C4 - Transfer learning using Multi- or Hyperspectral data

Hi there,

in my understanding, most of the CNN architectures out there are pre-trained with RGB images (e.g. 64x64x3) like imagenet an similar. Is it possible to use these on lets say multispectral imagery (e.g. 64x64x5)? Instead of just changing the top layers to my classes of interest, do I have to initialize the first layers somehow?

For a remote sensing task:
Any suggestions to also add normalized digital surface models representing the objects height as an additional dimension?

Would be really nice to get some recommendations here :slight_smile:

Here is an article I found discussing implementing transfer learning from RGB to multi-channel, I hope it helps.