Prof. Dr.-Ing. Jörg Dörr holds the chair of “Digital Farming” at TU Kaiserslautern. In this position, he is responsible for the research programs of the institute and has been leading the Smart Farming program at the institute since 2020. His work at the TU as well as at Fraunhofer IESE focuses on software and systems engineering, especially for applications in digital farming. Jörg Dörr has extensive knowledge in the areas of software and systems engineering, requirements engineering, and data usage control. He is the author of more than 150 academic and industry-related publications.
The list of applications of AI in agriculture seems infinite. What are your favourites?
Jörg Dörr: This is a difficult question because there are so many! Moving big objects autonomously is great. Another example is image recognition, i.e. using artificial intelligence not just to increase production and give better efficiency and higher yields, but also to achieve the sustainability goals of the Farm to Fork strategy and the European Green Deal.
One particularly interesting area is smart spraying: using a camera to detect and differentiate weeds from actual crops, spraying on the spot exactly where the weed is.
AI can also make a big contribution in giving decision support to the farmer: where to apply what kind of amount of material to the field. For example, we can measure the biomass in a certain area of the field using hyperspectral cameras from satellites. If we see a big dark green area, that is in high biomass, and you will get a high yield. If we have a yellow or reddish colour, do we add more fertilizer there? My students will say, yes! But that is not necessarily the right decision, because you have to know if there is the potential for yield, or if it is an area with bad soil and you will not get yield no matter what, and the fertilizer you put there will just run through the soil.
With data about the soil, data from space or drones, data about the NDVI index and from hyperspectral cameras about current biomass, and maybe even about the different kinds of plants you can give a real benefit to the farmer. Farmers know their fields very well, they have 20-30 years of experience, but they cannot have the experience of thousands of farmers in the world. That is where AI has the potential to analyse all the data from all the experience of all the famers together, and it can recommend a particular course of action to try. If it succeeds, maybe it will improve yields even further.
A third example, which may be a slight twist, is using AI to harvest something else: the opinions of farmers. We can analyse a huge amount of data on social media of how people think about digital farming technology. This is a completely different way of using AI to harvest opinions of farmers all over the world and use that to improve products.
Is there a difference in the ease of implementation between object-recognition technologies and decision support technologies?
Jörg Dörr: There are certainly challenges in the implementation of both types of technologies. For object recognition, if you have the ideal conditions such as a shielded area with good weather, then there is no big problem. But for instance, fog or dust from the harvest can pose challenges for object recognition, and there is still work to do for it to be completely reliable.
On decision support, the problem is with the data needed for the AI. In this case, the problems are 1.) how to do quality assurance for the AI of these decision support systems, and 2.) how to build trust. People ask: do I trust this system? Why should the farmer trust the system if he or she doesn’t understand it? Normally, if I wanted to explain a system to you, I would tell you the rules. Yes, if the soil is bad quality and if there is not a high biomass in the field, it doesn’t make sense to put fertiliser on it. But if I say, “the AI decided so because of the data it had”, the farmer may not believe or trust it. What does he or she know if this kind of AI makes the right or wrong decision? If the farmer follows the advice, who takes the responsibility if there is a significantly lower yield? So more transparency and education is needed.
Quality assurance and feedback on decision support must have a long delay and be data-intensive?
Jörg Dörr: This brings me to my main point about challenges, which is how do you do the quality assurance of such an AI-based system?
Let’s say you have data for the last 5 years. Now you do something differently, and you get the same yield. The next question is, are all the parameters the same? No, this year we had a slightly different weather. Is the change in yield due to your changed parameter, or is it due to the slightly different weather? Or is it even due to slight differences in the soil properties because, for example, you changed the crop rotation cycle? It is difficult to tell.
Another problem is with ‘parcel experiments’, where you take the same field, and you do one approach here and a second approach there. In a 5-10 years’ timeframe the climate has already changed. So, we cannot compare the field to itself 5 years ago, but rather to one in a different region that had a similar climate at that time. This is really tricky, but you can then do a data comparison and learning from the data can be much more valid.
Furthermore, we don’t have an assessment framework for comparing two products, or even on what parameters to judge them. If you have two solutions, how should a farmer judge which one is better? Research is desperately needed on a general framework for the assessment of digital farming solutions, including for the assessment of sustainability, social, economic, ecological, technical, quality and interoperability criteria.
Are people less likely to adopt a technology if they don’t feel it will benefit them? How do you deal with this?
Jörg Dörr: Profitability is usually among the top one or two reasons for not buying or adopting digital farming technology. But I think it’s the wrong approach to only talk about the financial perspective: you have to set out a multi-dimensional perspective on digital components. The question is not only how much do you save on the monetary value of resources, but also how does it add to sustainability? Because sooner or later that will also be rewarded.
Of course, people calculate that if they use a system, they can apply less material and therefore they have to buy less fertiliser or herbicide. How much money will that save? But other questions should be: how much will it contribute towards sustainability goals? How much will it decrease the workforce? How much less wear of the machinery will there be because I use it less? I think several economic models are missing calculating all of that.
Are there differences in implementation of AI in the agricultural sector in Europe vs. other regions in the world?
Jörg Dörr: Definitely. We did a study in 2019 about scouting the agricultural autonomous machinery market and how autonomous systems will evolve. There was a big part on market segmentation: Western Europe vs. Eastern Europe, Asia vs. the USA and Australia. The differences lie in the farm sizes, the role of data sovereignty and privacy and the kinds of attitudes people have. Also, the legal systems are different: if we can run an autonomous system, we still have issues with liability, like who pays for what and who sues whom. The policymakers also make a big difference. All of these factors influence the use of AI in autonomous systems.
So, what makes AI so exciting in the field of agriculture? What makes it lend itself well to this field?
Jörg Dörr: What makes AI so exciting in this field is that it helps us in an environment that is so much context dependent. The problem with agriculture and farming is that no two farms are identical. The farm size is different, the farmer has a different background, the soil has different properties, etc. It is quite hard to express all that in an ’if-then’ logic like we did in the old days, when we engineered these systems by simple rules. If you have a lot of data, a lot of context-dependency, and if the rules of reality are not known – and this is very true in agriculture – then AI brings its true power.
The good news is that the technological solutions are there, and the data is there. But the technology is not widespread because people have serious questions, e.g. about data sovereignty. If all farmers would share all their data, then the research industry and governments would have enough material to implement it. But the question is how to convince farmers. The best it to give them control and transparency, and it needs to be encouraged by policymakers. It’s not that we can’t do it – it’s about convincing the industry and the farmers to use it.