Abstract
In this paper, an unsupervised unmixing approach based on superpixel representation combined with regional partitioning is presented. A reduced-size image representation is obtained using superpixel segmentation where each superpixel is represented by its mean spectra. The superpixel image representation is then partitioned into regions using quadtree segmentation based on the Shannon entropy. Spectral endmembers are extracted from each region that corresponds to a leaf of the quadtree and combined using clustering into endmember classes. The proposed approach is tested and validated using the HYDICE Urban and ROSIS Pavia data sets. Different levels of qualitative and quantitative assessments are performed based on the available reference data. The proposed approach is also compared with global (no-regional quadtree segmentation) and with pixel-based (no-superpixel representation) unsupervised unmixing approaches. Qualitative assessment was based primarily on agreement with spatial distribution of materials obtained from a reference classification map. Quantitative assessment was based on comparing classification maps generated from abundance maps using winner takes it all with a 50% threshold and a reference classification map. High agreement with the reference classification map was obtained by the proposed approach as evidenced by high kappa values (over 70%). The proposed approach outperforms global unsupervised unmixing approaches with and without superpixel representation that do not account for regional information. The agreement performance of the proposed approach is slightly better when compared to the pixel-based approached using quadtree segmentation. However, the proposed approach resulted in significant computational savings due to the use of the superpixel representation.
Highlights
Hyperspectral imaging (HSI) systems collect spectral and spatial image data simultaneously in hundreds of narrow contiguous spectral bands
Quadtree segmentation was applied to the full hyperspectral image, and the number of spectral endmembers was determined for each tile using the Gram approach
This work proposed a new spatial-spectral approach for unmixing analysis using superpixel representation combined with quadtree segmentation
Summary
Hyperspectral imaging (HSI) systems collect spectral and spatial image data simultaneously in hundreds of narrow contiguous spectral bands. These signatures will be hard to extract using the geometric endmember extraction approaches described previously Because of their relevance to our work, we only review spatial-spectral approaches that use spatial sub-setting to improve the detection of small and low contrast endmembers. Uniform tiling is used in [17] for unmixing where the tiling artifact in the abundances is avoided by doing global computation of the abundances using all extracted spectral endmembers Another approach that uses image spatial sub-setting for unmixing is presented in [18,19], where spatial sub-setting of the image is done using quadtree partitioning (QT). The linear mixing model (LMM) represents a pixel as the convex combination of the spectral signatures of the endmembers multiplied by its fractional area coverage (or abundance) as follows:. The spatial relations are not considered when unfolding the hyperspectral image into a matrix as described in Equation (2)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have