Hierarchical CNN Classification of Hyperspectral Images Based on 3-D Attention Soft Augmentation

The distinction of similar classes has always been the core issue in image classification. In this article, a new hierarchical classification process based on three-dimensional attention soft augmentation (HC-3DAA) is proposed to improve the accuracy of classifiers, especially for the accuracy between similar classes. In HC-3DAA processing, the merging matrix is first constructed based on the validation confusion matrix to measure the similarity among different classes. Specifically, the 3-D attention soft augmentation module combined with CutMix is designed for guiding the network model to focus on the key discriminative features. Then, the extracted 3-D feature differences between similar classes are inserted into the attention module for the reclassification to get higher classification accuracy. To evaluate the performance of HC-3DAA, CutMix models with different feature dimensions and the HC module are separately discussed on two widely used hyperspectral datasets. Two different classifiers 3-D convolutional neural network and ResNet are included in the comparative analysis. Besides, experimental results also demonstrate that the proposed HC-3DAA outperforms several state-of-the-art attention-based methods.