ATLAS: Hyperspectral Image Compression via Adaptive Transfer of Low-rank Abundance Splatting

LoR-SGS showed that 2D Gaussian splatting can compress hyperspectral images (HSI) by factorizing the cube into shared endmember spectra and low-rank abundance maps, so each Gaussian stores $K$ abundance coefficients instead of $\lambda$ bands. However, it does not employ densification, learned quantization, or training strategies that RGB splatting codecs use. We conduct a transfer study of nine such techniques in the HSI domain. Five transfer: LSQ+ quantization of abundances and covariances, distortion-driven densification, two-stage quantization-aware training, learnable endmembers, and gradient-guided initialization. Four fail: content-aware filters, position quantization, progressive quantization, and spatial delta coding; the failures show that HSI abundance maps are spatially smooth, low-rank fields in which quantization error is amplified multiplicatively through spectral band coupling. The optimized pipeline, ATLAS, beats published LoR-SGS on 6 of 8 rate-distortion targets ($+1.20$~dB average PSNR), and a scaling study on scenes up to $2372 \times 2196 \times 128$ reaches 40.1~dB at 0.034 bits per pixel per band on a single GPU without tiling. Decoded cubes match or exceed original-cube classification accuracy at 106:1 to 280:1, and classifying the bitstream's abundance maps directly matches full-spectrum accuracy without reconstructing the cube.