Hyperspectral image classification based on multi-scale information compensation

By virtue of their effective capabilities of deep feature extraction, deep learning methods, especially convolutional neural network (CNN) algorithms have exhibited significant potential for hyperspectral image (HSI) classification. However, in information propagation, the amount of semantic information constantly increases while detailed information that representing small-scale local features of object vanishes. In addition, the scale singularity of the extracted feature restricts improvement in the accuracy of classification. In this letter, a multiscale CNN-based module called information compensation is proposed that can maintain detailed information while preserving semantic information by integrating the original input with more abstract hierarchical learning feature maps. An end-to-end network based on this module that takes advantage of spatial information is proposed for HSI classification. Furthermore, in the foundation of this network, spectral information is considered to develop a spatial-spectral hybrid framework. Experimental results on two-public HSI datasets show that the proposed networks achieve competitive performance compared with other state-of-the-art CNN-based methods.