CNN based hyperspectral image classification using unsupervised band selection and structure-preserving spatial features

Hyperspectral image (HSI) consists of hundreds of contiguous spectral bands, which can be used in the classification of different objects on earth. The inclusion of both spectral and as well as spatial features is essential for better classification accuracy. Extraction of spectral and spatial information without preserving the intrinsic structure of the data will downscale classification accuracy. To address this issue, we have proposed a method that uses unsupervised band selection called optimal neighboring reconstruction (ONR), which extracts a subset of spectral bands to linearly reconstruct the original data with minimum loss and Structure-Preserving Recursive Filter (SPRF) to extract spatial features. Then we have adopted Convolutional Neural Networks (CNN) with different sets of convolutional, pooling, and fully connected layers for classification of the data. To test the performance of proposed method, experiments are conducted with three benchmarks HSI data sets Indian pines, University of Pavia, and Salinas. These experiments reveal that the proposed method performed better classification accuracy over state-of-art methods in terms of standard metrics like Overall Accuracy (OA), Average Accuracy (AA), and kappa coefficient (k). The proposed method has attained OA's of 99.9%, 98.9%, and 99% for the three datasets, respectively.