Spectral-spatial classification for hyperspectral image based on a single GRU

Abstract Deep learning methods have been successfully used to extract deep features of many hyperspectral tasks. Multiple neural networks have been introduced in the classification of hyperspectral images, such as convolutional neural network (CNN) and recurrent neural network (RNN). In this study, we offer a different perspective on addressing the hyperspectral pixel-level classification task. Most existing methods utilize complex models for this task, but the efficiency of these methods is often ignored. Based on this observation, we propose an effective tiny model for spectral-spatial classification on hyperspectral images based on a single gate recurrent unit (GRU). In our approach, the core GRU can learn spectral correlation within a whole spectrum input, and the spatial information can be fused as the initial hidden state of the GRU. By this way, spectral and spatial features are calculated and expanded together in a single GRU. By comparing the different utilization patterns of RNN with a variety of spatial information fusion methods, our approach demonstrates a competitive advantage in both accuracy and efficiency.