A Fast Approximate Spectral Unmixing Algorithm Based on Segmentation

Jing Ke, Yi Guo, Arcot Sowmya; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017, pp. 66-72

Abstract


Efficient processing of spectral unmixing is a challenging problem in high-resolution satellite data analysis. The decomposition of a pixel into a linear combination of pure spectra into their corresponding proportions is often very time-consuming. In this paper, a fast unmixing algorithm is proposed based on classifying pixels into a full unmixing group for subset selection requiring intensive computational procedures and a partial unmixing group for proportion estimation with known spectra endmembers. The classification is based on n-band spectral segmentation using the quick-shift algorithm. A subset selection algorithm applied on real satellite data evaluates accuracy and approximation, and experimental results show significant performance acceleration compared with the original algorithm. Parallelization strategies are also presented and verified on NVIDIA GTX TITAN X.

Related Material


[pdf]
[bibtex]
@InProceedings{Ke_2017_CVPR_Workshops,
author = {Ke, Jing and Guo, Yi and Sowmya, Arcot},
title = {A Fast Approximate Spectral Unmixing Algorithm Based on Segmentation},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {July},
year = {2017}
}