The use of genetic search algorithms (GA) as spectral band selectors is popular in the field of remote sensing. Nevertheless, class information that has been used in the existing research for testing the performance of the GA-based band selector is broad (i.e. Anderson's level I or II). This means that each class possesses distinct spectral characteristics from one another, and it is relatively easy for the band selector to find spectral bands that maintain high spectral separability between classes. None of the existing studies has tested the band selector on class information that possesses very similar spectral characteristics (e.g. species-level data). A question therefore remains if the band selector can deal with such complexity. As a result, the key hypothesis of this research is that the GA-based band selector can be used for selecting a meaningful subset of spectral bands that maintains spectral separability between species classes. The testing data in use are very high-dimensional, spectrometer records that comprise 2151 bands of leaf spectra of 16 tropical mangrove species. The results turned out that the GA-based band selector was able to cope with spectral similarity at the species level. It meaningfully selected spectral bands that related to principal physio-chemical properties of plants, and, simultaneously, maintained the separability between species classes at a high level.
|Number of pages||11|
|Journal||ISPRS Journal of Photogrammetry and Remote Sensing|
|Publication status||Published - Aug 2007|
- hyper spectral
- remote sensing