use torchgeo Samplers to select images with some non-0 values

[MohajaneM]

Hello everyone ,

I would like to ask how to select images with some non-0 values ( I mean to delete images where no labels presented ) using Torchgeo using Samplers. Thank you

[adamjstewart]

This feature does not currently exist in TorchGeo, but it's something we plan to add. Basically, we want to support weighted sampling where the user can provide a list of weights with which each image can be sampled. Obviously, images with a weight of 0 will be skipped entirely.

[MohajaneM]

Thank you for your reply .
I am using Sentinel 2 with 6 bands and my labels raster file is composed of 8 classes , but whene I save the patchs and masks some have zero labels and do like to skip them for the training

[adamjstewart]

You could always manually delete or move them, that's the only workaround at the moment.

[hkristen]

I had the same issue and could solve it with a custom RandomGeoSampler that only samples when the ROI intersects with a provided vector label = when there are labels in the roi. Check out this GIST for it:
https://gist.github.com/hkristen/4b580b203a6ec1cb7e6d1c927f296e1a

-> it's sufficiently fast for my use, but ofc you need your labels in vector format for this to work.

I was also trying to achieve this with raster labels only, basically checking each sample that was created if it contains pixel values > 0, but it was kinda slow. If you only have raster labels, this might be a good option if you can make it fast enough with some numpy magic or so.

[MohajaneM]

Hello @hkristen.
Thank you very much for your reply.
Yes actually I am using a raster and I did delete theme manually.
Thanks again