Generative adversarial network based data augmentation for hyperspectral images target detection using convolutional neural networks

Hyperspectral images acquired by remote sensing systems have high ability for target detection due to providing valuable spectral data in a wide range of wavelengths. In addition to spectral information, these images have valuable spatial information that is effective in accurate discrimination of target from background. One of the main challenges of target detection is the limited number of target training samples. To deal with this challenge, a detector based on Convolutional Neural Network (CNN) is introduced in this paper. Instead of using the single pixels, the cubic patches with spatial information in two dimensions and spectral information in the third dimension are used as the detector input. CNN with high capability in extracting hierarchical spatial features achieves good accuracy in the output. In order to solve the problem of the limited number of training samples, the generative adversarial network (GAN) has been used to generate fake cubic patches similar to the considered cubic patches around the real target pixels. The effect of data augmentation with GAN has been evaluated in the CNN detector with different amounts of data augmentation in both cases of full dimension and reduced dimension. The experimental results show the higher accuracy of CNN compared to the widely used detectors, as well as accuracy improvement of the CNN detector using data augmentation implemented by GAN.