A Supervised Nonlinear Spectral Unmixing Method by Means of Neural Networks

Spectral unmixing (SU) technique is a conventional approach aiming at tackling the mixed pixel issue in hyperspectral imagery. In this study, we investigate the application of the convolutional networks (ConvNets) in SU. In fact, we propose two ConvNets, namely 1-D and 3-D in order to employ spectral and spectral-spatial information of the hyperspectral images, respectively. The proposed configurations are adapted by the VGG-Net that showed increasing depth with very small convolution filters (3×3) results in a substantial improvement. More importantly, rather than using the common mean squared error as the objective function, spectral information divergence is employed. Also, other network parameters, such as the number of kernels and kernel sizes are set according to the best findings in state-of-the-art methods in order to validate the network architecture. To demonstrate the superiority of the proposed ConvNets, some experiments are carried out on two well-known real hyperspectral data sets, namely Samson and Jasper Ridge, and the results are compared with some celebrated HSU approach in the literature. Moreover, the visual and quantitative assessments of the presented ConvNets display the necessity of incorporating spatial and spectral information in HSU.