Innovating plant phenotyping using hyperspectral imaging integrated with machine learning

November 27, 2018
a diagram of how researchers took hyperspectral images of wheat lines and adding vegetation binary mask to it

Hyperspectral images were taken of wheat lines. A vegetation binary mask was created to separate vegetation pixels from background pixels in a two-step process.

Machine learning is reshaping the way we live. It’s used today in online sales recommendations, content curation, and even for detecting earthquakes, fraud, and plagiarism.

For University of Minnesota doctoral student Ali Moghimi, machine learning is an opportunity to innovate plant phenotyping methods. He and a team of researchers, including department professors Ce Yang and Peter Marchetto, developed a new method that applies machine learning with hyperspectral imaging to rank salt tolerance of various wheat lines.

Conventional methods to assess and select salt-tolerant crops rely on labor intensive and time-consuming methods, which often poorly correlated to salinity stress alone. The researchers’ proposed method saves time and labor, and it can quantify the salt tolerance of wheat lines within one day after salt treatment, even before symptoms can be seen.

“Breeders and plant geneticists would not need to wait about two weeks after treatment to measure aerial and root biomass for each individual plant,” Moghimi said. “All they need to do is to take images of the plant and feed them to the proposed pipeline, which will rank the lines based on their tolerance in a quantitative, interpretable, and non-invasive manner.”

According to their paper published in Frontiers in Plant Science, using hyperspectral imaging has drawn attention for plant phenotyping because it allows researchers to see the inside of plants such as tissue structure, pigments, and water content and investigate how they are affected by environmental variables. Studies on hyperspectral imaging has been done, however, there is limited research on how to handle, process, and analyze those hyperspectral images. This is where machine learning can help.

“Machine learning algorithms are a promising approach to analyze large datasets generated by sophisticated imaging sensors,” the paper stated.

The researchers grew four different varieties of wheat lines in a hydroponic system with control and salt treatments. A hydroponic system was used to ensure uniform conditions. Plus, it allowed researchers to add salt to the plants without affecting its surroundings.

Hyperspectral images of the wheat lines were taken about 24 hours after the salt application. A closer analysis of the images showed that the salt had already affected the plants despite the lack of visual symptoms. Images were analyzed using image processing and machine learning algorithms, including a new original algorithm referred to as the vector-wise similarity measurement to calculate the similarity of all pixels to the salt endmember.

Moghimi and his team also took it one step further. In a separate study, published in IEEE Access, they identified the most informative spectral features for salt tolerance assessment.

In hyperspectral imaging, hundreds of wavelengths are scanned by the sensor. However, a significant portion of these wavelengths are redundant or irrelevant, depending on the desired application.

“Among the hundreds of wavelengths captured, only a small set are relevant to what you’re looking for,” Moghimi said. “By narrowing them down to only those that are desired for a certain phenotype, it reduces the challenges and complexity of hyperspectral image analysis. Also, a multispectral camera can be designed for that particular application to leverage the advantages incorporated with multispectral imaging.”

The new method can be applied to other plant phenotypes as well, and Moghimi is motivated to continue building on this concept. With the help of other researchers, he envisions developing a digital library of spectral information related to plant diseases and stresses, which can be beneficial and accessible to breeders, growers, farmers, researchers, and others interested in plant health around the world.

“It’s an exciting research due to its multidisciplinary nature,” he said.

The study was funded by the USDA-Agricultural Research Service, the National Science Foundation, and the Minnesota Agricultural Experiment Station. Additional financial support was provided by the MnDRIVE Initiative.