en de

Online Magazine

AI scares off wolves

With the reintroduction of wolves to Germany, wolf depredations are also on the rise. Anna Förster and her team from the University of Bremen are developing an AI-supported pasture fence that will recognize wolves and scare them off individually. A conversation about deterrence measures and how they can chase away elephants as well as wolves.


Mrs Förster, what is the current situation in Germany regarding wolves and grazing animals?
In the monitoring year 2021/2022, the Federal Ministry for the Environment registered a total of 161 wolf packs, 43 wolf pairs and 21 territorial lone wolves in Germany – in 2020/21, there were still 157 packs, 27 pairs and 19 lone animals. Especially here in the north, there are quite a few wolves, and we regularly hear about wolf attacks. Last year there were several large incidents in which around 30 sheep were killed in one night.

As we all know, this has led to the formation of two camps: One says that conspicuous wolves should be shot immediately, the other demands that they should not be bothered in any way. We are somewhere in between: In our opinion, we have to learn again how to deal with wolves and how to live alongside them. But at the same time, we have to make sure that the amount of conflict stays minimal.

What are the current recommendations for a supposedly "wolf-proof" pasture fence?
There are recommendations adapted to different grazing animals in terms of minimum height and electrical voltage – for sheep and goats, for example, it’s 120 centimetre high fences with at least 4000 volts. In addition, there are specifications for the number of wires and the undermining protection, i.e., how deep the fence must be buried.

The security specifications for fences are sufficient in theory, but fail in reality: for a gate, for example, you cannot create undermining protection.

It seems, however, these recommendations are not always sufficient. Why is that?
The specifications are sufficient in theory, but fail in reality: for example, it is not possible to create undermining protection for a gate. Or it can happen that a footpath leads past the fence and there are benches – which serve as a kind of "stepping stone" for the wolf. In some cases, such as in the case of travelling shepherds, the requirements are generally not applicable.
Such individual cases are often not properly resolved. And you can't check every single fence for safety – that's far too expensive, and there aren't that many wolf experts.

How exactly would your solution contribute to making a fence wolf-proof?
Our fence, or rather the deep learning model associated with it, would detect on camera images when wolves approach the pasture and trigger various measures to scare them off.

First, we train the deep-learning model to recognize wolves from images. To do this, we are currently working with various wildlife parks. We hang our cameras at their enclosures and then collect images – in different weather conditions and at different times of the day and night.

We have already tested several variants for image recognition:
On the one hand, thermal imaging cameras; but these have turned out to be completely useless. First of all, a wolf does not radiate heat very strongly because its fur insulates so well. Furthermore, especially in summer, various elements in the environment are warmed up during the day, such as stones or puddles. These are still so hot hours after sunset that they radiate more than a wolf. At most, you could still distinguish whether a heat source is moving or not, but since you can't see any real shapes, but simply herds of heat, you still don't know whether it's a wolf, a deer or even a human being.

Another method we are trying out is a laser rangefinder. It calculates the distance between an object and the measuring device and then reconstructs a 3D shape. This works very well at night or in rain and snow – if a snowflake lands on a camera, it can no longer take pictures. This does not happen to the laser rangefinder. The only downside: nothing but the shape is registered.

The most promising method is a night vision camera. During the day you have a normal color image, and at night a black-and-white image. On both, you can recognize a wolf very well and could even distinguish individual wolves, depending on the situation.

But I wouldn't call this fence absolutely "wolf-proof" either: a very hungry wolf can't be stopped for long by any kind of fencing. Thus, we play for time and try to keep the wolf occupied until other measures can be taken. This may well work for several nights.


AI offers new solutions not only to live alongside wolves, but also in many other areas of sustainability. Especially companies, who increasingly recognize the importance of sustainable production cycles and business models, can make use of it.

Read in this report how you as a company can become more sustainable with the help of AI and what benefits this brings.

And in a second step, these images would be analyzed by the deep learning model?
Exactly. The images are loaded into a Raspberry memory that runs the deep learning model. This means that the image analysis does not take place in the cloud but with the help of edge computing on site and in real time. On the one hand, this is an advantage in terms of data security, and on the other hand, we are independent of external networks.

So now the Deep Learning model has recognized a wolf. How does it inform the owner of the grazing animals? And how is the wolf afterwards scared off?
Our test users already have precise ideas about how they want to be alerted when a wolf appears. For example, they first want to confirm themselves that there is indeed a wolf in the picture before further measures are triggered or authorities are alerted.

If it is confirmed to be a wolf, we start a series of deterrence measures. The whole situation should be unfamiliar and unpleasant for the wolf, but not cause any permanent damage. One of these measures is based on ultrasound, to which wolves react much more sensitively than humans. In addition, one can also disorientate the wolf with the use of strobe-like lights.

Of course, there is a risk that a wolf will get used to the same deterrence measure over time. That's why we try to apply them in a random sequence, so that he is deterred by new stimuli over and over again.

When our AI detects a wolf, we start a series of deterrence measures. For that we for example use ultrasound or strobe-like lights.

How do you ensure that these deterrence measures do not disturb the grazing animals?
We test all measures on wolves as well as on grazing animals. However, grazing animals are by far not as skittish as one might think. You can also train them to ignore flickering lights or noise over time, while it still has a deterrent effect on the wolf.

Could the mAIn fence also be used to deter other wildlife?
In principle, you can train the fence just as well on wild boar, lynx or bears as on wolves – on all animals that could be a problem for grazing animals. In some areas, this also includes free-roaming dogs.

And wildlife detection can of course be thought of even more broadly: we currently have a project in Sri Lanka where we want to build a kind of virtual fence that recognizes when a herd of elephants is approaching a village by means of ground vibrations. It is not uncommon for such herds to destroy entire villages. In fact, it would be easy to drive them away, for example by banging pots together and making noise. But to be able to do that, of course, you have to know that they are coming.

In this case, however, we don't work with cameras, but with acceleration sensors in the earth. And the detection is then done by a neural network or a support vector machine.

In principle, you can train the fence just as well on wild boar, lynx or bears as on wolves – on all animals that could be a problem for grazing animals.

If your AI solution is so efficient at scaring wolves off – do you still need a fence at all?
Theoretically, you wouldn't need a fence to protect the grazing animals if you had our AI solution. However, without a fence you have the problem that you can't control in which direction a wolf takes flight. Besides, the grazing animals still have to be contained.

But what we are thinking about is to combine our solution with a very simple pasture fence – then you don't have to build a massive, expensive metal fence, especially not in places like the mountains where that is not possible. The fence then only serves to fence in the herd and as a psychological barrier for the wolf so that it runs away from the herd. And the actual wolf protection is taken over by our AI solution alone.

The discussion about the wolf is extremely emotionally charged. Could the mAIn fence help relax it a little?
It's difficult to say. First and foremost, I think it's good that we're adding another dimension to the current pro and con camps. It would be nice if the discussion would move away from this either-or, black-and-white, and if other researchers would point out more possibilities to deal with it. For example, in other contexts where our solution works less well. I hope that we have given an impulse to this.

About Anna Förster

Prof. Dr. Anna Förster has been a professor in the field of Sustainable Communication Networks at the University of Bremen for 8 years. She combines AI techniques such as machine learning with wireless communication protocols and applications. This way, she supports sustainable development, e.g. in agriculture or also in the field of environmental monitoring. Förster has a PhD in Computer Science from the University of Lugano, CH.


Data analytics Machine learning

7 habits to shorten the time-to-value in process mining
AI in medicine Data analytics

The smart drinking cup
AI in business Data analytics Machine learning

How can banks become truly AI-driven?