An efficient two-layer classification approach for hyperspectral images

Abstract

Different 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 detection 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 generated 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.