Spectral-spatial adversarial network for nonlinear hyperspectral unmixing of imbalanced datasets

With its successful application in various fields, hyperspectral unmixing (HU) technology has received extensive attention in remote sensing processing. Recently, various autoencoders based on the linear mixing model (LMM) have been proposed to provide a feasible unsupervised solution for HU. However, the ability of autoencoders to exploit the prior properties of the latent spatial distribution is limited, and ground objects with a small number may be under-fitted during the training process on an imbalanced dataset. Moreover, LMM may not be applicable for scenarios where multiple scattering occurs. To address the above problems, we develop a novel spectralspatial adversarial autoencoder (SSAAE) for nonlinear HU of imbalanced datasets. First, the superpixel segmentation technique is employed to partition the image into a collection of locally homogeneous regions where the abundances in the same region are assumed to follow the same multivariate Gaussian distribution. Second, in the adversarial process, the region-based prior distribution is imposed on the aggregated posterior of abundances estimated by the encoder. Third, the generalized bilinear model (GBM) with an unfolding form is integrated into a decoder to extract endmembers and estimate the quadratic interactions. Finally, we employ k-means to roughly cluster the image and count the sample weights to implement the cost-sensitive loss function to overcome the detrimental problem caused by imbalanced data. Comparison experiments conducted on both synthetic and realworld datasets are employed to certify that SSAAE outperforms seven other state-of-the-art approaches and prominently enhances the robustness and accuracy of the unmixing performance.