Determination of Relevant Hyperspectral Bands Using a Spectrally constrained CNN
Ulrike Pestel-Schiller, Kai Hu, Daniel Gritzner, Jörn Östermann · 2021
For transmitting the large amount of hyperspectral image (HSI) data over a small data link from a small platform to the ground, an efficient data compression with low computational cost has to be done at the platform. Additionally, spectral band reduction interpreted as preprocessing of the compression is reasonable. We present a method for hyperspectral band reduction using a modified convolutional neural network (CNN) which retains the information about the spectral origin from layer to layer until it can be assigned directly to the classes to be classified. The relevant bands for each class are determined. Experimental verification shows that the network architecture using only the relevant bands has improved stability and results in a better overall performance.