Research Highlights

Processing Big Pictures: Training Artificial Intelligence for Agriculture

In this article, you’ll find details on:

• Ohio State University research is using pictures to create training datasets to apply artificial intelligence to specific agronomic decisions.
• The research is investigating how to gather and process data, both to create the training datasets and eventually to drive AI applications in the field.
• Training AI to mimic subject matter experts could dramatically reduce use of some inputs, saving costs while maintaining or improving yield.

By Laura Temple

If a picture is worth 1,000 words, then 60 million pictures must be worth 60 billion words. That’s an incredible amount of information. 

What if farmers could apply that volume of information to in-field decisions in real time? How might that improve yields and profitability? 

Scott Shearer, professor and chair of the Department of Food, Agricultural and Biological Engineering at Ohio State University, is tackling those questions. He leads research focused on turning information from images collected in the field into artificial intelligence (AI) training datasets that can help farmers save input costs and improve yields.

His team can collect 60 million such images in one day. 

“We are figuring out how to generate intelligence to power future intelligence,” he says.

That requires processing and interpreting the images in a way that effectively trains AI for the desired task. 

Shearer works closely with the Intelligent Cyberinfrastructure with Computational Learning in the Environment, or ICICLE, to do this. ICICLE is a multi-institutional, multi-disciplinary AI institute funded by the National Science Foundation and led by Ohio State University that aims to democratize AI, making it more accessible to everyone.

“Digital agriculture is a key focus area for ICICLE,” Shearer says. “We are working on how agriculture can tie into ICICLE’s work to make money for farmers.”

His process starts with figuring out what agronomic challenges would benefit from applied AI. The Ohio Soybean Council is investing soy checkoff dollars in allowing Shearer to develop use cases.

“For example, we can use AI to identify the difference between magnesium and sulfur deficiencies in soybeans,” he explains. “In-season fungicide applications are another example. Is fungicide the right thing at the right time? Can it be applied selectively in high-risk areas?”

He notes that the intelligence to answer these questions is separate from the hardware components that apply it.

“There are 24 companies around the world with hardware components aimed at improving input use, like technology to spray herbicides only on weeds,” he says. “The hardware needs the intelligence to make decisions, based on all the relevant factors.”

Taking Intelligent Pictures

For each use case, Shearer and his team determine where cameras need to be positioned to gather the right information. For example, the cameras need different perspectives to evaluate crop residue for tillage decisions than to examine pod set in soybeans. Cameras can see residue from above the field, but they need to be inside the crop canopy to evaluate pod set.

The team uses 18 GoPro cameras, often positioned on equipment or drones, to get them to the right places. Shearer’s team also determines the frequency that pictures need to be taken. Each pixel of data from these cameras represents 0.5 square millimeters, or one-thousandth of a square inch. Such high resolution allows the cameras to “see” in great detail.

These pictures gather the intelligence needed both to train AI, and then to instruct AI-enabled hardware to execute the give task. 

The Processing Pipeline 

The GoPro cameras connect wirelessly to a graphic processing unit tied to a file server housed in 24 hard drives. The very large raw photo files create more than 3 million gigabytes of data. 

“Part of creating an AI application is determining where collected raw data will be processed and relayed to the equipment,” Shearer explains. 

He says that the cloud, where most data from cell phones and personal computers is stored, is very slow at processing this volume of information, especially if connections are unreliable in the field. 

Another option is “the edge,” which refers to placing the processing capacity near the data collection point. With in-field applications, it could mean putting the graphic processing unit on the implement or tractor, so it can turn images into data the AI-trained hardware can act on.

Shearer’s use case research allows his team to explore options and figure out what will work best in the field.

AI Training

While outlining how AI can impact an application, Shearer and team are also creating training datasets. These datasets feed into the ICICLE platform to create intelligence.

“Once we decide how and how often to capture images, we can effectively gather large amounts of data quickly,” Shearer says. “The bottleneck comes from getting this data to computer scientists.” 

Shearer explains three types of AI learning.

• Supervised learning has humans point out and define what parts of each picture the computer should “learn.”
• Unsupervised learning allows the computer to teach itself based on what it gathers from the data provided.
• Semi-supervised learning uses human input to tell the computer what to look for in categories but then allows the computer to categorize data itself.

Research indicates that the average person processes as much as 74 gigabytes of data per day.¹ That’s just a small fraction of the volume of data captured by the GoPro cameras.

Thus, Shearer’s team uses semi-supervised learning, which he says is more robust and accurate because of human guidance, but requires less human input than supervised learning. In recent trials, Shearer and team collect pictures during the day and their hard drives to send the files to ICICLE overnight. 

His plan for 2025 research includes starting to gather training images from cooperating farmers’ fields, rather than just research farms. Ultimately, this research is creating ways to help farmers access huge amounts of agronomic information while in the field. 

“Our goal is to train AI to mimic a subject matter expert,” he says. “Accomplishing this could create a 60 to 70% reduction in some crop inputs.”

¹ How Much Information Does the Human Brain Learn Every Day? Medium, June 1, 2022 

Additional Resources:

Combining Human and Artificial Intelligence for Input Decisions – SRIN article

Intelligent Cyberinfrastructure with Computational Learning in the Environment (ICICLE) – website

How a National Image Repository Can Transform Agriculture – GROW article

Meet the Principal Investigator: Scott Shearer

Published: Apr 28, 2025