Feature Mixture Generative Adversarial Network for Data Augmentation on Small Sample Hyperspectral Data

With the development of remote sensing technology, remote sensing data has been widely used in agriculture, medicine, military, and other fields. However, due to the disadvantages of the high cost of data collection and high redundancy, regression experiments using remote sensing data have serious overfitting problems. It limits its application in practical work. To alleviate this problem, we propose a generative adversarial network to generate remote sensing signals. In this paper, a feature mixing module was proposed to reduce the bias of the discriminator for different signals, thereby increasing the diversity of generated data. At the same time, spectral normalization is utilized to improve the stability during generation, which makes the generated data closer to the real signal. After a series of ablation experiments on small-sample remote sensing data, it is proved that the data generated by the generative adversarial network significantly improves the diversity of data and effectively alleviates the over-fitting problem based on ensuring the reliability of the data.