ParHybNet: Parallel Hybrid Network for Hyperspectral Image Classification

Deep Learning approaches, particularly Convolutional Neural Networks, have recently proven to be adequate for the Hyperspectral Image Classification (HSIC). These CNN-based techniques have performed admirably in a variety of applications. It is known to all that the dimension of the input image is lowered after each convolution process, which may result in feature loss. However, the scarcity of training samples remains a key barrier in HSI classification, significantly affecting classification performance. In order to address this issue, Parallel Hybrid Network (ParHybNet) has been proposed, an HSI classification (HSIC) approach, which employs two identical hybrid spectral-spatial feature extractors in parallel. The HSI cube is sent via a Triplet-Attention module in the proposed model before being subjected to feature extraction by two parallel feature extractors, and the resultant feature maps from these parallel extractors are then concatenated. For classification, the combined feature maps are passed through a sequence of fully connected layers before being sent to a softmax activation layer. The method proposed has been compared with other well-established methods for performance analysis using three well-known and publicly available datasets of hyperspectral images: Salinas Scene, University of Pavia, and Indian Pines. The proposed approach outperforms existing approaches in terms of Overall Accuracy, Actual Accuracy (AA) and Kappa metrics most of the time.