Spectral Constrained Generative Adversarial Network for Hyperspectral Compression

Lossy compression exhibits remarkable capabilities in handling large volumes of data. However, information loss can affect spectral characteristics and spatial information to various degrees during hyperspectral compression. Therefore, it is essential to restrict the range of spectral changes, as each spectral curve corresponds to a distinct semantic context. In this article, we propose a spectral-constrained generative adversarial network (SCGAN) for hyperspectral compression. Specifically, SCGAN integrates compression and classification tasks within a unified framework. During generative adversarial learning, SCGAN uses a classification map to guide the generation of global spectral information. To deal with different hyperspectral images (HSIs) in one model, a three-stage training strategy is leveraged. Experiments conducted on three public HSI datasets illustrate that the proposed SCGAN effectively narrows the semantic gap. For instance, the average overall classification accuracy of SCGAN on Pavia University is above 0.9, which is the closest to the classification accuracy achieved with the original HSIs, even at a low bitrate of 0.05 bits per pixel per band.