetation or sudden changes in the weather,” the paper reads. “Therefore, captured images not only contain a variety of animals but frequently consist only of rocks and vegetation.”
Publicly available on GitHub, the software is free to use and can be adjusted to map other species when trained on different datasets. When designing the algorithm, Guo specifically chose the program’s language to be accessible to the general public, including those with minimal machine learning experience.
“I use Python … for the machine learning because everyone uses (it to) develop a lot of programs, so you don’t want be coding them, just using them,” Guo said. “So the Python for me, it’s the most useful language.”
The algorithm, trained on a k-medoids clustering model, groups the various images together based on the measured degree of similarity between paired images — a unique feature compared to other programs on the market.
“This algorithm is different from other methods that assume all images in a dataset are from known individuals and thus regard the animal ID problem as a retrieval identification task,” the paper reads. “The clustering is performed based on the degree of similarity between all image pairs in the dataset with the result validated using an expanded definition of the silhouette score.”
Clustering definitions of the leopard’s identifiable features are further verified through integrating a silhouette score that is defined separately from each image and compared against other identified leopard clusters.
“The silhouette score measures the degree to which an item belongs to the cluster it has been assigned by comparing the difference between the mean similarity scores for its cluster with the nearest neighbor cluster,” the paper reads.
Looking into her future, Guo is beginning to formulate her dissertation’s final stretch.
“I think I am very close,” Guo said. “Currently, I’m writing maybe half of the paper. Hopefully I will publish this summer so I can do the final defense this fall; that’s my plan.”
As she continues to advance her machine learning technology, Guo’s passion for the possibilities and achievements through engineering remains unwavering.
“I think (that) engineering is (a) very beautiful science because it’s, like, (a way) to help the science better, and using the current or developing the current algorithm (is a way) to do that,” Guo said.
Reach Katie Fisher at science@collegian.com or on Twitter @CSUCollegian.
