Classification of Hyperspectral Imagery Using Spectral-Spatial Residual Attention Network

Due to the rapid progress in data-driven computational imaging, hyperspectral image classification has become the center for attention. Hyperspectral image as a three dimensional image consists of 2D spatial information and 1D spectral bands. To classify hyperspectral imagery, it is essential to explore robust representation learning solutions. Convolutional Neural Networks have reported marvelous results on various image classification problem. However, convolutional-based methods are not able to capture sufficient contextual information and ignores the potential of long-term memory. Existing data-driven hyperspectral image classification methods give priority of each pixel equally. Thus, in this paper, we proposed a residual attention network capable of learning both short-term and long-term dependencies focusing on the relevant part of hyperspectral imagery. Our featured solution demonstrates the superiority of various information processing leveraging pair-wise self-attention mechanism. Extensive experiments on University of Pavia shows that our proposed attention-aware deep learning method outperforms other data-driven strategies in terms of various metrics.