Hyperspectral images are a series of images, which are captured for a specific region over a range of wavelengths. This makes the classification process computationally more expensive. For reducing the computational complexity, instead of considering all bands, it is essential to select the most informative bands. In this paper, a probabilistic histogram-based band selection approach is proposed. Here, adjacent band fusion with a class-specific deviation is computed followed by extraction of fused band intra- and inter-class histogram properties, to rank the bands with ensemble probability. In both the steps, median measure is used to half the total dimension. So finally, one-fourth of the optimal bands are obtained. Both spectral and spatial features of the reduced bands are considered for classification using KNN with different distance measures. Performance measures like accuracy and execution time are compared. Even by considering only 5% of optimal bands, the proposed approach maintains reference accuracy with reduced computational complexity.
展开▼
