A novel spectral-spatial singular spectrum analysis technique for near real-time in-situ feature extraction in hyperspectral imaging

Fu, Hang and Sun, Genyun and Zabalza, Jaime and Zhang, Aizhu and Ren, Jinchang and Jia, Xiuping (2020) A novel spectral-spatial singular spectrum analysis technique for near real-time in-situ feature extraction in hyperspectral imaging. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 13. pp. 2214-2225. ISSN 2151-1535

[img]
Preview
Text (Fu-etal-IEEE-2020- spectrum-analysis-technique-for-near-real-time-in-situ-feature-extraction-in-hyperspectral-imaging)
Fu_etal_IEEE_2020_spectrum_analysis_technique_for_near_real_time_in_situ_feature_extraction_in_hyperspectral_imaging.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo

Download (4MB)| Preview

    Abstract

    As a cutting-edge technique for denoising and feature extraction, singular spectrum analysis (SSA) has been applied successfully for feature mining in hyperspectral images (HSI). However, when applying SSA for in situ feature extraction in HSI, conventional pixel-based 1-D SSA fails to produce satisfactory results, while the band-image-based 2D-SSA is also infeasible especially for the popularly used line-scan mode. To tackle these challenges, in this article, a novel 1.5D-SSA approach is proposed for in situ spectral-spatial feature extraction in HSI, where pixels from a small window are used as spatial information. For each sequentially acquired pixel, similar pixels are located from a window centered at the pixel to form an extended trajectory matrix for feature extraction. Classification results on two well-known benchmark HSI datasets and an actual urban scene dataset have demonstrated that the proposed 1.5D-SSA achieves the superior performance compared with several state-of-the-art spectral and spatial methods. In addition, the near real-time implementation in aligning to the HSI acquisition process can meet the requirement of online image analysis for more efficient feature extraction than the conventional offline workflow.