Using image segmentation to assess the quality of seed collections for conservation

Fig. 1. Seeds are x-rayed when entering the seed bank to see what proportion are infested, full or empty. These three images illustrate the variability in size and shape of seeds.

Seed collections can be variable in quality. Many collected seeds can be empty, partially developed or even infested with insects on arrival at the seed bank (Fig. 1). To assess the quality of the collections before they are stored long term in the seed bank, X-rays can be used to view the inside of the seed without damaging it.

We developed an automated method of counting and labelling seeds based on two commonly used deep learning image segmentation algorithms: YOLOv8 and Mask R-CNN. We also developed a web application that allows X-ray images, individually or in batches, to be uploaded (Fig. 2). The segmentation algorithms then separate the different parts of the seed: seed coat, endosperm, infestation and empty interior. This allows the algorithm to calculate the percentage fullness of each seed, and, based on a userdefined threshold, classify seeds as full, part-full or empty. If a seed contains any sign of infestation, it is classified as infested, even if some of the endosperm remains. The user can then verify all classifications and amend them if desired. Finally, the user can export the classification as a csv, and as a classified vector image. 
​

Fig. 2. Seed x-ray app user interface. Seeds can be uploaded on the left, and a model choice used (YOLOv8 or MaskRCNN). The right panel is used to visualise and annotate outputs.

​Overall, we observed that the YOLOv8 algorithm had faster inference times than the Mask R-CNN one and performed better on images without infestations and synthetic images. However, the Mask R-CNN algorithm performed better on images with infestations and on images with overlapping seeds, despite slightly slower inference times. We are now working to make the tool more accessible.

K.L. Dhanjal-Adams (RGB Kew), presenting work from a student project by G. Duffy, V. Gyasi, A. Holmes, A. Morton, C. Roberts, O. Rosen & I. Stoyanov (Imperial College London; MSc in Artificial Intelligence and co-supervised by R.A. Craven) publised in Samara issue 40 

Acknowledgements
Special thanks to the Kew Madagascar Conservation Centre for their hard work collecting the seeds used in the analysis and for kindly sharing data for app development. Special thanks also go to Alice Hudson, Tim Pearce, Elinor Breman, Roberta Dayrell, Pablo Gómez Barreiro, Sharon Balding and the entire seed curation team for feedback, discussion and ideas, as well as your many hours X-raying seeds.