Back to search

FFL-JA-Forskningsmidlene for jordbruk og matindustri

Reliable and efficient high-throughput phenotyping to accelerate genetic gains in Norwegian plant breeding (virtual phenomics; vPheno)

Alternative title: Større avlingsframgang i norsk kornforedling gjennom pålitelig og high-throughput fenotyping (virtual phenomics; vPheno)

Awarded: NOK 6.4 mill.

This project is developing and testing new technologies in both image analysis and genomic prediction that will facilitate a more precise selection of new plant varieties. This will give the plant breeder access to more precise data on the growth and development of the plants. We have the last five field seasons (2017 to 2021) conducted field trials both at Vollebekk research station in Ås and at Staur research station in Stange. At both locations, a panel of about 300 spring wheat lines have been tested for agronomic traits and grain yield. Drone images have been taken repeatedly at about weekly intervals from emergence to maturity. Parameters from these images are being compared with manual measurements. In the last four seasons, a validation panel of 300 new spring wheat breeding lines were also tested at both locations. The use of machine learning-based variable selection methods has resulted in promising grain yield predictions from the multispectral drone images, and this work was published in 2021. The last two field seasons we have performed parallel testing of two different multispectral cameras at weekly intervals during the field season to compare their data quality with regards to yield prediction. These results show that a much cheaper and user-friendly drone with an integrated multispectral camera provides comparable data to the more advanced and expensive camera we used earlier. This we have now documented in a scientific publication that will contribute to making multispectral imaging more accessible to plant breeders with limited experience in drone flying. In 2021 we have worked to improve the trait estimation methodology for plant height and heading date from drone images, and in collaboration with Boston University tried out promising methods for handling shadow effects from clouds in the orthomosaics from drone images. In collaboration with CIMMYT in Mexico we have developed efficient and quite reliable statistical models for use in genomic selection that combines phenotypic data with marker data and multispectral information from drone images. Several models that incorporate correlated traits, genotype-by-environment interactions and weather data have been tested out and published. Testing of these models on field trial data from the project shows promising results of improving prediction accuracy by integrating multispectral vegetation data in the genomic prediction models. Recently, we have also compered these classical statistical models with models based on deep learning and artificial intelligence. Results so far show that these models require more data to achieve the same reliability and robustness as the simpler statistical models. We have in the project also conducted yield trials with 24 historical wheat cultivars to study the plant physiological basis of grain yield increase in Norwegian wheat breeding. At Vollebekk, the cultivars have been planted over six years (2016 to 2021) at two nitrogen fertilization levels corresponding to today?s common practice (150 kg N per ha) and half of that (75 kg N per ha). In addition to grain yield and agronomic traits, we have in some of these trials also investigated yield components and plant physiological traits. In addition, repeated drone imaging with multispectral camera and close-up imaging with the field robot was performed. These trials clearly document a steady progress in grain yield over the past five decades, with the number of grains per spike as a main driving factor for the yield increase. There is also tendency with the modern cultivars showing a longer grain-filling period. A manuscript on this was submitted for publication in 2021. In the project we have developed new ways to visualize data from field trials in "virtual reality". A demo-version has already been completed by Making View that combine the three-dimensional models from drone images with multispectral data and close-up images from the robot and other data that are being gathered in the field trials. This is a tool that enables the plant breeder to visit the field trials virtually after the end of the field season and select what type of images and information to be displayed. In order to handle the massive amounts of data produced in the project, we have initiated the development of a structured database for storage and processing of the phenotypic data in MySQL. Tools for easy and efficient data sharing, data visualisation and pre-processing of the data are currently being implemented through a web application working with the aforementioned database. We are also making further progress in developing deep learning models to automatically detect and count wheat heads based on close-up images from the field robot.

-

New approaches are necessary to meet the goals of increased food production. Plant breeding can play a key role by developing cultivars with higher yield potential. New technologies like genomic selection and high-throughput phenotyping offer possibilities to increase genetic gains through more precise selection and shortening of the breeding cycle. However, considerable research is needed in terms of theoretical developments, statistical modeling and technical solutions to achieve this in practice. By bringing in world-leading expertise in statistical modeling and image analysis, we will develop novel statistical models to extract biologically relevant information from hyperspectral images. The work will consist of developing reliable methods for capturing high-resolution images of field plots, and utilizing novel computational solutions to integrate top view images from drones with close-up images from robots to build 3D models that retain the original resolution and hyperspectral information. Computational algorithms will then be used to extract important physical and physiological traits from these 3D models that can be used directly as selection tools in plant breeding. By coupling hyperspectral data with grain yield and other direct measurements, statistical prediction models will be developed that plant breeders can use in early-generation selection to increase yield gains. Considerable efforts will be spent on developing efficient computational solutions to manage the large amounts of data that will be generated, and finding intuitive ways of displaying relevant information to the plant breeder. User-friendly solutions will be developed through direct involvement of Graminor plant breeders in the project. By utilizing virtual reality technology, our ultimate goal is to "take the field to the breeder" and let the plant breeder observe the field plots and associated data through VR goggles.

Publications from Cristin

No publications found

Funding scheme:

FFL-JA-Forskningsmidlene for jordbruk og matindustri