Effective and Efficient Dimensionality Reduction of Hyperspectral Image using CNN and LSTM network

Convolutional neural networks (CNN) which is a feature-based machine learning algorithm is very popular in hyperspectral image (HSI) classification. CNN exploits the spatial relationship between HIS. However, HSI intrinsically have a sequence-based data structure called the spectral features. Combining spectral and spatial information offers a more comprehensive classification approach. 3D-CNN can exploit Spatial-spectral relationship but can be computationally expensive. LSTM, an important branch of the deep learning family, is mainly designed to handle Sequential data. In this paper we propose a model that uses the 1D CNN and 2D-CNN for extracting the spatial features and a LSTM for extracting the spectral features. Experimental results show that our method outperforms the accuracies reported in the existing CNN and LSTM based methods.