A new deep convolutional neural network for fast hyperspectral image classification Review

Basically, this paper imply a 3D CNN.

Points 1:Three ways to apply CNN on remote sensing image

1. Extract only spectral information: Nx1 dimension Vector, N represents the band of specture.
2. Extract only spatial information: Consider neighboring pixels of original pixel.
3. Fusion the above two models to improve accuracy.

Points 2: Proposal CNN:

Points 3: Data processing

In image processing part, divide the who hyperspectral image into several dxdxn patches where d is the width and height of the batch which n is the total No. of bands.

Highlight:

in order to suppression overfitting and handle more information in each patch, this paper rise a simple idea to mirror the d/2 pixels outwards each edge pixels.