CropGym: AI drives smarter farming with harvest and soil quality predictions

AI expert Hilmy Baja at Wageningen University and Research (WUR) has developed an AI tool that can predict harvest and soil quality to help farmers sow and fertilize crops at the right time for optimum yields. The tool, CropGym, makes calculations for the future based on historical data and current measurements. It can help farmers achieve environmental goals and policymakers explore policy options.

Baja is building the AI model CropGym during his PhD research at the recently launched chair group Artificial Intelligence at WUR. The tool also aims to encourage farmers to use digital tools with AI in their decision making.

“CropGym is an interesting tool because it is adaptive and contextual, in the sense that the AI tool can adapt its recommendations based on changes in the weather, predicted weather, soil conditions, or crop conditions,” Baja tells Food Ingredients First.

“These properties come with the advantages of Reinforcement Learning (RL), which I use to train the AI model in CropGym. One just needs to supply the current (or historical) weather, crop, or soil data to it and it will tell you at that moment whether to fertilize or not. You can also plan ahead by giving it weather forecasts.”

Baja will present the tool at the Dies Natalis celebration at WUR on March 7, 2025, which will focus on the theme of “Artificial Intelligence for Sustainable Futures.”

Training the AI

WUR notes that RL is crucial to the functioning of the AI model, which is based on built-in “punishments and rewards.” The AI learns to determine the best strategy since Baja has programmed that using a lot of artificial fertilizer is “negative,” and a good harvest is “positive.”

The AI tool adapts recommendations based on weather, soil, and crop conditions, says Baja.The computer scientist says he has trained CropGym with RL, which is “inherently capable of adapting to droughts or floods,” given that it was trained on relevant data.

“For example, we can give inputs on historical drought data and train CropGym so it can learn what fertilization practices work for such conditions.”

He recently experimented with a function that lets the farmer know “when to take a measurement in the field,” according to WUR. The input helped him provide more reliable results.

However, according to Baja, the uncertainty never disappears completely. “The model is not a crystal ball, but a tool to compare options.”

Supporting policymaking 

Baja says policymakers can use the AI model to determine rules to achieve a goal, and it can “ultimately be a tool to explore possible policy options.”

“CropGym can discover fertilization practices based on the objectives that we set it to, or how we “reward” the agent. So one thing that can be done, for instance, is to set an objective for the CropGym AI to achieve an optimal Nitrogen Use Efficiency (NUE).”

After discovering a certain practice that can always achieve this optimal value, policymakers may create certain policies based on this discovered information from the AI model, he explains.

Assessing climate change scenarios

Baja tells us climate change scenarios are great for looking at “what-if” situations and assessing how they impact fertilization practices. 

“CropGym can discover the best fertilization practice for each scenario and can be used to plan ahead for other farm management decisions, such as how much fertilizer to buy in the coming year.”

CropGym can also be used to support policymakers in deciding rules for agricultural goals.He says AI is useful for looking 20 or 50 years ahead. “You can then immediately include the effect of climate change by testing different scenarios.”

“With more or less warming, the farmer can see how he can further optimize the harvest, achieve environmental goals and keep the soil healthy.”

Leveraging a “digital twin”

Baja is currently working on an EU-funded project called Smart Droplets to explore how drones and IoT can be integrated into the model. He plans to utilize a “digital twin” approach with autonomous retrofitted tractors equipped with IoT sensors for interoperable communication with CropGym. 

“The main idea is that when CropGym pushes out a fertilization recommendation, the retrofitted tractor will carry out a fertilization mission autonomously. Everything about the farm can be monitored within the server through the digital twin.”

Following a test in Spain, the project is now moving to Lithuania to “test the model in practice,” in April on field wheat, according to WUR. 

The date for fertilization, the optimal period for crop protection, and the required quantities of additives are yet to be determined. 

Overcoming limitations

Baja says CropGym can currently simulate the fertilization of crops such as rice, wheat, potatoes, and other C3 crops.

“However, it is not capable yet of fertilizing C4 crops (maize, sorghum, millet, etc.), since this is a limitation of the WOFOST crop growth model, which is the backbone of CropGym.”

“In this case, we require further expertise from plant scientists and agronomists to develop the WOFOST model further,” he concludes.