Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
Unsupervised methods for the classification of hyperspectral images with low spatial resolution
Villa A., Chanussot J., Benediktsson J., Jutten C., Dambreville R.  Pattern Recognition 46 (6): 1556-1568, 2013. Type: Article
Date Reviewed: Oct 15 2014

Spectral imaging constitutes a fascinating type of data. Organized as a set of grayscale images, with each image recording the reflected light of the same scene under a different light wavelength range, a spectral image can reveal significantly more than a color image. Slight dissimilarities between materials can be easily identified by pinpointing the one or more wavelengths for which the grayscale images expose a contrast. Given this, it is expected that spectral imaging will continue to expand its areas of applicability. Detecting materials in the scene reduces to matching spectra (pixel vectors formed of pixels in the same location throughout the collection of grayscale images) with vectors known to represent the materials.

Yet, fundamental problems continue to exist. One such problem is that of mixed pixels, that is, pixels that, due to limited spatial resolution, cover more than one material, and are thus a mixture of materials. In that case, previous techniques would merely produce the formulas and parameters with which the materials were mixed. Discovering which specific part of the pixel corresponded to which material is a much more complex job and often required fusion of the spectral data with other, less expressive, yet of higher spatial resolution, imagery. Unfortunately, such additional data is not often available.

The method described here takes a different route. Following the determination of abundances through which the mixed pixels are formed, the image is superscaled (that is, each pixel is replaced by a number of pixels). Next, the exact location of the materials within the pixel is determined through the adaptation of classic (simulated annealing) and the development of new (pixel swapping) techniques. The result is a classification of the materials done at a resolution higher than the original. The results are quite impressive, significantly surpassing previous work.

Overall, this is a paper that promises to open new directions in the research and applications of spectral data. Written in a clear scientific style and supported by a strong list of references, the paper should be a must read for any interested researcher, student, or practitioner.

Reviewer:  Stefan Robila Review #: CR142831 (1501-0099)
Bookmark and Share
  Editor Recommended
Featured Reviewer
 
 
Classifier Design And Evaluation (I.5.2 ... )
 
 
General (I.5.0 )
 
Would you recommend this review?
yes
no
Other reviews under "Classifier Design And Evaluation": Date
A survey on ensemble learning for data stream classification
Gomes H., Barddal J., Enembreck F., Bifet A.  ACM Computing Surveys 50(2): 1-36, 2017. Type: Article
Jun 16 2017
Biological signals classification and analysis
Kiasaleh K.,  Springer International Publishing, New York, NY, 2015. 380 pp. Type: Book (978-3-642548-78-9)
Jul 1 2016
Personality in speech: assessment and automatic classification
Polzehl T.,  Springer Publishing Company, Incorporated, New York, NY, 2014. 176 pp. Type: Book (978-3-319095-15-8)
Nov 12 2015
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright © 2000-2017 ThinkLoud, Inc.
Terms of Use
| Privacy Policy