Exploring Spectral and Spatial Features Using a Hybrid Approach Combining Stacked AutoEncoder and a Novel Convolutional Neural Network for Hyperspectral Image Classification

With the introduction of high-resolution hyperspectral sensors, hyperspectral images have become one of the paramount mediums of collecting information from remote places. Owing to the enormous dimension of spectral bands and high correlation between the bands, proper classification of hyperspectral images suffers seriously. Furthermore, proper exploration of merged spectral and spatial features remains challenging for traditional approaches. Keeping the above challenges in mind, we have proposed a properly tuned Stacked AutoEncoder(SAE) and a novel Convolutional neural network (CNN) architecture that simultaneously considers the output of all the convolutional blocks. We have used a benchmark hyperspectral dataset called KSC center. Experimental results have shown that our method has achieved an average accuracy of 99.50%, surpassing other state-of-the-art approaches significantly.