Read Multispectral data for Remote Sensing using LightlyDataset

[imantha-das]

Not so much as an issue, but I am trying to use Self-Supervised Learning on a remote sensing dataset and wondering if there is a way to read Multispectral data which has beyond the 3 RGB Bands. The data I have compiled is from Sentinel2A and has 13 bands, so I believe that LightlyDataset cannot read these as they are loaded through PIL from what I understand.

Any idea how I can extend the LightlyDataset to more than the 3 RGB bands ?. Also if this is possible will it still cause issues with other functions such as transformations (e.g DinoTransforms) ?.
I am speculating here, but will things like color jitter or rotations be applicable in such cases.

[guarin]

Hi! I suggest to use TorchGeo for multispectral data. They have a example using SimCLR here: https://github.com/microsoft/torchgeo/blob/main/torchgeo/trainers/simclr.py I believe it is only needed for data loading/augmentations. The actual model implementation should not depend on the number of bands.