Hybrid Unmixing Based on Adaptive Region Segmentation for Hyperspectral Imagery
Xiangrong Zhang,Jingyan Zhang,Chen Li,Cai Cheng,Licheng Jiao,Huiyu Zhou
DOI: https://doi.org/10.1109/tgrs.2018.2815044
IF: 8.2
2018-01-01
IEEE Transactions on Geoscience and Remote Sensing
Abstract:Unmixing is an important issue of hyperspectral images. Most unmixing methods adopt linear mixing models for simplicity. However, multiple scattering usually occurs between vegetation and soil in a bilinear scene. Thus, nonlinear mixing problems which are difficult to be solved should be taken into consideration under this circumstance. In practice, both linear and nonlinear spectral mixtures exist in hyperspectral scenes. Considering the characteristics of different regions in images, we propose a hybrid unmixing algorithm for hyperspectral images based on region adaptive segmentation. Our method uses a standard K-means clustering algorithm to obtain different regions, including homogeneous regions and detailed regions. The model of the homogeneous regions is assumed to be linear, which will be pursued using the method of sparse-constrained nonnegative matrix factorization (NMF), and the mixing in the detailed regions is assumed to be based on a nonlinear model. We also propose a new nonlinear unmixing method, called graph-regularized semi-NMF, which considers the manifold structure of hyperspectral data as the unmixing method to deal with the detailed regions. Finally, by combining the two regions, we obtain the abundance of the whole hyperspectral image. The proposed method can not only achieve more precise abundance but also be good at keeping the edge information of the bilinear abundance. The experimental results on both synthetic and real data also show that the proposed method is effective for improving the unmixing accuracy of hyperspectral remote-sensing images.