Improving deep hyperspectral image classification performance with spectral unmixing

Recent advances in neural networks have made great progress the hyperspectral image (HSI) classification. However, overfitting effect, which is mainly caused by complicated model structure and small training set, remains a major concern. Reducing complexity of could prevent to some extent, but also declines networks’ ability express more abstract features. Enlarging set difficult, for high expense acquisition manual labeling. In this paper, we propose an abundance-based multi-HSI classification method. Firstly, convert every HSI from spectral domain abundance dataset-specific autoencoder. Secondly, representations multiple HSIs are collected form enlarged dataset. Lastly, train classifier employ predict over all involved datasets. Different spectra that usually highly mixed, features representative reduced dimension with less noise. This benefits proposed method simple classifiers data, expect issues. The effectiveness verified ablation study comparative experiments.