Sparse demixing of hyperspectral images

Area of Science:

• Geoscience
• Remote Sensing
• Signal Processing

Background:

• Hyperspectral imaging relies on linear mixing models (LMM) where spectra are combinations of endmembers.
• Traditional methods assume pixels combine most endmembers, leading to dense abundance vectors.
• This limits the accurate identification of underlying material compositions.

Purpose of the Study:

• To introduce a novel sparse demixing (SD) method for hyperspectral image analysis.
• To develop an algorithm that assumes pixels are combinations of only a few endmembers, resulting in sparse abundance vectors.
• To improve the accuracy of endmember extraction and abundance estimation in hyperspectral data.

Main Methods:

• Developed sparse demixing (SD), an orthogonal matching pursuit-like algorithm for calculating sparse abundances.
• Combined SD with dictionary learning for automated endmember extraction.
• Applied the method to airborne visible/infrared imaging spectrometer data from Cuprite, NV.

Main Results:

• SD significantly outperforms existing L(1) demixing algorithms.
• The performance of L(1) algorithms is adversely affected by the angles between endmembers, a limitation addressed by SD.
• Endmembers extracted using SD and dictionary learning favorably compare with USGS spectral library signatures.

Conclusions:

• Sparse demixing (SD) offers a more accurate approach to analyzing hyperspectral data compared to traditional methods.
• Automated endmember extraction using SD and dictionary learning is effective for real-world datasets.
• The developed method enhances the potential of hyperspectral imaging for material identification and mapping.