This will likely break out into many different issues, but wanted to start with an ideal end state, and then from there we can go back to what should be in the MVP.
The overall goal of this application is to be a key tool in a complete feedback loop from users making use of FTW data, seeing that the model performance is poor in the area they care about, spending time to improve / correct the data, and submitting it to be incorporate into the core FTW data, for testing and training.
For FTW benchmark data '2.0' we'll be doing everything on the ftw-grid. All the training data will be cut from there, and anything in that form will be compatible. So the core idea of this app will be to let users select a specific grid cell from ftw-grid and 'complete' it - delineate every field in it, so it can go directly into testing or training data.
The editing tools should be good enough that a user who wants to start from scratch can quickly edit an area, with all the capabilities that advanced GIS editing tools enable, but tailored for this specific use case. The biggest one is 'split' - be able to draw an outline and then just draw a line across the field boundaries to create individual fields. And generally snapping should work well. I'm not an editing expert, so feel free to add more here @ahocevar
The other thing the app should do is enable 'AI assistance'. The biggest one here is to run FTW models (in the same way as the inference app to start - we could explore some in browser stuff too) and then be able to 'start' with those results, correcting it to make it better. It could also be interesting to experiment with some lighter weight CV/ML tools, like segment anything, to help users work in the browser.
There should be some 'submission' process where the output goes into a queue for review, and if successful it goes into an 'ftw community dataset' area, and that can be immediately used by people making ftw models, and it will undergo further review for incorporation into the FTW dataset.
There is a whole set of functionality around managing users, reviews, scoring and submissions. I think we ideally have a pretty clear / clean architecture for the core 'editing' application, so it can easily be reused / remixed by others. But to do this well we will need the 'other' stuff - let users log in and see what they've done, assign tasks, prioritize reviews, have some queue to review what others have done. We should discuss what we can do for the MVP, and also how to architect things so the editing can be fairly standalone.
The one bit of functionality that would be nice to keep in scope is some basic 'review' thing. Like instead of the only path being to do a complete editing of a grid cell there should be the ability to just 'score' a given grid cell. This could be users who just score the results of an FTW model - they don't have the ability to correct, but they could flag that it is bad. Ideally we have some examples of like a scale of 1 to 10, to help people pick a right score. It could also be users reviewing editing submissions of others, to make sure it's high quality.
Long term we'd like to easily make other editing apps, for buildings, for crop classification, for land classification, etc. I'm not a fan of trying to build the most generic application from the start - I prefer to build for purpose and then extract from there. But it is likely worth knowing for any architecture choices that could make that harder in the future.
This will likely break out into many different issues, but wanted to start with an ideal end state, and then from there we can go back to what should be in the MVP.
The overall goal of this application is to be a key tool in a complete feedback loop from users making use of FTW data, seeing that the model performance is poor in the area they care about, spending time to improve / correct the data, and submitting it to be incorporate into the core FTW data, for testing and training.
For FTW benchmark data '2.0' we'll be doing everything on the ftw-grid. All the training data will be cut from there, and anything in that form will be compatible. So the core idea of this app will be to let users select a specific grid cell from ftw-grid and 'complete' it - delineate every field in it, so it can go directly into testing or training data.
The editing tools should be good enough that a user who wants to start from scratch can quickly edit an area, with all the capabilities that advanced GIS editing tools enable, but tailored for this specific use case. The biggest one is 'split' - be able to draw an outline and then just draw a line across the field boundaries to create individual fields. And generally snapping should work well. I'm not an editing expert, so feel free to add more here @ahocevar
The other thing the app should do is enable 'AI assistance'. The biggest one here is to run FTW models (in the same way as the inference app to start - we could explore some in browser stuff too) and then be able to 'start' with those results, correcting it to make it better. It could also be interesting to experiment with some lighter weight CV/ML tools, like segment anything, to help users work in the browser.
There should be some 'submission' process where the output goes into a queue for review, and if successful it goes into an 'ftw community dataset' area, and that can be immediately used by people making ftw models, and it will undergo further review for incorporation into the FTW dataset.
There is a whole set of functionality around managing users, reviews, scoring and submissions. I think we ideally have a pretty clear / clean architecture for the core 'editing' application, so it can easily be reused / remixed by others. But to do this well we will need the 'other' stuff - let users log in and see what they've done, assign tasks, prioritize reviews, have some queue to review what others have done. We should discuss what we can do for the MVP, and also how to architect things so the editing can be fairly standalone.
The one bit of functionality that would be nice to keep in scope is some basic 'review' thing. Like instead of the only path being to do a complete editing of a grid cell there should be the ability to just 'score' a given grid cell. This could be users who just score the results of an FTW model - they don't have the ability to correct, but they could flag that it is bad. Ideally we have some examples of like a scale of 1 to 10, to help people pick a right score. It could also be users reviewing editing submissions of others, to make sure it's high quality.
Long term we'd like to easily make other editing apps, for buildings, for crop classification, for land classification, etc. I'm not a fan of trying to build the most generic application from the start - I prefer to build for purpose and then extract from there. But it is likely worth knowing for any architecture choices that could make that harder in the future.