Agriculture: Laying the foundations for hoeing
Research projects show how pesticide use could be reduced with drone images, artificial intelligence and hacking robots.
Flying drone and controller in front of a sorghum environment
(Image: Tobias Hase/StMELF)
If you believe the praise of innovation evangelists, politicians and, of course, providers, artificial intelligence is supposed to be useful for all sorts of things. But what does it look like in practice when AI applications meet reality and have to prove their usefulness?
Start-ups are springing up in the slipstream of ChatGPT, Gemini, Claude and LLaMA. The truth lies in the field, as a research project from Bavaria shows. Herbicides are still used in large quantities in agriculture, despite organic farming and stricter limits. Their use is to be reduced. In Bavaria, for example, the use of pesticides is to be halved by 2028. Can digitalization help? Approaches are being researched – and are very promising.
Today's agriculture has already switched to digital technology in many areas. From soil and weather data to seed optimization and crop yield measurement –, hardly anything works without computers in today's agricultural machinery zoo. The range of digital agricultural equipment extends from large harvesting monsters to small hoeing robots that drive across the field largely autonomously. And it is supposed to hoe what, from a farmer's point of view, doesn't belong there – technically known as weeds. But how is the machine supposed to know exactly what should not grow there?
Sorghum is a type of millet and an energy crop. According to data from the Food and Agriculture Organization of the United Nations (FAO), it is the fifth most important grain in the world in terms of acreage (after wheat, maize, rice and barley). Sorghum is becoming increasingly popular among vegans and vegetarians. Among German farmers, the plant is very popular as a renewable raw material for biogas plants, which requires less water than maize, can cope well with periods of drought and does not need much fertilizer. The disadvantage: sorghum grows relatively slowly at the beginning. This leads to other plants overgrowing it. How can farmers prevent this without using masses of herbicides or spending the day in the field with a hoe?
Videos by heise
AI until the hacker comes
In a multi-year joint project, scientists from Weihenstephan-Triesdorf University of Applied Sciences, the Technical University of Munich, the Technology and Promotion Center for Renewable Resources, and the Bavarian State Institute for Agriculture have researched weed control. They used the example of the sorghum plant. The ingredients: Flying drones, AI and robotics, three test fields and the sorghum millet. "Can we use drones to fly over fields and map weed herds? Can we then use intelligent robotics to remove them more efficiently?" says Dominik Grimm, describing the question. This is because the hoeing robots have not yet been able to clear fields hectare by hectare. "Go there and remove the weeds" - generating this information from drone images is the goal, explains Grimm, Professor of Bioinformatics at Weihenstephan-Triesdorf University of Applied Sciences and the Technical University of Munich.
The difficulties involved should not be underestimated: False positives lead to lower crop yields, as do false negatives if the wrong herb is then growing in the field. Over time, the plants change their appearance, sometimes dramatically, both the desirable and the undesirable ones. There is also another problem: hoeing too much could lead to soil erosion, especially on slopes, which is also undesirable. In other words, a colorful bouquet of challenges of various kinds.
No GPU clusters required
The very first difficulty, however, lies in the use of the drones themselves. For their research project, the researchers deliberately opted for commercially available hardware: small commercial drones were used, two DJI models in different price categories.
They also focused on existing machine learning algorithms such as UNet, DeepLabv3+ and FCN for the image recognition models used and tested, explains Dominik Grimm. Pre-trained feature datasets from general, non-agricultural image data were used for this purpose.
For this use case, it would be disproportionate and not useful to use non-pre-trained algorithms. The model selection allowed the researchers to use relatively narrow graphics cards from the consumer segment with 24 gigabytes of memory. It would even work on an RTX3060 if the parameters were scaled down a little, says bioinformatician Grimm: "We don't need an H100 super GPU cluster to produce good results."
