High-resolution satellite imagery, modern digital cameras and drones have opened up new possibilities for surveying remote wildlife populations. But the huge volume of data created presents its own unique challenge for researchers; could recent advances in machine learning technology provide the answer?
In a recent paper in Remote Sensing in Ecology and Conservation, an interdisciplinary team of zoologists and computer scientists from the University of Oxford and the University of Bath used machine learning to identify African elephants in high-resolution satellite images from Addo Elephant Park in South Africa and the Masai Mara in Kenya.
“Elephant surveys and censuses are the bedrock of status reporting of elephant populations,” says Ben Okita-Ouma, co-chair of the IUCN-SSC African Elephant Specialist Group and director of policy and planning at Save the Elephants, a Nairobi-based NGO. “Understanding population status helps in addressing contemporary management and conservation needs of not only the animals but their habitats too.”
Researchers have been exploring ways to survey animals through satellite imagery for 20 years. So far, efforts have mainly focussed on marine environments where a simple backdrop makes it easier to identify species such as whales. Some studies have looked for other useful information in satellite imagery such as measuring guano stains as a way to estimate the size of penguin colonies. In most studies to date, human observers have manually sifted through satellite images to identify animals, creating a natural limit on how much data can be processed.
The key issues with surveying animals through satellite imagery are the resolution of the images and the volume of data to be processed. In both these areas, improvements in technology are making satellite imagery surveys an increasingly viable option.
“In two decades, [satellite image] resolution has improved 200% but we still require higher resolution to enable us to monitor more species,” says Isla Duporge, a researcher at the Wildlife Conservation Research Unit at the University of Oxford and co-lead author of the study.
Duporge and her colleagues “trained” a convolutional neural network (CNN), a form of machine learning, to identify African elephants from satellite images. For the first time, the team were able to show that CNNs can identify large-bodied animals in the complex background of a bush and scrub habitat, performing comparably well to human observers.
“It has great potential for elephant detection in remote open areas, where the operation of aircraft or drones gives logistical challenges,” says Richard Lamprey, a remote-sensing specialist consulting for Save the Elephants.
But satellite surveying is not without its challenges. A lot of elephant habitat includes areas of thick cover where elephants like to retreat to in the middle of the day, says Okita-Ouma. While the CNN identified the same number of animals as the humans did in the satellite images, that does not mean that all animals in the images were identified.
Lamprey has also been trialing a new method of elephant surveying called oblique camera counting (OCC) for Save the Elephants . In traditional aerial surveys, fixed-wing aircraft are flown at a set altitude along transects, and observers count elephants during the flight. For his setup, Lamprey has replaced the observers with digital cameras and analyzed the images on the ground.
“The [OCC] surveys indicate that traditional counts with observers have sometimes missed over 50% of large mammals with the result that previous wildlife population estimates are biased too low,” Lamprey says. “This is welcome news at a time of pessimism for many species in Africa.”
There are, however, a few key differences in the images produced by OCC surveys compared to satellite images. Oblique refers to the angle that the images are captured at — by taking images at a 57° angle rather than the typical direct overhead satellite view, there is a higher chance of being able to capture animals stood under trees.
Even more significantly, OCC images are at a resolution of 3-5 centimeters (1-2 inches), where each pixel is equivalent to 3-5 cm on the ground, compared to 31 cm (12 in) for the WorldView-4 satellite. At this resolution, OCC imagery can be used to identify a number of other smaller species, while satellite surveying is currently limited to the larger-bodied species.
Much like the satellite image method, the OCC method results in a high volume of data that needs to be analyzed. When Lamprey surveyed a 9,560-square–kilometer (3,690-square-mile) section of Tsavo National Park in Kenya, 81,000 images were produced. Lamprey and his team have also been experimenting with machine learning to help process images and have had good success with elephants; smaller-bodied species continue to prove more of a challenge.
Another potential issue is the cost of satellite imagery. At full commercial prices, satellite surveying that at present can only identify large-bodied animals would be more expensive than aerial surveys that can identify multiple species at once.
“Maxar [owners of the WorldView satellites] are very generous with image grants for non-commercial purposes,” Duporge says, “but yes, currently cost of imagery is a central limitation.”
Despite these drawbacks, Duporge says there are still many potential advantages to surveying using satellite images. Satellite imagery is completely non-invasive, with no human presence required on the ground. This is particularly important for heavily poached populations, where animals may actively hide from human presence, and is also apt in a global pandemic when getting on the ground is a challenge for researchers.
Satellites can image remote and inaccessible areas where it’s not possible to go for geographic or political reasons. The regular flight paths of satellites allow regular repeat surveys. And, in 2021 Maxar is launching a constellation of six new satellites called WorldView Legion that will pass the same location 15 times a day, heightening the chance of surveying animals when they are active, and offering more angles of view than the traditional direct top-down view.
As image resolution, satellite coverage, machine learning and computer power all continue to improve, working with remote-sensing experts like Duporge to survey populations via satellite may become an increasingly practical option for conservationists.
Duporge, I., Isupova, O., Reece, S., Macdonald, D. W., & Wang, T. (2020). Using very‐high‐resolution satellite imagery and deep learning to detect and count African elephants in heterogeneous landscapes. Remote Sensing in Ecology and Conservation. doi:10.1002/rse2.195
Lamprey, R., Pope, F., Ngene, S., Norton-Griffiths, M., Frederick, H., Okita-Ouma, B., & Douglas-Hamilton, I. (2020). Comparing an automated high-definition oblique camera system to rear-seat-observers in a wildlife survey in Tsavo, Kenya: Taking multi-species aerial counts to the next level. Biological Conservation, 241, 108243. doi:10.1016/j.biocon.2019.108243
FEEDBACK: Use this form to send a message to the author of this post. If you want to post a public comment, you can do that at the bottom of the page.
You may republish this article, so long as you credit the authors and Mongabay, and do not change the text. Please include a link back to the original article.