A fully convolutional network with channel and spatial attention for hyperspectral image classification

Deep learning method has achieved great success in the hyperspectral image (HSI) classification. However, the existing deep learning methods with local HSI patch as input can only use the local spatial-spectral information to determine the class of samples. In this paper, the fully convolutional network framework is proposed to classify the whole HSI scene. The whole HSI is taken as input of the deep network, and then the classification result of the whole HSI scene is output. In this way, there is no need to slice the original data cube, and the processing efficiency is greatly improved. In order to make better use of the global context information, we introduce channel and spatial attention model to enhance the discriminability of features. The classification experiments on two real HSI datasets show that the proposed method could achieve higher classification accuracy than the patch-based deep learning method, and the training and testing speed are faster.

Automated summary

A new deep learning method has been proposed for successful classification of hyperspectral images, utilizing a fully convolutional network framework with whole HSI input to achieve faster training and testing speeds. The method also incorporates channel and spatial attention models to enhance feature discriminability and improve classification accuracy over patch-based deep learning methods.