LoR-SGS++: Residual spectral-spatial refinement for low-rank Gaussian hyperspectral image compression

Abstract

Hyperspectral imaging acquires rich scene information by capturing hundreds of contiguous bands across ultra-violet, visible, infrared, and other spectral ranges, making it valuable for remote sensing and material analysis. However, the resulting data volumes demand efficient compression to reduce storage and transmission costs while preserving reconstruction quality. Implicit neural representations (INRs) can represent HSIs compactly but struggle to model their spatial locality, often incurring high GPU memory cost and slow encoding. LoR-SGS addresses these limitations by coupling low-rank spectral decomposition with 2-D Gaussian splatting, yet two performance-critical bottlenecks remain. First, Gaussian rasterization introduces local coefficient inconsistency across neighboring pixels. Second, the fixed NMF-initialized spectral basis causes spectral mismatch that cannot be corrected during optimization. To address these two bottlenecks, we propose LoR-SGS++, which introduces lightweight, zero-initialized correction modules. Spatial Coefficient Refinement (SCR) applies a convolutional residual to the rasterized coefficient map to improve local spatial consistency, while Low-Rank Endmember Calibration (LEC) introduces a scene-adaptive low-rank update to the spectral basis. Both modules are gated by learnable scalars initialized to zero, so that the codec starts from the baseline behavior and activates corrections only as optimization requires. Evaluations on four benchmark datasets show that LoR-SGS++ achieves consistently strong performance in PSNR, MS-SSIM, and SAM across all tested scenes. At matched bitrates, the proposed method surpasses LoR-SGS by over 3 dB in PSNR on the most challenging scene while adding fewer than 0.001 M parameters.