Abstract
Deep neural networks, an emerging paradigm in deep learning, have proven to make feature extraction from remote sensing data easier. Deep learning has been shown to be capable of effectively classifying hyperspectral images (HSI). Deep convolutional neural networks (CNNs) are one of the most effective approaches for HSI classification. The deep learning architecture needs to be capable of providing a better spatial-spectral classification performance. In increasing the depth of layers in CNN might lead to overfitting issues. As spatial-spectral information is not correlated along different layers, hence information is lost. This paper attempts to solve these problems by presenting an enhanced-CNN. Initially, proposed e-CNN method explores the merging of the outputs of successive two layers within the huge convolutional block and the merged feature extract outcome is fed as the input to the next layer, which renders relevant feature extraction. Then, from low-level layers to deep high-level layers, spectral-spatial features are retrieved by concatenating the spectral features to four-stage spatial features. Finally, to communicate with hybridized extracted feature information, a 1 x 1 convolution layer is used throughout the block. With the limited training samples and the provided pixel-size the proposed e-CNN model works much effectively. In order to obtain the standard generalizing ability of classification an adaptive AdaBound optimization method is used. Finally, HSI classification is performed with the enhanced CNN model. The existing models and optimizers (SGD, AdaGrad, AdaDelta, Adam) are used to compare the results. The experiments were carried out on widely used HSI datasets (i.e., Indian Pines and Salinas) and the result of the proposed e-CNN model with AdaBound optimizer obtains ~2% higher accuracy compared with existing methods. Optimization result of e-CNN model with AdaBound optimizer have the highest classification accuracy in the least amount of time.