Abstract
Hyperspectral image (HSI) classification using convolutional neural
network requires a lot of training samples, which is not always
available. Consequently, decreases the classification accuracy due to
the overfitting problem. Many studies have been conducted to solve the
issue; however, they failed to solve it entirely. Therefore, we proposed
a new approach to classify HSI with few training samples using a
convolutional neural network in that context. The proposed approach
employed an extended morphological profile cube (EMPC) to extract rich
spectral-spatial features and then used a 3D densely connected network
for classification. Besides, we used sparse principal component analysis
to reduce the high spectral dimension of HSI. Experiments results on
Indian Pines (IP) and University of Pavia (UP) datasets proved the
efficiency of the proposed approach. It increased the OA by 2.61% -
13% and the Kappa coefficient by 2.68% - 15:51% on IP dataset and
increased the OA by 0.17% - 11% and the Kappa coefficient by 0.23% -
19% on UP dataset, which is superior to some state-of-art methods.