Hey dj, sorry for the late reply. There is no great way to adapt Prithvi to
accept 4 input channels without blowing away the existing architecture. One
trick could be to broadcast your NIR to 3 channels (RGB + NIR + NIR + NIR)
to meet that 6-channel requirement. In my experience, Prithvi is a bit
difficult to use for those who are new to DL. Hence why I put together this
guide for those who might be struggling to use it. I honestly suggest
looking at torchgeo https://github.com/microsoft/torchgeo and fastai
https://github.com/fastai/fastai for your geospatial deep learning project.
If you can adapt your data to be 3-band then there is a great selection of
pre-trained ImageNet models to use. Good luck dude!
On Thu, May 23, 2024 at 5:03 AM dj1994 @.***> wrote:
I'm totally new in DL and RS. What im trying to do is to finetune Prithvi
model for one specific task. The main idea is first, I will Fine-tune the
model with my own dataset that consist in Satellite imagery (RGB+NIR). My
question is if there is any problem by finetuning the model with a
different type of data in this case from MS(6 Bands) from the original
model to (4 Bands). Do you have any recommendation or suggestion for this?
Hey dj, sorry for the late reply. There is no great way to adapt Prithvi to accept 4 input channels without blowing away the existing architecture. One trick could be to broadcast your NIR to 3 channels (RGB + NIR + NIR + NIR) to meet that 6-channel requirement. In my experience, Prithvi is a bit difficult to use for those who are new to DL. Hence why I put together this guide for those who might be struggling to use it. I honestly suggest looking at torchgeo https://github.com/microsoft/torchgeo and fastai https://github.com/fastai/fastai for your geospatial deep learning project. If you can adapt your data to be 3-band then there is a great selection of pre-trained ImageNet models to use. Good luck dude!
On Thu, May 23, 2024 at 5:03 AM dj1994 @.***> wrote: