在过去几年里,卷积神经网络已经在高光谱图像分类上取得良好的效果,然而高光谱图像的高维性和卷积神经网络对所有波段的平等处理,限制了这些方法性能。本文提出了一种端到端的光谱空间注意力双边网络(SSABN),直接将原始图像3D块作为输入数据,而不需要进行预处理。首先,通过光谱空间注意力模块从原始数据中增强有用波段,抑制无效波段。然后设计双边网络两条路径。其中,空间路径用于提取空间信息,上下文路径用于提供更大的感受野,并通过特征融合模块有效的结合特征。实验结果表明,SSABN在三个公开数据集上取得了更高的分类精度,同时有效的减少训练时间。

In the past few years, convolutional neural networks have demonstrated good performance in hyperspectral image (HSI) classification. However, due to the high dimensionality of HSI and the equal treatment of all bands by convolutional neural networks, the performance of these methods is limited. In this paper, we propose an end-to-end Spectral-Spatial Attention Bilateral Network (SSABN), which uses the original 3D blocks as input data without preprocessing. First, the spectral-spatial attention module is used for enhance the useful band from the original data and suppress the invalid band. Then bilateral network is designed. The spatial path is used to extract spatial information, and the context path is used to provide a larger receptive field. The feature fusion module is used to effectively combine features. Experimental results demonstrate that the SSABN has achieved superior classification accuracy on the three data sets of Indian Pines, Pavia University, and Salinas, while effectively reducing training time.