Novel spectral redundancy suppression framework for high-speed multispectral image compression

Deep learning-based methods have improved multispectral image compression by addressing spatial redundancy, but managing spectral redundancy remains a challenge, limiting practical applications due to reduced downlink capacity and high computational complexity. To overcome these, this paper proposes NSRF-MIC (Novel Spectral Redundancy Suppression Framework for High-Speed Multispectral Image Compression), which integrates a Dynamic Spectral Channel Aggregation (DSCA) module and a Context Modeling with Windmill-shaped Non-uniform Channel (CMWNC) for joint spatial and spectral redundancy modeling. The DSCA uses adaptive fusion to capture complementary features, while the CMWNC employs a windmill-shaped attention mask to enhance entropy modeling and local dependencies. The method, validated on Landsat-8, Sentinel-2, and Gaofen-1 datasets, improves compression performance, maintaining efficient encoding, decoding, and computational speed.