Date

2018

Author

Kütük, Mustafa

Metadata

Show full item record

Item Usage Stats

241
views

254
downloads

Cite This

Spectral signature-based methods which form the mainstream in hyperspectral target detection can be classified mainly into three categories. These are the background modeling methods, subspace projection based methods, and hybrid methods that combine linear unmixing with abundance estimation. A common characteristic of all these methods is to classify each pixel of the hyperspectral image as a target or background while ignoring the spatial relations between neighboring pixels. Integration of contextual information defined over neighboring pixels can, however, suppress the noise on the individual pixels and yield better detection. In this thesis study, the baseline superpixel extraction algorithms which are previously developed for RGB images, namely the Simple Linear Iterative Clustering (SLIC) algorithm and boundary update-based superpixel extraction method, are first adapted to hyperspectral images. Then their extraction performances are compared in terms of the metrics which are boundary recall and undersegmentation error. After the selection of the boundary update-based superpixel extraction algorithm due to its better performance, different target detection methods performing over superpixels are proposed. The proposed methods utilize superpixel representatives instead of pixels for background modeling, matching and abundance estimation. The experiments suggest that using superpixels for target detection improves the detection performances in terms of precision-recall curves compared to the baseline methods using only pixels.

Subject Keywords

Hyperspectral imaging., Spectral imaging., Remote sensing.

URI

http://etd.lib.metu.edu.tr/upload/12622703/index.pdf
https://hdl.handle.net/11511/27528

Collections

Graduate School of Natural and Applied Sciences, Thesis