AI against bee mortality
Winners of the national AI contest announced: The main prize goes to tech enthusiast and hobby beekeeper Sebastian Steppuhn for his clever mite detection system.
(Image: Ale Zea / BWKI)
Varroa mites are considered to be one of the main causes of bee mortality. Sebastian Steppuhn, a student from Pforzheim, has built a clever device to detect and treat infested bees – and was awarded the main prize of 1500 euros at the German Federal Artificial Intelligence Competition (BWKI) as well as an internship at the robotics and automation company FANUC.
This year's BWKI, organized by the Tübingen AI Center for the sixth time, was held under the motto "What freedom does AI give you?". Last Friday, ten teams were able to travel to Tübingen and present their projects – to both the jury and the interested public. There was also a supporting program with an AI makerspace, a holobox, an interactive marble run and an alumni get-together, in other words a relaxed event with an event character and eager anticipation of the decisions.
Federal President Frank-Walter Steinmeier also attended the event and was impressed, and not just by the technical achievements: "The most admirable thing is the young people – Girls and boys – who dedicate themselves to their developments with great creativity, courage and imagination, and most of them do so in their free time alongside their normal schooling."
The mite in the beehive
It is only 1.1 millimetres long and 1.6 millimetres wide – but it is a major problem: the Varroa mite attaches itself to the thorax or back of adult honey bees and sucks on their fat bodies. In doing so, it can transmit viruses or other pathogens. Infested bees become weak and die. However, the parasites cause particularly devastating damage in the brood cells in which the young bees grow up. Without treatment, the infested colony dies within a few years. To combat the mite, the entire colony must be treated, usually with formic acid.
(Image: Ale Zea / BWKI)
Because this method also exposes many healthy bees to unnecessary treatment, hobby beekeeper Sebastian Steppuhn set to work and developed an intelligent sluice for his beehive that recognizes infested bees and automatically sorts them out for treatment. He won over the jury of this year's BWKI with his well thought-out system, which covers and solves the many facets of the problem from start to finish.
The start was anything but easy. Steppuhn had constructed a box with a narrow transparent passageway. He placed this in front of the entrance to his beehive so that he could systematically observe the animals using a camera mounted above it. His plan was to detect the parasites using a classic machine learning algorithm and send the infested bees directly to a "treatment room". A suitable machine learning model for such real-time applications was quickly found in YOLO – but there was no training material available for his special problem.
Training data: sowing and harvesting
In order to find and then classify a significant number of images of varroa mites on bee bodies in his own video material with the naked eye, Steppuhn would have had to invest many, many tedious hours. So he came up with a trick: He generated a small synthetic data collection by mounting mite images on bee images and producing various variations from them. He then used these examples to train the YOLO system to locate at least a few frames with real images of infested bees in his videos.
From these valuable realistic training samples, he was able to generate further variations, and so on, until he finally had a comprehensive data set covering all possible angles and appearances of the mite-bee pair. A cleverly chosen combination of Plexiglas and lighting for the passageway ensured low-interference and, in particular, low-reflection images. The fully trained mite detector proceeds in two stages: First it marks the back and abdomen, then it searches for the telltale patterns within this region of interest (ROI) – similar to what happens in real-time license plate recognition, by the way, where the search area is first narrowed down to the vehicle and then to the license plate.
(Image: Ale Zea / BWKI)
Recognizing infested bees is great, but without a suitable sorting mechanism it remains an academic exercise. The 17-year-old was also unable to fall back on tried and tested methods to solve this problem. There are already machines that sort out fungus-infested cereal grains on an industrial scale. However, such high-speed processes are not suitable for the sensitive insects; after all, the aim is to treat them, not kill them. So he quickly devised a compressed air mechanism that gently diverts the patients into a separate container. Many parts of his beehive sluice came from a 3D printer, and the tech beekeeper connected his home garden to high-speed internet.
Videos by heise
In Steppuhn's own hive with 25,000 bees, the system ultimately separated 18 specimens with mites, with around 80 misses. At first glance, this ratio does not seem particularly good, but it is a common phenomenon in mass screenings. In such scenarios, success is measured by a benefit-damage calculation, which in this case would be in favor of the system: not having to treat 24,900 healthy bees and having prevented the spread of the mite.
Clever traffic lights, courageous experiments
The winners of the other categories also scored points with systems that worked and were thought through to the end: Leonie Weiss from Regensburg was awarded the prize in the special category "AI for Good" for her intelligent TrafficAid traffic light control system. The system recognizes vehicles of all kinds and only switches to red when it is really necessary –, thus avoiding unnecessary waiting times and reducing CO₂ emissions. The 18-year-old also took safety and fairness aspects into account in the control system so that the respective waiting times do not diverge too much if there are a lot of vehicles coming from one direction and only a few from the other.
(Image: Ale Zea / BWKI)
The "KI aemazing" project by Anna Perkovic and Nicholas Dahlke from Lörrach in Baden-Württemberg was virtually predestined for the "No risk, no fun!" category. The two 17-year-olds had set themselves the goal of solving a still unsolved scientific mystery with the help of AI: Why does hot water freeze faster than cold water? What sounds quite harmless in the project description turned out to be an extremely ambitious experiment with a very elaborate experimental setup. On one side of a tube, water and oil were injected together to form 0.5 millimeter droplets coated in a film of oil. The test continued through a minus 40 degree cold cooling system, behind which a previously trained image recognition system classified each droplet: frozen or not. In fact, it turned out that significantly more droplets froze when hot water was used than when cold water was used. The young researchers certainly benefited from the equipment at the Phaenovum student research center in Lörrach, but also from their organizational skills: they borrowed various special devices from companies.
(Image: Ale Zea / BWKI)
The special "AI Research" prize was awarded to the work of Lorenz Rutkevic from Leer in Lower Saxony. His system optimizes and analyses microscopic images of cells. On the one hand, it should also enable less precise digital microscopes to deliver pin-sharp images. Secondly, the algorithms trained and adapted by Rutkevic support doctors in diagnosing diseases by searching for typical patterns and marking or segmenting them in the image. The audience award also went to Lower Saxony, namely to the "FolderCopter" by sixteen-year-old Peter Fuchs from Hanover: a language model that runs locally on the computer and answers questions about one's own collection of documents and knowledge.
Strict refrigerator
The intelligent refrigerator created by 15-year-old Fabian Then from Aidlingen was definitely very entertaining. The system used sample data to learn which foods need to be stored in which compartment so that they last as long as possible – and was not fooled by the Federal President: "The pineapple shouldn't go in the fridge," the fridge said when Steinmeier wanted to find out which climate zone of the appliance was best for the fruit.
heise Medien is a BWKI cooperation partner in 2024. c't editor Andrea Trinkwalder was part of the jury this year.
(atr)
