By Marty Graham, Contributor
With a drone in the air and the pilot working off an algorithm meant to find and identify poachers via thermal images sent from the drone, AirShepherd and University of Southern California doctoral student Elizabeth Bondi are changing the way wildlife rangers try to stop poaching.
Working at night when poachers creep around, the drone pilot spends hours studying thermal imagery where animals and humans appear sharp white. When the pilot sees poachers, wildlife rangers are dispatched.
Bondi and her team from the Center for Artificial Intelligence (AI) in Society wanted to make the drone pilot’s job easier and more effective. Even the best pilots had a hard time distinguishing animals from humans when seen from higher altitudes or as the drone flew quickly on patrol. So the team set out to create an algorithm that can do the spotting in real time, and dispatch rangers to catch poachers and save animals.
But they faced a challenge that is common in the data science trade: Machine learning involves feeding thousands upon thousands of bits of information—in this case images—to the machine learning algorithm, and that means humans have the arduous task of labeling the imagery so the machine can learn what to look for. The images have to be labeled accurately and with context. The tiny white blob that moves—is it a human or an animal? If people teach the machine with sloppily-labeled imagery, accuracy will decline.
Profitable Targets
Poaching is a profitable enterprise in central and southern national parks in Africa, where thousands of elephants, tigers, and rhinos have been hunted and killed for their tusks, skins, and horns (up to $70,000 per kilogram, higher than the value of gold).
The World Wildlife Fund says poaching is the fifth most profitable illegal activity worldwide, generating an estimated $10 billion dollars annually, and that poaching in Africa is increasing exponentially—even as the victim species of elephants, tigers, and rhinos teeter toward extinction.
According to AirShepherd, an African elephant is killed every 15 minutes on average; the population of African elephants that was more than one million in the 1970s lingers around 400,000 today. Because of poachers, Africa’s white rhino population is exactly two females, and they will not reproduce. The last Javan rhino in Vietnam was killed in 2010.
The four national parks in Zimbabwe, Malawi, and South Africa served by AirShepherd range from Liwonde, Malawi’s 212 square miles, to Hwange, Zimbabwe’s 5,657 square miles. The reserves generally have few rangers per square mile. And because a number of rangers throughout Africa have been killedby poachers, the rangers deploy in groups of no fewer than four when poachers are spotted. Despite the limited workforce and resources spread across vast savannah, deploying rangers to crimes in progress is critically important. That’s where Bondi’s AI algorithm comes in.
The Labor in Labels
Success or failure of the machine learning algorithm to help rangers accurately detect poachers depends on the accuracy and consistency of labeling. “We started with labeling historic videos and, once finished, were able to fine-tune an existing deep learning algorithm that would help do image detection,” Bondi said. “Once we had that, we were able to bring that information to our thermal imaging which is a little different than the imaging people are used to.”
The team of about a half dozen people labeled content in 39,000 videos frame by frame in about six months. “We focused on size, temperature, and shape, she explained. “We draw a little box around the image and label it human or animal.”
Each of the 39,000 images has several labels and color-coded boxes defining animals and humans, Bondi noted. For all its utility in surveillance settings, temperature-based thermal imaging didn’t help as much as hoped.
“The differences in body temperature [between poachers and humans] aren’t much, so it’s not that useful. Shape and size are factors, but when your altitude increases, size is hard to get right. A few things were easy, like groups of many, many bright dots meant we can assume it’s a herd.”
“We’re still relying on shape for the most part, but in the future it would be interesting to be able to look at motion, how the shapes move.” —Elizabeth Bondi, doctoral student
Bondi has a wish list: getting to the ability to crunch massive amounts of imagery and feed it to the algorithm to detect different types of movement—an extraordinarily data-intensive approach.
“We’re still relying on shape for the most part, but in the future it would be interesting to be able to look at motion, how the shapes move,” Bondi said. “An elephant or a rhino move very differently than humans.”
The tedium and absolute need for consistency and context in labeling is a problem throughout the AI landscape. Arguably, the largest development bottleneck in machine learning today is getting labeled training data, Stanford University researchers say.
Companies like Hive and Google’s Fluid Annotation, that promise well-labeled data sets, have popped up in the past few years, while data science professionals continue to try to shift the task of labeling from humans to machines. Stanford, MIT, and many other data science schools have research and projects underway to teach neural networks to label data. In China, cheap labor is driving the nation’s AI ambitions, but at the cost of tedious jobs for thousands of workers. And labeling is one of the most frequently asked questions on AI forums:
Q. How can I label data quickly without writing so many source code line?
A. You can’t.
With the labeling work behind them, Bondi’s team and AirShepherd are seeing good results: Rangers are catching more poachers, either before they act or when they’re in the midst of killing animals; and a large number of attacks have been foiled because rangers found where the poachers were setting traps.
“Now we can identify poachers and their location in minutes, and the rangers can go straight to where they are,” Bondi said. “It is a breakthrough and one that the poorly-funded preserves can afford.”