Transformer-based lossy hyperspectral satellite data compression

Hyperspectral satellite sensors generate vast amounts of data, making efficient compression crucial for storage, transmission, and downstream analysis. We propose SpectralNet-X, a hybrid convolution–transformer autoencoder that operates purely in the spectral domain. A 1D convolutional projection captures local spectral smoothness, while transformer layers with Rotary Positional Embeddings model long-range dependencies. A small set of learnable queries performs cross-attention pooling, yielding a compact latent space that serves as the compression bottleneck. To improve stability and reconstruction accuracy, we pretrain the encoder using a masked spectral reconstruction objective before fine-tuning the full autoencoder. Experiments on PRISMA hyperspectral data demonstrate that SpectralNet-X consistently outperforms a pure transformer baseline (HyCoT) and achieves competitive spectral fidelity, although a convolutional baseline (A1D-CAE) remains superior. These findings highlight the potential of hybrid CNN–transformer architectures for hyperspectral compression and motivate future research on scaling data, refining loss functions, and extending evaluation.