A Correlation based Band Selection Approach for Hyperspectral Image Classification

The recent advancement in remote sensing has made it possible to capture hyper spectral images with more than hundred bands which include spectrum beyond the visible range as well. This increased number of spectral dimension gives detailed information about the objects and hence increases the classification accuracy. But at the same time it also increases the computational complexity. So, reducing the number of bands without much compromising the information content has been a challenge in the field of hyper spectral image classification. This paper attempts to address a correlation based approach for band selection. This approach entails calculation of the correlation among the bands of the hyper spectral image and subsequent selection of those bands having correlation less than a threshold value. The experimental results obtained, have shown that with only a very limited number of bands we can achieve accuracy closer to that of using all the bands.