Gaussian Process Inspired Neural Networks for Spectral Unmixing Dataset Augmentation

Abstract

Hyperspectral imaging is increasingly used for product monitoring in industrial processes. Spectral unmixing is an important task in this context. As in many other areas of signal processing, neural networks also provide promising results for spectral unmixing. Unfortunately, it is very time-consuming to prepare labelled training data for the neural networks. To address this problem, this paper presents a method where small training datasets are augmented to improve spectral unmixing performance. Inspired by Gaussian processes, simple neural networks are trained which are capable of generating additional training data. These are similar to the original training data but cover areas in the continuous label space that are not covered by the original data.