Fifty thousand nature shots in two days: The Bavarian Forest National Park is simplifying its wildlife monitoring with Microsoft Azure and AI

The challenge: Categorizing images manually takes an enormous amount of time

A gust of evening wind blows leaves through the light barrier. Click. The shutter release is triggered again. After nightfall, a mother lynx quietly slinks along the trail with two cubs in tow. Click. The next morning, a student sits down at her desk to sift through the previous day’s images. She manually categorizes them on her computer according to whether they show animals, humans, or nothing, and enters them into a database. A brief smile lights up her face when she spots the lynx family, but disappears again when she glances at the long list of images still waiting to be manually categorized.

“Our motto is ‘let nature be nature,’” says Prof. Marco Heurich, Head of National Park Monitoring and Animal Management in the Bavarian Forest National Park, and Professor of Wildlife Ecology at the University of Freiburg. “Our main task is to observe and analyze. We don’t intervene unless it’s to protect the wildlife in the Park.” To this end, the park managers rely on camera traps to help them monitor animal populations and document their numbers accurately. The local lynx population, for instance, is currently being monitored by a total of 120 camera traps on both the German and Czech sides of the forest. This helps the team establish not only the number of lynx, but also their home range and the health of the population. In turn, these insights form the basis for potential management measures: “The most important and time-consuming part of this process is categorizing the image material,” Heurich says. “After all, not every image shows an animal—camera traps are tripped any time something moves, regardless of whether that something is an animal, a human, a vehicle, or leaves blowing in the wind.”

In the past, Heurich and his team relied on students to manually sort out the relevant wildlife shots from the plethora of images. This was a time-consuming process—just categorizing the images in the lynx monitoring project took around 30 working days. And the lynx isn’t the only animal the park monitors: “Our camera traps take up to a million images a year for the overall wildlife monitoring project,” says Dr. Christian Fiderer, Wildlife Monitoring Officer in the Bavarian Forest National Park. “Manually categorizing and analyzing them would take one and a half years. During this time, the animal population could change again, and then the numbers would be outdated. We need to be able to respond to changes in the population right away if we are to ensure that we get more timely results.” So the Bavarian Forest National Park decided to automate image categorization using a new database solution, artificial intelligence (AI), and Microsoft Azure.

The solution: Using Microsoft Azure and AI to preclassify 50,000 images in two days

The new solution is built on open source data management and MegaDetector, an AI solution. Developed by the Microsoft AI for Earth Team, MegaDetector is an open source solution for nature and species conservation. Both solutions operate entirely in the cloud—on Azure Virtual Machines. This allows the solution to be used anywhere the Azure cloud platform is available, which is why the solution is already being used in other national parks in Germany. It can potentially be used anywhere in the world it might be needed. MegaDetector handles the actual image recognition: it uses AI to recognize what the images depict and to categorize them in the database as human, animal, or empty image—and does it at breathtaking speed.

Since the camera traps are also triggered by people, data protection is an important issue for those in charge: “Before the photos come to us or the students for detailed consideration, artificial intelligence blurs any humans to make them complete unrecognizable,” Heurich says. “This makes it impossible to identify those people, thus protecting personal data.”

And the AI is learning more than just image categorization: “Wildlife monitoring is also about knowing how many deer per hectare live in the National Park—that is, their population density,” Fiderer says. “To determine this, we need to know the distance between the photographed object and the camera so we can unequivocally attribute animals to a camera radius and ultimately avoid double-counting. In the past, we estimated this distance manually using distance markers in the captured images. We wanted to make this step faster and easier.” What used to take two minutes per image and be extremely error-prone can now be done a hundred times faster and with far greater precision. For each camera trap location, the AI was trained on image points that were previously measured manually. In this way, it learned to estimate depth in images and can thus discern the distance between an animal and the camera.

Using Microsoft Azure and MegaDetector, the Bavarian Forest National Park created a data platform that can now provide data-driven answers to relevant questions: How do, say, humans impact the local nature? Or: How does the proliferation of flora and fauna affect forestry outside the park? If, for instance, red deer migrate out of the National Park into neighboring regions, those regions might see an increase in browsing—that is, animals biting off buds, leaves, or twigs—and in the number of traffic accidents involving damage caused by wild animals. To counteract this, red deer populations are carefully controlled. “But if natural predators, such as wolves, should return to the Bavarian Forest, less hunting is required,” Heurich says. The faster image analysis enables them to take quick and timely action to maintain the balance both inside and outside the park.

 

But the Bavarian Forest National Park hasn’t yet exploited the full potential of AI for the future, Heurich says: “We’re just starting to take advantage of the possibilities. In the future, our AI should be able to recognize not only the predefined categories animal, human, and empty image, but also different species and even individual animals—say, based on their coat. This will allow us to study nature even more accurately and to focus on protecting nature, species, and biodiversity.”