The status quo
Imagine sitting in the back of a Cessna 182, flying 300ft above ground at a groundspeed of over 110 miles per hour. You're a rear-seat observer (RSO) staring out the window to your right, scanning the ground for elephants, elephants that have likely wallowed in the mud or taken a dust bath and are more camouflaged than one would expect. You finally see one, then a second, and a third - a large mother and her two young calves. Now, as this plane is zooming by, you need to count as many in the herd as you can.
The plane is flying in straight path traversing a section, or transects, of the national park, in an African savanna you are flying over. Your goal is to count the number of elephants you see in a specific field of view, called a strip width, on each side of the plane. Your colleague on the left side of the plane, another RSO, is doing the same. Once all the transects are complete and you and your fellow RSO have recorded all the observations, you can extrapolate that number to arrive at an elephant population estimate for the entire park, based on what percentage of the park the transects covered and other factors like total available habitat. This method is called a sample count.
Uncertainty surrounds such population estimates, as high as 30% even. Determining a precise number is difficult if not impossible when covering so much land. But it's this method, and others like it, that have been used to examine African savanna elephants for decades, including for 90% of the Great Elephant Census
. (Different methods are used to count African forest elephants for the Great Elephant Census Forest Initiative, which we will explain in a future post!)
Sometimes, the method makes use of photos instead. Photos are taken as the plane flies these transects, and the pressure of counting in real-time is relieved. The photos are reviewed, manually, one by one and elephants are counted. However, geospatial data needs to be taken into account (i.e. overlapping photos), to make sure the elephants aren't double counted here either. The process is time-intensive, tedious, and not to mention expensive. A survey of one park can result in tens of thousands of photos, where less than 1% of the photos may contain any elephants, and you need to find someone or some group to review all these images.
Why even count these species?
Too many animal species are endangered these days; many are protected by various national and international laws and regulations, like the U.S. Endangered Species Act (ESA) and the Convention on International Trade in Endangered Species of Wild Fauna and Flora (CITES
). Determining endangered status requires understanding the current population status of a species, as well as the population trends and threats. Counts and trends give governments, nations, organizations, the information needed to better design and enact policies and programs to protect the species that need it most and in the ways they need it most. Under the ESA and CITES, government agencies are mandated to have the best scientific data available on such species; however, they are often restrained by the resources, costs, and time required of the status quo method described above.
So what can we do?
Sometimes people ask why don't we just tag all the animals and track them. While this has considerable value for understanding habitat use, dispersal and migration, it is of limited value in inferring overall population status and trends. Tagging an animal is dangerous for the animal and the human. Many animals need to be tranquilized to be tagged, creating a stressful situation for them. If you spook a large animal before it's been tranquilized, it may charge and attack or worse be harmed or die. Not to mention the logistical side of recharging the tracker/sensor, replacing batteries, or handling other malfunctions. The effort would also need to be continuous, constantly searching out and tagging any newborn animals as well. On top of this, it’s expensive and at the end of the day only a small number of animals in a population will ever be tagged.
Now, with the advances of machine learning, what if we can train a computer to find and count these species for us? That's what the latest project from Vulcan and the new Vulcan Machine Learning Center for Impact (VMLCI) aims to do. With our project, “Modernizing Wildlife Surveys with Machine Learning” (MWS), we're working with Wild Me
, a non-profit based in Portland, Oregon and colleagues in East Africa to develop a platform specifically trained on aerial imagery, the kind that is used to generate these population estimates. If all goes well, we'll be able to release the platform, open source, to allow for other organizations and entities to use it, or even add to it, by training it with their data - it could be a system that can derive population estimates for elephants and eventually other large game, like giraffe, in African savannas to polar bears in Arctic tundra.
MWS is not without its challenges though - collecting the data alone, can be time consuming and expensive. We need thousands of photos of a species taken from the vantage point that we need the platform to work at. Then the photos need to be annotated, or labeled, to train the machine learning model. While the above image is meant to illustrate what mean by annotate, the elephants in the image are actually appear much larger than what we need to train against. Take the next image for example. The elephant bounded by the blue square only takes up a 10x10 pixel space, and we need to train a machine learning model to know that 10x10 pixel blob is an elephant, but that other blob is a boulder.
Even with these challenges, we look forward to making population estimates easier, faster, and cheaper to obtain and at the same time, decrease uncertainty around these estimates to better detect trends in population changes for the various species. Stay tuned for future posts around MWS and wildlife conservation!