Figure 9.

Convolutional filters and feature maps of an automated 2D-CNN feature extractor (a) and 3D-CNN feature extractor (b) for 24 training HSIs and 16 testing HSIs. From the original dimension 512 × 512 pixels × 203 bands (width × height × spectral bands), hyperspectral data cubes were reduced to a new and lower-dimensional data space of 16 × 16 pixels × 64 filters (width × height × convolutional filters) by the 2D-CNN feature extractor and 16 × 16 pixels × 5 bands × 64 filters (width × height × spectral bands × convolutional filters) by 3D-CNN feature extractor. The reduced spatial features (16 × 16) were plotted against the convolutional filter values (a random selection from 64 units). The graduated scale represents the convolutional values.