Developing machine learning algorithms
EMILI is working with the National Research Council of Canada, and the University of Winnipeg to develop image collections of pea plants and their root systems in order to apply machine learning to better differentiate pea plants from weeds and test the associations between Rhizobium inoculation and pea growth. Rapid phenotyping analyses from machine learning will allow pea breeders to select for increased resilience and higher yields faster than previously possible.
Machine learning for plant phenotyping
Plant phenotyping is essential to the development of new and resilient crop varieties. Selecting resilient genetics is the first line of defense to stabilize yield through changing environments. Our crop varieties must continue to grow with the needs of growers. Constant improvement is required to continually produce tolerant and beneficial traits while improving crop yield, harvestability, quality and value. The move towards high-throughput digital phenotyping through computer vision will help modernize crop development efforts for more efficient and effective selection of traits to benefit Canadian agriculture.
We are establishing a large image collection of field peas from emergence to maturity to capture above-ground plant biomass in the field as well as their root systems. Imagery and ground-truthing agronomic measurements will be used to train machine learning algorithms so we can move from traditional and manual trait-based measurements to more efficient automated phenotyping pipelines.
Traditional methods of phenotyping rely on laborious manual measurements subject to human error. Rapid phenotyping analyses allow breeders to stay ahead of yield-limiting factors and contribute to the success of Western Canadian farms. Our collaboration is working to develop an algorithm that utilizes machine learning to capture high throughput plant phenotyping data. We are collecting images of 12 field pea varieties on Innovation Farms at various growth stages through successive field seasons. Our aim is to use computer vision for estimating phenotypes such as biomass and yield and differentiating between pea growth and weed formation.
Together, we are developing methods to assess rhizobium root nodules and root system traits. This includes establishing metrics applicable for optimizing pea nitrogen fixation and crop yield. Root images from rhizoboxes at the National Research Council of Canada in Saskatoon will be utilized by University of Winnipeg and University of Manitoba researchers to develop an automated machine learning pipeline for measuring root and nodule development over time. The research aims to develop tools for the selection of beneficial below-ground traits for new pea variety development and ultimately, contribute to the goal of optimizing and reducing inputs for Canadian farmers.