A Multi-Layer Classifier for Hyperspectral Images
Event Type
TimeTuesday, July 306:30pm - 8:30pm
LocationCrystal Foyer and Crystal B
DescriptionDifferent from regular RGB images that only store red, green,
and blue band values for each pixel, hyperspectral images are rich with
information from the large portion of the spectrum, storing numerous
spectral band values within each pixel. An efficient, two-layer region de-
tection framework for hyperspectral images is introduced in this paper.
The proposed framework aims to automatically identify various regions
within a hyperspectral image by providing a classification for each pixel
of the image, associating them to distinct regions. The first layer of the
system includes two new classifiers, and is responsible for generating
probability scores as the “new feature set” of the original dataset. The
second layer works as an ensemble classifier and combines the newly gen-
erated features to estimate the region of the sample. Experimental results
show that the proposed system can produce accurate classifications with
an average area under the ROC curve of 0.98 over all regions. This result
indicates the higher accuracy of the proposed system compared to some
other well-known classifiers. Also, the speedup of the proposed classifier
performs very well compared with more sophisticated methods.