SpectralMAE: Spectral Masked Autoencoder for Hyperspectral Remote Sensing Image Reconstruction

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral imaging is crucial for feature identification but faces spatial-spectral resolution trade-offs.
• Hardware solutions are costly and environment-dependent.
• Existing methods struggle with data limitations and generalizability, especially with missing or contaminated spectral bands.

Purpose of the Study:

• To introduce SpectralMAE, a novel spectral reconstruction model.
• To address the limitations of fixed mapping and data-hungry approaches in spectral reconstruction.
• To improve the utilization of arbitrary input band combinations for hyperspectral data.

Main Methods:

• Developed SpectralMAE, a masked autoencoder architecture for spectral dimensions.
• Employed a self-supervised learning paradigm for spectral reconstruction.
• Implemented a combined random masking pre-training and fixed masking fine-tuning strategy.

Main Results:

• SpectralMAE demonstrated superior performance over state-of-the-art methods.
• Achieved improved qualitative and quantitative metrics across five remote sensing datasets.
• Showcased enhanced generalizability and data utilization capabilities.

Conclusions:

• SpectralMAE offers a robust and efficient solution for hyperspectral reconstruction.
• The self-supervised masked autoencoder approach overcomes limitations of previous methods.
• This model advances the field of spectral reconstruction for remote sensing applications.