What is the best strategy for a successful harvest with minimal fertiliser and pesticide use? An AI tool can assist farmers and policymakers in making informed choices for a sustainable future. Researcher Hilmy Baja will present this tool at the Dies Natalis on 7 March 2025.
In April, field trials with winter wheat will start in Lithuania. The exact fertilisation date is still unknown, as are the optimal dates for crop protection, along with the necessary quantities. These moments and amounts will be calculated by an advanced AI model (Artificial Intelligence). Following a test in Spain, for which results are not yet available, this Horizon EU project Smart Droplets is now moving to Lithuania to test the model in practice.
AI expert Hilmy Baja is developing an AI model called CropGym during his PhD research at the newly established Artificial Intelligence Group. "Farmers everywhere, of course, look at the weather to decide the best time for sowing or fertilising, but they hardly use digital tools. In the Netherlands, some tools are available, but they lack built-in AI," says Baja, who previously applied computer science to other technical challenges. In his current research at WUR, he found the combination of technology and social relevance.
During my Engineering studies, for example, I learned how to control a motor. That was interesting too, but I wanted to work on more socially relevant issues, such as sustainable farming
AI predicts yield and soil quality
The AI model makes future predictions based on historical data and current measurements. How do farmers' choices today influence future crop yields? Or soil quality? This requires weather and growth data. Using weather records from past decades, it is also possible to simulate future weather conditions. Baja explains, "A farmer can then see what happens with more or less fertiliser or pesticides. What impact does that have on yield and nutrient levels in the soil? The idea is that users can test different strategies."
A crucial feature of the AI model is its built-in 'penalties and rewards.' This is called Reinforcement Learning. The AI learns to determine the best strategy because Baja has programmed it to treat excessive fertiliser use as a negative reward and a good harvest as a positive reward. The model includes additional rewards and penalties, and long-term effects are also considered. "AI is useful for looking ahead 20 to 50 years. This allows farmers to immediately factor in the impact of climate change by testing different scenarios. With varying levels of warming, a farmer can explore how to optimise yields, achieve environmental goals, and maintain soil health."
Certainty for farmers and policymakers
"Farmers want to minimise risks and make smart investment decisions for the future," Baja continues. This means that the model must provide sufficient reliability for farmers to trust it. "Farmers will not simply adopt AI as a tool. That is why we keep improving the model to enhance its reliability. This is our challenge as researchers."
Recently, Baja experimented with an approach that notifies farmers when a field measurement is needed. Providing input at the right time has been shown to improve the model's performance, making its outcomes more reliable. However, Baja notes that uncertainty will never be entirely eliminated: "The model is not a crystal ball, but a tool to compare options."
As the final step in his PhD research, Baja aims to make the model explain the reasoning behind its recommendations. "A farmer will then understand why the model suggests fertilising at a specific moment, ensuring that the minimum required amount likely delivers the best result. A policymaker can use the model to set regulations to achieve specific goals. Ultimately, this type of AI can be a tool to explore possible policy choices."
