A new band selection framework for hyperspectral remote sensing image classification.

Dimensionality Reduction (DR) is an indispensable step to enhance classifier accuracy with data redundancy in hyperspectral images (HSI). This paper proposes a framework for DR that combines band selection (BS) and effective spatial features. The conventional clustering methods for BS typically face hard encounters when we have a less data items matched to the dimensionality of the accompanying feature space. So, to fully mine the effective information, BS is established using dual partitioning and ranking. The bands from the dual partitioning have undergone informative band selection via ranking. The reduced band subset is then given to a hemispherical reflectance-based spatial filter. Then, finally, a Convolutional Neural Network (CNN) is used for effective classification by incorporating three-dimensional convolutions. On a set of three hyperspectral datasets - Indian Pines, Salinas, and KSC, the proposed method was tested with different state-of-the-art techniques. The classification results are compared using quantitative and qualitative measures. The reported overall accuracy is 99.92% on Indian Pines, 99.94% on Salinas, and 97.23% on the KSC dataset. Also, the Mean Spectral Divergence values are 42.4, 63.75, and 41.2 on the three datasets respectively, which signifies the effectiveness of band selection. The results have clearly shown the impact of the band selection proposed and can be utilized for a wide variety of applications.