Food for the Future: How Artificial Intelligence Can Improve Drought Resistance

By Kevin Xie, 2018-2021 FFAR Fellow

The surface of a corn leaf. The stomata cells are light green and sink deeper into the leaf surface.

The surface of a corn leaf. The stomata cells are light green and sink deeper into the leaf surface.

We breathe in oxygen and breathe out carbon dioxide (CO2); plants do the opposite. But how, exactly, do plants “breathe”? My interest in plants traces way back to when I was in grade school. I was given some ugly seeds to sow in pots on the balcony of my home and was amazed when spectacular flowers grew over the following months. Since then, I have enjoyed growing plants as a hobby and later further redirected my research interest into crops to help breeding for the future warmer and drier environment. Understanding how plants “breathe” is a key step in knowing how efficient they can produce.

The “mouths” of plants are called stomata. These cells on leaf surfaces form pores that allow CO2 to enter the inner leaf space where photosynthesis occurs. However, it’s not a situation of “the more mouths, the better”. A critical trade-off exists. As CO2 is flowing into the leaf, water vapor inevitably flows out. The loss of water vapor from leaves accounts for more than 95 percent of water uptake by plants. If the water supply runs out, the stomata must close to prevent the plant from drying out. This prevents photosynthesis from making sugars that fuel the plant, which can lead to crop failure.

Part of my research as a graduate student at the University of Illinois and as a 2018-2021 FFAR Fellow, is to decrease stomata conductance through plant breeding to create crops that lose less water and are more drought-tolerant. My work is focused on this question: Can we find a good balance point where plants capture the most CO2 possible while using the least possible amount of water? Farmers are most eager to know the answers to this question so that they could save investment in irrigation and have more stable harvest when a dry year occurs.

To answer this question, we as researchers must accurately measure the number and size of stomatal pores. For a long time, measuring these traits was complicated, tedious and time-consuming. It involved painting the leaf with nail polish and peeling off an imprint of the cells in the nail polish, followed by endless manual counting and measuring under the microscope. This strategy was inefficient and made it complicated for researchers to accurately collect data.

This is where machine learning can offer a more efficient solution. In my research, I labeled the stomata and fed the coordinates into the computer along with this image. The computer extracted the features of the images, enabling the computer to recognize what stomata features look like. This process was repeated and improved over various cycles.

After letting the program run for about one day, I finally had a mathematical model that could be used to identify and count stomata in new images. The summary of the outputs gave me the location coordinates of each stomatal pore and its size. With that in hand, my team was able to scale-up and easily run the model on thousands of images. The real-world implications for this new process are profound and will significantly impact food and agriculture research. For example, I am now using this tool to identify regions of DNA in the corn genome where genetics drives variation in the number of stomata and water use efficiency in different varieties of the crop. This is a necessary step towards downstream gene function evaluation and integration to existing elite germplasms, allowing them to gain better performance in drought-tolerance and water use efficiency.

The backbone of the algorithm is Mask R-CNN, which is one of the latest computer science programs designed for object detection. This well-annotated framework allowed me, a crop sciences graduate student with no background in coding, to implement the algorithm, build my own model and answer my scientific questions. For anyone who is interested in similar topics (do you want to track down your pet dog from a home camera?), learn some basic Python and Linux and give it a try yourself!

None of this would have happened without the amazing support from both FFAR and Bayer. Being a part of the first cohort of FFAR Fellows has been an incredibly exciting and rewarding experience.