Deep learning-based phenotyping of crop seed banks and herbaria

Biodiversity underpins long term sustainability of agriculture as well as ecosystems. Current commercial
crops have a extremely narrow genetic base, which is a vulnerability. Seed banks and Herbaria represent
a concentrated resource, collected over many decades and include many heritage varieties with potential
value in the face of climate change and other pressures. The challenge is to extract this information in a
systematic manner that useful to researchers, breeders and agriculture.
We hypothesize that these physical archival materials can be mined for variation in useful traits using
non- or minimally destructive approaches, that include image-based phenotyping, and this variation
related to the underlying genetics.
To test this, well documented grass and cereal collections are available in the AberInnovation Seed
Biobank and in the IBERS' Crop Herbarium. The specific objectives are to:
Obj1. Machine learning tools
To develop robust deep learning protocols and test, a set of manually labelled images from selected test
grasses will be used to train neural networks to recognise and measure features. We have recently
applied deep learning to identify and count organ from both vouchers and uCT scanning (Masters project,
Fig). These deep learning networks will be extended to quantify a wider variety of traits at different
levels of image resolution. For example, higher resolution imaging and deep learning will be applied to
microscopic features such as stomata and vessels that support water transport and gaseous exchange
processes and other minimally destructive modalities will be evaluated (X-Ray Fluorescence imaging for
elemental analyses, metabolomics for nutritive content, etc).
Obj2. De novo test-case experiments
To formally address whether variation in these traits can be linked to underlying genetic variation (and
therefore be useful to breeders), the DR will use vouchers made from genetically unstructured
populations that have been previously created by crossing selected pairs of oats and of ryegrass. These
morphologically diverse and genetically characterised populations provide a rigorous test-case for
validation of the approach. The student will build quantitative models to account for the genotypic
contribution to variation in features contributing to yield and quality, and compare to models produced
using conventional data.
Obj3. Phenotyping the grass and oat herbaria
A very extensive collection of grass species is populated by original wild samples (with geographic
coordinates) as well as duplicates grown locally under defined agronomy. These will be imaged and,
where appropriate, uCT scanned for seed and stem traits. The samples are crossed referenced to
breeding ledgers and in many cases their pedigree can be traced through to current commercial varieties
while seed can be recovered and regrown as necessary. This physical collection of >10000 samples
provides an unexploited resource to examine genetic and environmental effects across time. The student
will focus on the ryegrass and oats within the collection aiming to exploit advanced genetic analytical
techniques such as GWAS.
Justification and Likely Outcomes: This crop-focused project brings together the Phenomics Centre and
Seed Bank at Aberystwyth with Computer Sciences in Surrey who have previously worked with Kew
Gardens on biodiversity. This combination of expertise combined with large datasets from genetically
defined species will allow design of optimised approaches to extract and interpret the information
content of more extensive collections, available in Botanical Gardens from across the world, and
ultimately extending beyond the grasses to other crops and into wider questions associated with
biodiversity.