A GAN based method for cross-scene classification of hyperspectral scenes captured by different sensors

Abstract

Labeling samples in hyperspectral images is time-consuming and labor-intensive. Domain adaptation methods seek to address this challenge by transferring the knowledge from a labeled source domain to an unlabeled target domain, enabling classification with minimal or no labeled samples in the target domain. This is achieved by mitigating the domain shift caused by differences in sensing conditions. However, most of the existing works implement domain adaptation techniques on homogeneous hyperspectral data where both source and target are acquired by the same sensor and contain an equal number of spectral bands. The present paper proposes an end-to-end network, Generative Adversarial Network for Heterogeneous Domain Adaptation (GANHDA), capable of handling domain adaptation between target and source scenes captured by different sensors with varying spectral and spatial resolutions, resulting in non-equivalent data representations across domains. GANHDA leverages adversarial training, a bi-classifier, variational autoencoders, and graph regularization to transfer high-level conceptual knowledge from the source to the target domain, aiming for improved classification performance. This approach is applied to two heterogeneous hyperspectral datasets, namely RPaviaU-DPaviaC and EHangzhou-RPaviaHR. All source labels are used for training, while only 5 pixels per class from the target are used for training. The results are promising and we achieved an overall accuracy of 90.16% for RPaviaU-DPaviaC and 99.12% for EHangzhou-RPaviaHR, outperforming previous methods. Our code Implementation can be found at https://github.com/amirmah/HSI_GANHDA.