Aarhus University Seal / Aarhus Universitets segl

Genetic analysis and image-based phenotyping of white clover

Publikation: Bog/antologi/afhandling/rapportPh.d.-afhandlingForskning

  • Marni Tausen
White clover (Trifolium repens L.) is a highly successful temperate species, commonly used in clover-grass pasture mixtures for grazing. White clover has a symbiotic relationship with nitrogen-fixing bacteria called Rhizobia. Therefore, white clover is commonly used in organic farming as a natural source of nitrogen in clover/grass mixtures.

White clover is a recent allotetraploid hybrid of relatives of Trifolium occidentale, a coastal species, and Trifolium pallecens, an alpine species. Hybridization has been estimated to have occurred between ~15.000 to 28.000 years ago during the last ice age. Circumstances of the hybridization event were looked into by comparing mutation accumulation simulations to 200 genotyped white clover plants and estimating the effective population size history using pairwise sequentially Markovian coalescent (PSMC) and multiple sequentially Markovian coalescent (MSMC) on four whole-genome sequenced clover individuals. PSMC and MSMC found no strong genetic bottlenecks in the history from the hybridization event, and the simulations found that a strong bottleneck did not fit well with the variation found in the clover population. A single hybridization event was unlikely, and progenitors of white clover seemed to be readily hybridizing.

Large-scale phenotyping experiments require the processing of large quantities of data, and extensive manual labor is required when using traditional techniques. The automatic capture of images for phenotyping experiments allows for growth rate estimates. We conducted a large-scale clover-rhizobia interaction experiment containing 3600 plants in two rounds. The experiment was captured by a camera system with 180 cameras, each capturing ten plants. The experiment yielded around 355.000 images over 146 experimental days. The automatic acquisition of multi-plant images requires the detection and identification of the individual plants and segmentation of the plants for measurements.

We implemented Greenotyper, a piece of standalone python software, to detect and analyze plants on multi-plant images. Object detection was done through the Tensorflow object detection API, using the faster-rcnn inceptionv2-resnet network. The object detection had a very high bounding box accuracy of 98.7% mAP and allowed for the successful detection of 3568 plants out of 3600. We trained a U-net using 40 ground truth masks for the segmentation of white clovers on the images. The U-net got an Intersection over Union IoU of 0.84 and average precision (AP) of 0.96, performing better than traditional methods such as thresholding. All of the images were processed in around 24-36 hours using a large computing cluster.
OriginalsprogEngelsk
UdgivelsesstedAarhus
ForlagAarhus University
Antal sider180
StatusUdgivet - 2020

Se relationer på Aarhus Universitet Citationsformater

ID: 193507829