Thank you - this explanation helped me a lot.
1 Like
I’ve had this thought as well. If the CV could be accurately trained to recognize cultivated organisms, it might be the only way to really solve this problem. However, there are a few challenges here. Depending on the evidence uploaded, there can be little or no indication whether a plant is cultivated. Or, that evidence might be present in photos other than the first, which as I understand it, the CV ignores.
Also, we’d need to train iNaturalist users to discern which observations are cultivated vs wild, which, in my experience, they are not very accurate at correctly labeling. This is understandable, since evaluating cultivation status isn’t as interesting and people haven’t thought as deeply about it as they have about IDs. But if we want to use user labels as training data, this needs to be considered.
Overall, it might be hard for the CV to be accurate at this because it often involves making a judgment call involving the geographic setting, species ID (and its behavior in the local area), user habits, immediate placement in the landscape, etc. Not all of which are visual elements in a single photo.
1 Like
I agree. They could just add an announcement to the top of the users dash.
1 Like
As someone who focuses mostly on mammals, this is also very prevalent with people just uploading photos of their pets. While there is an argument for those that have escaped and established themselves within an ecosystem, some are very clearly pets.
1 Like
Yeah, this is a bit annoying. I can understand using a pet to kind of “test upload”, but folks usually don’t delete the observation after that.
1 Like
Very sorry for taking such a long break from this thread and coming back after people got bored. I’ve been thinking of different ways to remind people who use iNaturalist on what is and is not cultivated in hopes that it causes a fraction of improvement. Others here said it’s a losing battle if these reminders are not given and that this is just an acceptable loss. I don’t want to drop this topic.
Certainly there’s a balance here, right? Many researchers like me who would love to use as much wild-only observations to compile an algorithm or machine learning model that requires good data, however we don’t have the time or money to sift through and ask every user to clarify whether or not a questionable observation is of a wild or not-wild organism.
Our Options
- train observers better: get more ambassadors to recruit volunteer identifiers to look at the data quality of others’ observations and hope the “quality observation” to “total observation” ratio improves.
- incentivize identifiers to continue: this already being done through the improvements to seasoned identifiers’ UI and compensation for certain identifiers through their workplaces, but some other things that could help avoid identifier burnout may persuade them to continue identifying even though they may still have to tell observers to reconsider the captive/cultivated status of their observations. I don’t know how to make a person want to continue identifying other than money, a better UI, and a welcoming and supportive atmosphere, but there have to be more ways.
- make the reminder system: suggest, program, and just hope a pull request for a regular reminder system gets approved (not likely, and there’s no consensus on whether it will be in the site or in an email).
- automate captive/cultivated suggestion: suggest, program, and implement an expansion of the CV model that classifies whether or not an organism is captive or cultivated based on surrounding evidence. This is going to be a big ask, but let’s consider it. It’s not easy due to the nature of machine learning, but if the CV uses a multiclass model that predicts captive/cultivated status for different reasons (e.g. the classes are “wild,” “potted,” “planted in a row,” “indoors,” “collared,” “caged,” “near building,” etc.) the automatic suggestion would require the user to either accept applying the captive/cultivated status or dismiss the notice. It may or may not be integrated with the geomodel and might interfere with other existing models. The model will obviously never override the observer’s decision; the observer chooses to accept or dismiss the suggestion.
- Example 1: if a dog-shaped thing is detected by the model and it looks like a collar-shaped thing is around it, the model is likely to suggest to the user to mark the observation as captive/cultivated with the reason that it looks like the “dog” is “collared.”
- Example 2: if the model detects a green, leaf-shaped thing higher than a rectangular or cylindrical clay-colored thing in a well-lit area, the model is likely to suggest to the user to mark the observation as captive/cultivated with the reason that it looks like the “plant” is “potted.”
- Example 3: if anything evidence is taken in a background that looks well lit and has detectable patterning like tile, marble, floorboards, or wallpaper (easy-ish for high-quality CV models), the model is likely to suggest to the user to mark the observation as captive/cultivated with the reason that it was photographed “indoors.”
- find another more feasible idea
- give up
I am perfectly happy splitting these up into one or more feature requests if others suggest it is a good idea.
If iNaturalist Has a Reminder System
However, if iNaturalist implements a repeated reminder to users about what is and what is not captive/cultivated, the notification shouldn’t be overbearing or repetitive, so it should only be shown a few times for each user.
If Reminder System Is on Site/App
To make it most effective, it will have to:
- occupy the entire screen or otherwise remain at the center of the screen in the form of a div box or equivalent
- require the user to read the whole thing, and implement an “I agree and accept this interpretation of captive/cultivated” similarly to how some mobile websites force you to scroll through the entire terms of service before you can click on the “I agree and accept the Terms” button
- link to a more detailed page, possibly something on iNatHelp, discussing the concept in further detail using more verbose terms to encompass as many edge cases so people don’t have to keep asking on the forums and possibly merge the automatic non-wild deliberation explainer into it
- be in the language set by the user
- use stronger language explaining that it is a guideline, insinuating that it not a suggestion but not a rule either
We can and should debate these requirements. Nobody wants an annoying reminder system on the website or app to also fail to do its job.
If Reminder System Is Implemented in Email
If implemented here instead, it will need:
- similar basic requirements to the alternative
- explanations, not very verbose, of what terms like “wild,” “stray,” “feral,” “waif,” and others mean
- to explain the importance of marking things as captive/cultivated; if it’s subjectively important to people, they would be doing it more often
- reminder to archive or bookmark the message to re-read on occasion
A reminder system would be a terrible idea if it were implemented for everyone currently using iNat, and usually a person only needs to be told once or twice, quite sternly, to remember to do something.
Otherwise, I’ve given up on the idea that iNaturalist is also for data quality-sensitive research…where the people who benefit from iNaturalist by using its data for models and other aggregations will never be as prioritized as the average citizen scientist. That’s disheartening for conservation science where every finding of [reader’s favorite species with declining population] needs to be identified when many identifiers work on hundreds of species at a time and are swamped with fixing observations with other DQA issues.
iNat is a great place for tracking personal interaction with nature (which is fine, and I know I can’t just “change the mission” of an established platform), and so far that just means a wide network of people who like to show off which organisms they recorded, an extra source of data to funnel to GBIF, plus a few hundred or thousand publications using iNat data. If only this one data quality measure we’re so concerned about was monitored better because a lot of observations have to be discarded for research. I think a reminder system is fine if experienced users don’t get reminders.
Again, very sorry for the late response. I wanted time to think.
4 Likes