Low-Rank Spectral Continuity Modeling With Gaussian Splatting for Hyperspectral Image Compression

Lossy compression has become the mainstream approach for downlinking hyperspectral images (HSIs). However, most learning-based methods rely on implicit mappings that fail to preserve inherent structural priors of HSIs, while existing explicit methods lack end-to-end optimization and spectral continuity modeling. To address these issues, a low-rank spectral continuity modeling with a Gaussian splatting framework (LR-GS) for HSI compression is proposed. Singular value decomposition (SVD) and rank-revealing QR (RRQR) construct a compact low-rank spectral representation with band selection and stable spectral basis vector estimation. Representation coefficients are modeled as continuous functions using weighted Gaussian primitives. A 3-D total variation (3DTV) constraint is further introduced in the rate–distortion optimization framework to regularize spatial–spectral gradients and suppress artifacts. Experiments show superior performance at different bitrates. The proposed method achieves 36 dB PSNR, improving results by 22% over JPEG2000 and 5% over LoR-SGS at 0.2 bits per pixel per band (bpppb).