Iterative Sample Generation and Balance Approach for Improving Hyperspectral Remote Sensing Imagery Classification with Deep Learning Network

Sample augmentation is effective for improving the supervised performance of land-cover classification with hyperspectral remote sensed image (HRSI) when the training samples are limited. However, numerous existing methods have neglected, considering the interclass-imbalance problem in the process of sample augmentation. In this work, new sample generation and sample balance strategies were promoted and simultaneously combined into an iteration for balancing and improving classification performance with HRSI. First, a sample augmentation with superpixel's constraint (SASC) is designed to augment the initial training samples set to avoid the overfitting of a sample generation neural network. Second, sample generation based on generative adversarial network (SGGAN) was proposed to generate samples for each class. Then, the proposed SASC, SGGAN, and a pattern recognition neural network named 3 dimensions-convolutional neural network (3-D-CNN) are combined into an iterative classification process called iterative sample generation and balance (ISGB) for balancing the user's accuracy for each class and optimizing the classification performance. Experiments on four widely used HRSIs are performed. The results when compared with eight state-of-the-art methods based on few-shot learning and generative adversarial network (GAN) efficiently demonstrate the feasibility and superiorities of the proposed approach for improving land-cover classification performance when the initial samples are limited. Moreover, the comparisons of the standard deviation of the user's accuracies (SDUA) demonstrated the balancing ability of the proposed approach. The code of the proposed approach is available at https://github.com/ImgSciGroup/ISGBA.