Exploring machine learning as a tool to improve the accuracy and speed of orchid viability testing

The Orchidaceae is one of the largest plant families in the world, but also particularly at risk from the impacts of habitat loss and climate change due to their dependence upon other life forms (e.g., fungi) at different stages of their life cycle, and therefore the response of any associated organisms to these threats (Reiter et al., 2016; Fay, 2018). Almost half of orchid species assessed by IUCN (955 spp.) fall within a threatened category. Habitat loss, climate change and their dependence on other life forms at different stages of their life cycle calls for urgent conservation.

Seed banking provides one option to help with their conservation effort. However, many orchid seeds still present a challenge for traditional seed banking procedures. For example:

• They can be tricky to handle – orchid seeds are very small, think vanilla seeds in your custard or ice cream, so a lot of work on them has to be done under a microscope
• Many orchid seeds are short-lived – this means they require quick processing from the point of seed collection through to banking, generally less than 2 weeks
• Orchid seeds often require a specific mycorrhizal fungal association in order to germinate, if you don’t have the correct mycorrhizal fungi they will not germinate. Even if you do have the correct mycorrhizal fungi, germination can in some cases still take many years.

Key to effective seed banking is to make sure that the seeds being stored in the bank are alive and remain alive and able to germinate throughout their time being stored in the seed bank. 

For most species at the seed bank, we do a germination test to tell if the seeds are alive. But because of the germination time, seed size and mycorrhizal fungi requirement for germination this can be very tricky. So instead, a different method of testing viability is often used for orchids – a staining test. This works on the principle that some chemicals (stains) will change colour when they come into contact with living tissue, in this case the embryo of the seed, so depending on if the seed embryo changes colour you can tell whether the seed is dead or alive. There are a number of different staining tests which can be used depending on the original colour of the orchid seeds e.g., tetrazolium chloride or fluorescein diacetate.

Although the staining tests to check if orchid seeds are alive or dead is easier than a germination test, it still has a few challenges such as the size of the orchid seeds means everything has to be done under a microscope and because the test relies on a colour change it can also be subject to different colour interpretations by different people.

Our project aims to use machine learning to automate the process of analysing tetrazolium chloride orchid seed viability testing. This will involve:

• Developing a training dataset of annotated tetrazolium chloride orchid viability tests
• Using machine learning to create a model isolate each individual seed from each image
• Using machine learning to create a model to classify each seed as alive, dead or empty
• Delivering the machine learning models through a web app for use by other seeds banks around the world

Project Leads

Alice Hudson
Pablo Gómez Barreiro

Project team 

Ania Pajdo
Nicola Mills
Adam Richard-Bollans
Eren Karabey
Kate Gill
Sarah Gattiker
Sian McCabe
David Hickmott
Sarah Adams
Lucy Taylor
Victoria Philpott
Frances Stanley
Athena Tang

Acknowledgements

Orchid seed collections used in this project were collected by: 
Silo National des Graines Forestieres (SNGF) Madagascar 
Ministry of Agriculture, Lands, Housing & Environment Montserrat 
Instituto de Investigação Agrária de Moçambique (IIAM), Mozambique 
Departmento de Recursos Naturales y Ambientales, Puerto Rico & the British Virgin Islands National Parks Trust

Bloomberg Philanthropies