Hyperspectral Unmixing Based on Spectral and Sparse Deep Convolutional Neural Networks

Hyperspectral unmixing refers to the process of obtaining endmembers and abundance vectors through linear or nonlinear models. The traditional linear unmixing model assumes that each mixed pixel can be represented by a linear combination of endmembers. Considering real-world situations, a sparse constraint is normally added to the linear unmixing model. However, the linear model does not take into account that the spectrum of mixed pixels is not simply linearly mixed. To fully study the mixing characteristics of ground object spectra before being imaged by the sensor, we propose a supervised unmixing architecture based on a one-dimensional convolutional neural network (CNN) by considering the spectral information and the sparse characteristics in the mixed pixel. Since 1-D CNN only considers feature learning, we combine the traditional root-mean-square error (RMSE) and
${\ell _{1}}$
regularization in its loss function to minimize training error. The performance of our proposed unmixing model is assessed by comparing the unmixing results with three traditional linear sparse unmixing algorithms and the fuzzy ARTMAP neural network in a simulated dataset and three real datasets. The RMSE was used to verify the unmixing accuracy of the different methods. The results showed that the RMSE obtained by our proposed CNN-based method was the lowest among the methods on all three real datasets, proving the effectiveness and stability of the CNN in unmixing tasks.