As an ecologist, Joeri Zwerts often finds his feet in the mud of the jungle in Africa. In the countries of Congo, Gabon and Cameroon, Joeri investigates whether forestry with an FSC label helps to save endangered species. A fascinating but also primitive workplace, according to the NPO documentary Heroes of the Wilderness. In this article, Joeri talks about how data science is helping him further in his research.
Going back to the beginning, Joeri was looking for a better way to monitor wildlife in tropical forests, and that has everything to do with whether FSC-certified forestry is better for wildlife populations than regular forestry. "Customers like you and I pay extra for certified wood, so then you want it to make a difference." To survey wildlife, Joeri works primarily with camera traps, but a major drawback of these is the limited range of a camera. "I wanted to know if sound might be a better method. Sound has a longer range and thus that method might be much more effective and cheaper. However, to count animals in sound recordings you have to listen back to all the recordings. That takes too much time, which is why you want to automate that. Otherwise, you might as well put people in the forest and have them manually turf out sounds."
To supplement his own expertise, he sought contact with the Research Data Management Department (RDM). "RDM then tipped me off about the grants from the Applied Data Science focus area and with that I was able to pay for the hours of computer scientist Heysem Kaya. A fruitful collaboration in which we learned a lot from each other's fields. We did things that I couldn't have done on my own, and RDM developed software that they would never have thought of on their own. That's the nice thing about this multidisciplinary approach." Two, and nearly three, scientific publications have now resulted from the collaboration.
Customers like you and I pay extra for certified wood, so then you want it to make a difference.
Together with computer scientist Heysem, an algorithm was developed that detects the sounds of primates. The choice for primates is a pragmatic one, as they make a lot of noise and give a good picture of how well or badly the population of endangered species is doing. The detection algorithm requires training data. That sounds simple, but it isn't. "If you collect sound in a forest, you just have to be lucky that there are enough monkeys in that place."
Joeri found those training dates at a monkey sanctuary in Cameroon. A place where monkeys live in captivity with the great advantage of knowing for sure that there are monkeys there. All recorded sounds were labeled together with students. "We did that for 5 monkey species. With that data, we were able to train a fairly simple algorithm." The trained algorithm turned out to recognize primate sounds well.
But what if the recorded data came from a real tropical forest? So outside the shelter and with a huge complexity of environmental sounds. The algorithm still recognized too many sounds from the forest as primates when they were not, aka false positives. Joeri explains how they solved that: "We then pasted those complex jungle background sounds onto the monkey sounds and then re-trained the algorithm. That gave much better results and by mixing the different data we finally managed to develop a good monitoring method. With this method, anyone who has sound recordings of chimpanzees can use the algorithm we make openly available to monitor populations. We make the code and software user-friendly so that organizations like WWF can collect data with a simple sound recorder - and that can be as little as an old phone - and then use our algorithm. Our method further shows that with training data from, say, the local zoo, effective detection algorithms can be trained. At least if the animal makes enough noise."
"But we're not done yet. To improve the algorithm even further, we published the dataset in a computerscience challenge. Computer scientists around the world then try to make the algorithm better with state of the art techniques. In the world of computerscience this is a common way to improve techniques and gain fame by winning a challenge. For me, by the way, this was completely new. Now, together with Heysem and master students from Utrecht University, I am going to apply the state of the art techniques that came out of that challenge to our latest algorithm to increase the detection capacity even further. Again, we will make that public. So everything we develop, we do open source so that others can benefit from it."
With this method, anyone who has sound recordings of chimpanzees can use the algorithm we make openly available to monitor populations. We make the code and software user-friendly, so that organizations like WWF can collect data with a simple sound recorder-and that can be as little as an old phone-and then use our algorithm.
Joeri's project is a great example of what the Applied Data Science focus area was created for. In a previous article, Peter van der Heijden said about it, "I think that as UU we should empower researchers with tools and handles to apply data science in their own field. Ideally, when formulating a research question, a scientist should not be too guided by what he or she has mastered. It is better if you grab the content of the question by the head and then add methodology and expertise through others. That fits the current zeitgeist where people work in teams." Joeri also underlines this thought: "I think if UU invests in this kind of collaboration you can more easily achieve more. On my own I could not have done this. Now we've all become smarter."
