Mapping water tracks in the Arctic! #62
Replies: 4 comments 6 replies
-
oh wow, this is rad. Thank you so much for this message @jmdelvecchio - brightened my day for sure.. And i love seeing output like this. I have seen (recent) results from @dbuscombe-usgs using 4+ band imagery. i think its just having a another folder of 4th band images (in addition to image+label dirs). the i also want to say that you might have some success 'remapping' the classes via the config and collapsing the problem to a binary one - 'streaks' and 'nostreaks'. especially in the limit of small datasets, this has been helping us alot. (but more labeled data is always worth the most in terms of increasing model performance. for 19 images, i think your curve looks pretty rad!!) Keep us posted!!! |
Beta Was this translation helpful? Give feedback.
-
@jmdelvecchio this makes me v v happy, thanks for the message - you rock! We should better expose and properly document the 'custom satellite' unet. Initial trials would suggest that it is inferior to both the res-unet and unet implementations, but only on limited RGB-only datasets. So, I would encourage experimentation! (I don't think we've tested that model on >3 band images, but it's now on the to-do list) As @ebgoldstein said, I would second what @ebgoldstein suggested about playing with binary classes. Use the And yes, more data should really help the model generalize. Glad you got Doodler working for you! hyperparameter choices (such as batch size, learning rate, kernel size, and number of filters) can strongly improve results on small* datasets, but it is our experience that greater improvement is usually best achieved by labeling more, especially relatively 'difficult' images Keep us posted! We'll happily (and gratefully) add any clarifying details to the docs, too, just let us know using an Issue |
Beta Was this translation helpful? Give feedback.
-
Hey fam, just dropping in to report after a term of teaching I'm ready to return to this project, this time with a student who is paid to do all the clicking!! I'm going to start her on just Doodling the Planet RGB imagery while I hack away at adding bands. I have to figure out how to keep the pixels straight between different resolution datasets - 3 m Planet means upsampling 10 m S2 and downsampling 2 m ArcticDEM, right? I assume I can't mix pixel sizes between bands. |
Beta Was this translation helpful? Give feedback.
-
Hi, thanks for the update. Sounds like a fine plan! Yes, finding a common pixel size would be necessary using resampling. I imagine nearest neighbour resampling would work well Once you have a trained Gym model that you'd like to apply, I have codes that may help with the "geospatial" part .... I'm working up to a blog post or two. I'll ping you when they are completed |
Beta Was this translation helpful? Give feedback.
-
Hi Dan and Evan! Thank you so much for giving the world Doodler and Segmentation Zoo. Dan, I was working along with MLMondays in the fall, and to my delight this week I discovered this project. For my postdoc my goal is to automatically detect Arctic flowpaths known as water tracks and map their networks. After working with your example data (again, thank you for walking us through the process and setting up some pretty painless programs) I chopped up my PlanetScope imagery from western Alaska and I jumped right in!
I scribbled on a mere 18 images from my ~250 total for a June 2019 scene in western Alaska that included classes for snow, shrubs, "streaks" (what I'm calling not-quite-shrubby water tracks), rock, "background" (any non-textured tundra), and nodata. I used the "vanilla" settings and I'm so impressed with the results, considering I still don't know how to tune these models:
Like I said, I'm really new, but this is a great start. Open to suggestions, I know that accuracy is pretty low but to me it's amazing😂. Doodler in particular will be awesome for my student researchers.
I noticed that you're in the process of describing the satunet in the wiki, and I want to know more about that! I messed around with that (I had to edit the .py scripts in the src folder to "simple_satunet" instead of "custom_satunet" to get it to run, was that your intention? I can make an issue/pull request), but I still only gave it my 3 RGB bands and the results were definitely wonky. I know for a fact that both spectral indices and topographic metrics would vastly improve the model prediction, so I'm eagerly awaiting instructions/examples for 4+ band inputs. With Planet's Superdove 8-band images and ArcticDEM-derived topo metrics as inputs I think this model can work out really well.
Beta Was this translation helpful? Give feedback.
All reactions