As the human population increases and protein demand doubles, modern plant breeders must further optimize soybean plant architecture and per plant yield for modern farming systems.
Conventional techniques use imprecise visual scoring and laborious hand harvesting of single plants. Many important plant traits for modern farming systems are difficult to measure with current breeding tools, especially those related to complex physiological, structural, and environmental interactions.
Attempts to accurately measure these traits often require advanced technologies or highly labor-intensive methods. Because of these challenges to future food production, UTokyo associate professor Wei Guo intends to "open a new era of artificial intelligence (AI) driven plant phenomics for these valuable but hard to access traits."
With this purpose in mind, his lab teamed up with NARO soybean researcher Dr. Akito Kaga to design an image capture and AI analysis pipeline. Their technique enables much more precise and rapid measurement of single plant yield, plant architecture and seed localization with easily acquired in-field photographs or video. The study is published in the journal Plant Phenomics.
As Prof Wei Guo says, "Most efficiency enhancing AI agricultural applications require costly aerial or robotic platforms, while our much lower cost system could be used by soybean breeders with very modest financial resources."
UTokyo Ph.D. candidate Tang Li developed a novel image analysis pipeline that can automatically process and estimate the number and spatial distribution of soybean seeds on a plant in the field. The deep learning image analysis pipeline, called Multi Scale Attention Network (MSAnet) uses a multi-scale attention mechanism to help count seeds.
Li says "the most challenging aspect of designing MSANet was detecting only the foreground with minimal computation resources." After focusing attention on the foreground and making seed distribution heatmaps, various tasks are conducted on upsampled images, then the images are downsampled, matched with neighboring images and a loss function is applied to increase estimate confidence.
Finally, a kernel density algorithm is used to locate and count seeds, with more accurate results than any other existing pipeline. Then, easy to interpret graphs can be produced showing vertical seed distribution on individual plants that can be used by breeders to evaluate a variety of previously inaccessible traits on potential new varieties, or conduct genetic analysis on those novel traits.
Soybean breeders can use this new technique to directly select superior varieties for specific farming systems or for genetic analysis to identify the genetic regions of the soybean genome controlling vertical seed localization, plant architecture and height.
Click here to see more...