Generation of artificial data sets to train convolutional neural networks for spectral unmixing

This paper presents a method to generate training data for artificial neural networks for spectral unmixing. Therefor, only the spectra of the pure substances involved and, depending on the model used, a few spectra of mixed substances to determine the parameters are needed. With mixing models, which can also be used directly for spectral unmixing, large quantities of spectra can be generated for training. In contrast to the direct use of mixing models, where a spectrum per pure substance is used, this approach takes into account the spectrum variability by using different spectra of each pure substance. The property of artificial neural networks to learn significant features based on large amounts of data is exploited here.