Artificial intelligence bot works with expert human colleagues to improve Australian biodiversity data

Posted by AaronClausen

 5 Feb 2023

Australian Citizen Science platform NatureMapr has just turned on its first non-human, machine learning based moderator, known as Carbon AI (or “Carbon – Artificial Intelligence”). The Australian made and hosted technology is being rolled out right across Australia. The new technology will provide citizen scientists with lightning fast sighting identification while simultaneously increasing the quality of Australian biodiversity data.

Carbon AI is undergoing live testing and is being trained to identify common orchids, daisies, mammals and reptiles - at least to begin with. The non-human moderator works in unison with its expert human colleagues to participate in the identification workflow of uploaded records. Human moderators can disagree with its advice, which in turn trains the robot to become smarter over time.

Carbon AI is trained by the community and works for the community.

Category moderators are volunteer experts within the citizen science community that generously share their time and expertise within chosen categories of plant or animal species. Their backgrounds are diverse and include professional experts like retired CSIRO entomologist Kim Pullen and passionate former garbage collector, Stuart Harris, who fell in love with peacock spiders and has even had his own species named after him (See: Maratus harrisi).

“Moderators don’t have to have any formal background, they often start out as everyday citizen scientists who are passionate about nature and learn how to identify species over time” said Aaron Clausen, Founder NatureMapr. Category moderators are really the special ingredient that makes citizen science so successful, because they share so much time and invaluable expertise.

“They are amazing people who want to give back to their community and help others learn more about the plants and animals in their region. Moderators enjoy the challenge of trying to identify the thousands of sightings that come in, but it can also lead to a lot of work so we are always conscious of how precious their time is and how lucky we are to have access to them.” Mr Clausen said.

“The NatureMapr platform provides a flexible workflow that allows moderators to contribute as much or as little as they like. We are trying to better support these incredible people that make NatureMapr possible and do more to safeguard their time. Getting a machine to automatically identify the common species like magpies or eastern grey kangaroos, is one of the easiest ways we can free up our experts to focus on the tricky ones that a machine would struggle with - we don’t want to burn them out.” Mr Clausen said.

In March 2014, expert photographer and Field Guide to the Orchids of the ACT co-author Tony Wood joined Canberra Nature Map and began sharing his immense knowledge with the community through the identification of uploaded native orchid sightings.

It was controversial in 2014 because the locations of these sensitive plants had not been uploaded and managed successfully online previously – overlaying expert moderated safeguards to anybody’s plant and animal sightings was one key reason NatureMapr was created. Tony quickly gained the respect of the growing citizen science community and became NatureMapr’s first formal category moderator. Over almost 5 years, Tony generously shared his expertise to identify over 3200 native orchid sightings.

“One of the first categories we’re live testing is native orchids. Tony Wood, tragically lost his health battle in 2018 and passed away, leaving a huge hole in the community and in our knowledge of native orchids. But we’ve been able to load the expertise from Tony’s 3243 orchid identifications into Carbon AI’s machine learning model.

It means that the results of Tony’s effort and expertise can be taken forward to help future generations, forever. Knowing that kind of special knowledge isn’t lost is very powerful and is something that can be quite emotional to talk about.” Mr Clausen said.

NatureMapr encourages anybody with an interest in plants and/or animals to get in touch about learning to become a moderator. Anyone can learn to do it and it is one of the best ways to learn more about Australian biodiversity while helping to contribute to the real world outcomes that the data is used for.

Contact

Aaron Clausen, Founder NatureMapr
Mob: 0413 125 719
https://naturemapr.org

38 comments

   5 Feb 2023
Great move forward Aaron
   6 Feb 2023
Fantastic enhancement Aaron and very exciting to see it now working. I also love how it builds on the significant contribution of Tony Wood - what a great legacy! For those interested, take a look at some of the recent butterfly sightings to see the accurate suggestions already provided by @CarbonAI
AaronClausen wrote:
   6 Feb 2023
Thanks @MichaelMulvaney and @MatthewFrawley. Keen to get your expert eyes over the bot to help us train it and bed it in!
   9 Feb 2023
This is a very impressive improvement to our work place. Aaron, did you see my comment re ML on this sighting:- Harpobittacus australis (Hangingfly)
Mavis wrote:
   9 Feb 2023
Absolutely brilliant @AaronClausen
AaronClausen wrote:
   13 Feb 2023
Thanks @MatthewFrawley @michaelb @Mavis

@CarbonAI will now only make a suggestion if it is at least 50% confident in probability. It will scan ALL images within any given sighting and then go with the highest probability match found across any of the images within the sighting. It will only make one suggestion, being the suggestion with the highest probability of being correct.

@CarbonAI now runs on its own with zero intervention from the developers. Our friendly NatureMapr administrators can determine which categories he is best suited for and enable/disable him accordingly.
KylieWaldon wrote:
   14 Feb 2023
amazing. will be interesting to see which category/ies its most accurate in!
   14 Feb 2023
Good work Aaron, CarbonAI is progressing well.
AaronClausen wrote:
   17 Feb 2023
OK I've just turned on other arthropods and fish. They will start getting trained after midnight tonight. Maybe we can auto identify a few spiders.
MattM wrote:
   17 Feb 2023
Just curious what happens if the bot is trained on sightings that are later updated to another species? There are a few orchid species I need to review as there might be a few missidentifications.
AaronClausen wrote:
   17 Feb 2023
Very good question @MattM. Behind the scenes, when a sighting "changes species" and that sighting has already been used to train it's former/previous species, we will remove the image from the model and then it will be picked up and "re trained" under it's new species the next time the training runs again (currently around midnight every day).
KorinneM wrote:
   21 Feb 2023
A very cool thing you've created! Love it. A question - what happens with things like areas other than Canberra where there are no moderators for a category? Do the photos just sit as unidentified until a moderator can be found?
jks wrote:
   21 Feb 2023
@KorinneM great question and yes. We hope to grow so there are more community moderators across Australia. We are so lucky in Canberra to have a rich base of scientists and naturalists with combined knowledge surmountable to the distance to the moon or more. Other areas are not as dense in support so we encourage as many citizen scientists that back their ability to ID photo based records of a particular category, even out of their local area, to put their hand up. Our mate Carbon AI also helps out where they can annd gets better every day. Our dedicated human moderators are one of the most important cogs in what makes Nature Mapr data so valuable to science and conservation efforts. Also, It’s no pressure volunteering, you either are sure of an ID or you leave it. If you are busy… no worries!
AaronClausen wrote:
   22 Feb 2023
Very well said @jksmits

@KorinneM another change we're implementing soon is that moderators will be able to moderate across all regions much more easily. And the main naturemapr.org website is becoming much more of a single point of entry / one stop shop.

More to come!
KorinneM wrote:
   22 Feb 2023
Thank you @jksmits and @AaronClausen for the answers. Exciting developments!! Such a great site.
AaronClausen wrote:
   23 Feb 2023
This evening, we've granted @CarbonAI limited ability to confirm species identifications automatically, only in situations when it is at least 99.0% certain.

So far during testing, suggestions made with above 99.0% probability have been highly accurate. We will of course continue to monitor the bot's performance and accuracy.

Moderators please remember, if you don't agree with a suggestion or confirmation made by @CarbonAI, please make an alternative confirmation over the top of the bot's suggestion. This will reinforce and/or correct the model to continue learning from our expert human moderators.
RodDeb wrote:
   23 Feb 2023
@AaronClausen and @CarbonAI wow didn't know CarbonAI could do that how clever is that replying as well. Have meant to ask I can work out where AI comes from but where does Carbon come from?
Well done all CarbonAI is doing so well and so quick as well.
as per sighting 4471758
AaronClausen wrote:
   24 Feb 2023
Thanks @RodDeb - the bot is definitely progressing well with thanks to the guidance and continued efforts from our expert moderators.
AlisonMilton wrote:
   24 Feb 2023
Hi Aaron, CarbonAI is doing well, but needs a lot more training on spiders - :-)
AlisonMilton wrote:
   24 Feb 2023
Deb, Carbon comes from the fact that humans are known as carbon-based life forms (about 12% carbon). This is obviously a play on this fact.
RodDeb wrote:
   25 Feb 2023
Thanks Alison that makes sense that the carbon-based life forms are working with Artificial Intelligence and working well together. Appreciate your clarification.
AaronClausen wrote:
   25 Feb 2023
That's basically it in a nutshell @AlisonMilton @RodDeb - and you can thank @Gaelian for the name! He named it about 3 years ago when we were too busy to actually build it. Finally the day has come!
RodDeb wrote:
   25 Feb 2023
@AaronClausen @AlisonMilton @Gaelian thank you very much.
AaronClausen wrote:
   26 Feb 2023
Hi Moderators - small change tonight. If @CarbonAI makes a confirmation that you disagree with, please click "Undo" directly on his/her confirmation. This will downgrade the carbon AI confirmation back to merely a suggestion. And the bot will learn from it. This means that you can override and disagree with confirmations made by the AI bot now even if you don't necessarily know the correct species to identify said sighting with.
AaronClausen wrote:
   26 Feb 2023
   27 Feb 2023
Thanks @AaronClausen that's a good change. I'm curious if anyone has seen an incorrect @CarbonAI confirmation yet? I haven't thus far.
AaronClausen wrote:
   27 Feb 2023
Hey @MatthewFrawley this is the only one I've seen. This was a confirmation, now downgraded to a suggestion.

Aprasia parapulchella
JimL wrote:
   28 Feb 2023
I must get and read the Commentary more often! Explains the lot well.
I did wonder who this @CarbonAI was as they just seemed to be repeating my Suggestion.
Great work for those behind this! If it works as intended, this can be a potential time saver to support all of NatureMapper volunteer moderators and more!
KylieWaldon wrote:
   28 Feb 2023
I've only spotted three incorrect CarbonAI so far. Fuscous Honeyeater instead of White Plumed Honeyeater, White Naped Honeyeater instead of Blue Faced Honeyeater and Superb Parrot instead of immature Yellow Rosella. Most of these were not good photos.
AaronClausen wrote:
   28 Feb 2023
Good feedback @KylieWaldon - let's keep an eye on him/her!
AaronClausen wrote:
   28 Feb 2023
@JimL well said - yep that's our main goal with @CarbonAI - protect the precious time of our generous and amazing volunteer moderators.
DonFletcher wrote:
   12 Mar 2023
Like JimL, having read all these comments, I realise I should read the commentary more often, instead of allowing weeks and weeks to go by while I just keep responding to the automatic email notifications I get.

@AaronClausen I corrected one CarbonAI confirmation and one suggestion before reading this. I recall one was a Varanus varius that Carby had identified as V. rosenbergi and I think the other was probably some kind of small mammal. But meanwhile there have been vast numbers of correct IDs of mammals. I'm astonished how good it is. It is a fantastic development and you deserve a big fat medal for doing it! Hearty congratulations!

I have a little stash of images that I use to help people learn to ID local macropods - mainly monochrome (night) pics from camera traps so I will find more images like that for which the location is stored, then upload them to help CarbonAI get better.
AaronClausen wrote:
   12 Mar 2023
Excellent idea re: training CarbonAI @DonFletcher - exactly. Throwing as many images at it as possible with all kinds of varied lighting and angles will help really make the robot a lot more robust.

Maybe we need some kind of "announcement re-cap" email to be sent out once weekly or monthly or something that people can also unsubscribe from if it gets too much.

Thanks a lot for the awesome feedback. The team have worked super hard on this one so we genuinely appreciate it. That said, @CarbonAI is definitely still very much a work in progress and much more fine tuning to be done yet...
NeilCrisp wrote:
   12 Mar 2023
What a great citizen science initiative! Useful for people wanting to identify something they have found, and adding to an overall species map that is helpful for research. Also, training an AI with a very useful skill. Wins all round on so many levels!
AaronClausen wrote:
   12 Mar 2023
Hi @NeilCrisp thank you for the positive feedback and a warm welcome as well. It is great to get feedback like this, we have a lot more work to do but the team are doing amazing things... Are you in Brisbane?
AaronClausen wrote:
   25 Mar 2023
Hi everybody, we have made an addition to our privacy and terms of agreement page to cover the use of AI and machine learning within the platform:

"Training of AI / machine learning model(s)
NatureMapr utilises machine learning ("AI") for the purposes of automated identification of upload records. Images and/or meta data that you upload to NatureMapr may be used by NatureMapr as input to help train NatureMapr's machine learning model(s).

You can rest assured that NatureMapr will not provide access to its machine learning models to third parties or external organisations."

Or you can read the full page here: https://canberra.naturemapr.org/content/privacy
AaronClausen wrote:
   30 Mar 2023
Hi All,

A small update to CarbonAI's rules yesterday.

The community found that occasionally, CarbonAI would make a confirmation (which means above 99% probability/certainty), however it was still getting it wrong. We narrowed this down to the fact that within particular categories, an insufficient number of species were trained, leaving very little for CarbonAI to use for comparison purposes.

So we've added in another safeguard before CarbonAI is permitted to make a confirmation:
- At least 50% of all species within any given category must be successfully trained before CarbonAI is permitted to make a confirmation within that category

If you're not aware, we maintain the list of rules that the non-human moderator bot uses here:
CarbonAI
A.mitchell wrote:
   28 Apr 2023
@AaronClausen Hi Aaron, I am working on the WeedScan project which is also AI based. I am based at Black Mountain, CSIRO in Canberra. I would be interested in catching up and sharing notes.

Please Login or Register to comment.

2,204,154 sightings of 20,928 species in 9,221 locations from 12,761 contributors
CCA 3.0 | privacy
We acknowledge the Traditional Owners of this land and acknowledge their continuing connection to their culture. We pay our respects to their Elders past and present.