Accurate hyperspectral remote sensing information is essential for feature identification and detection. Nevertheless, the hyperspectral imaging mechanism poses challenges in balancing the trade-off between spatial and spectral resolution. Hardware improvements are cost-intensive and depend on strict environmental conditions and extra equipment. Recent spectral imaging methods have attempted to directly reconstruct hyperspectral information from widely available multispectral images. However, fixed mapping approaches used in previous spectral reconstruction models limit their reconstruction quality and generalizability, especially dealing with missing or contaminated bands. Moreover, data-hungry issues plague increasingly complex data-driven spectral reconstruction methods. This paper proposes SpectralMAE, a novel spectral reconstruction model that can take arbitrary combinations of bands as input and improve the utilization of data sources. In contrast to previous spectral reconstruction techniques, SpectralMAE explores the application of a self-supervised learning paradigm and proposes a masked autoencoder architecture for spectral dimensions. To further enhance the performance for specific sensor inputs, we propose a training strategy by combining random masking pre-training and fixed masking fine-tuning. Empirical evaluations on five remote sensing datasets demonstrate that SpectralMAE outperforms state-of-the-art methods in both qualitative and quantitative metrics.
Keywords: hyperspectral imaging; masked autoencoder; self-supervised learning; spectral reconstruction; transformer.