Hyperspectral image shadow compensation via cycle-consistent adversarial networks

Illumination variance and shadows are challenging problems in remote sensing and hyperspectral imaging applications. Shadow compensation can effectively enhance the accuracy of object detection and material classification. Most shadow compensation methods either require preprocessing to detect the shadow region, or extra knowledge collected from additional sensors. Supervised deep learning based methods require paired samples to train the network. To overcome these restrictions, this work proposes an effective cycle-consistent adversarial network for shadow compensation (SC-CycleGAN). This unsupervised method is able to automatically transfer spectra in shadow region to their nonshadow counterparts, without requiring paired training samples and the step of shadow detection. The superiority of the proposed scheme is confirmed with both laboratory-created labeled data and real airborne data.