Hyperspectral Image classification using pretrained model



I just started to work on the project which uses deep learning to classify health/sick plant
from hyperspectral images.
Since the dataset size is not so large(around few thousands), I am planning to utilize
pretrained models such as ResNet or VGG.

Existing models usually accept RGB or grayscale images, so I am thinking to add several layers to
the front(input part) of a pretrained model to reduce the channels of hyperspectral data to 1 or 3.
I saw many example adding layers to the end of pretrained models, but does anyone know whether adding new layers to the front also work well or not?

Thank you very much.

(Dennis Sakva) #2

You can also replace first conv layer with one that takes your number of channels and copy weights for missing channels from your initial weights (like from green channel) and make first conv layer trainable as well.