Soybean stress detection using UAVs and checkoff dollars

By Dusty Sonnenberg, CCA, Ohio Field Leader, a project of the Ohio Soybean Council and soybean check-off

Scouting soybean fields to detect stress factors impacting the crop is a never-ending process during the growing season. Stress caused by weather, disease, insects, and weeds are constantly challenging a soybean crop’s yield potential. The idea of using Artificial Intelligence (A.I.) to assess crop stress is being explored by researchers at The Ohio State University.

Scott Schearer, Professor and Chair of Food, Agricultural and Biological Engineering has lead the initiative, funded by the Ohio Soybean Council and soybean check-off dollars, to investigate the potential of using Unmanned Aerial Vehicles (UAVs), to aid in this task. UAVs, commonly referred to as ag drones, can be equipped with sensors to detect and map out the stress areas in fields, with the intent of being able to return to those areas to address the stress factor present.

“We started with the use of artificial intelligence to assess crop stress, namely weed pressure,” Schearer said. “We began thinking about if we could preserve glyphosate resistant soybean traits for multiple years if we could get the weed escapes under control. Often those escapes start as small areas, and the idea was to use the technology to attack those smaller areas early on in the life cycle of the resistance evolution and prevent them from developing.”

In this scouting process, multiple technologies are being employed.

“When we started, we were using an inexpensive microcomputer the size of a credit card with a technology called Raspberry Pi,” Schearer said. “This was inexpensive and provides all the capabilities of a laptop computer and was light weight enough to fly with. The technology evolved, and now a company, NVIDIA, is making graphical processing units (GPU’s) using the Raspberry Pi format that are are cost effective and can be paired with a cell phone type camera on a chip. When incorporated with the right intelligence, we can in essence collect weed images and process them on the drone and record the location of all the weed escapes.”

This research project has been ongoing.

“Part of what we have been doing the last couple of years has been developing that intelligence,” Schearer said. “We are collecting a lot of images and we are training convolutional neural network classifiers. Once trained, we can fly it on the drone and then generate those maps of the weed escape areas. That has led to the most recent round in which we are now attempting to act on those maps and treat those escape areas.”

Schearer feels that the main goal of the research project is being accomplished, that of creating technology able to scout, identify and map out weed escapes. The task now is to refine the technology and improve the results so that areas are not improperly identified or possibly missed.

“We are on track in terms of what we had envisioned to be successful,” Schearer said. “We still have to work on refinement of the algorithms. Sometimes we detect weed escapes that don’t exist, and sometimes we miss some. That goes back to the magnitude of the data sets. The more data and classified images we have, the better the neural network classifiers. The numbers of images we are working with is in the thousands. If we can get those numbers to the tens of thousands and hundreds of thousands, then that classifier begins doing a much better job. We are trying to make certain that we are presenting to that classifier all possible combinations of what we might see in terms of weed escapes. We are primarily focusing on marestail, water hemp and giant ragweed. We are attempting to create a library of images that is comprehensive enough to encompass all the exceptions that we might see in Ohio and in the future, expand that technology to the rest of the Midwest.”