Feature reduction in spectral unmixing using neural networks

Spectral unmixing (SU) is a standard approach to solve the mixed pixel problem in hyperspectral (HS) images. In this study, the application of feature reduction in SU using multi-layer perceptron (MLP) with some data-independent approaches is investigated. MLP is a popular artificial neural network that can learn complex nonlinear relationships between the endmembers and the abundance fractions in HS images if it is properly trained. So far, various approaches have been introduced to extract training samples from the data itself. Since it is not possible to access the actual abundance fractions of materials in real HS images, MLP training becomes complicated. Due to a large number of bands in HS images, complexity and large training time are some of the remaining problems that would be investigated in this study. In order to overcome the problem of unavailability of the actual abundance fractions, a synthetic library is generated based on scene mixture models. And some data-independent approaches, such as discrete cosine transform and discrete wavelet transform are utilized to reduce the complexity and the training time of the MLP. The experimental results are provided using both synthetic and real datasets with different mixture models. The results show the acceptable estimated abundance fractions with root mean square error, up to 0.0008 in the linear dataset and 0.0062 in the nonlinear dataset.