Learned Hyperspectral Image Compression With Coarse-to-Fine Spatial–Spectral Prediction Network

Abstract

While deep-learned hyperspectral image (HSI) compression has achieved remarkable progress, existing methods typically encode latent representations indiscriminately. This entangled paradigm fails to separate global structural priors from local spectral details, thereby bottlenecking the overall spectral fidelity. To address this issue, we propose a coarse-to-fine spatial-spectral prediction network (CFSP-Net) that strictly decouples HSI reconstruction into two progressive stages: structural estimation and spatial–spectral refinement. Specifically, a lightweight multiscale pooling network first extracts low-frequency priors to provide stable global semantic guidance. Subsequently, a full-resolution fine prediction network (FPN) progressively restores high-frequency local textures and complex spectral correlations via a novel spatial–spectral refinement block, which integrates window-based spatial aggregation and hierarchical cumulative modeling. Furthermore, a fine-grained residual quantization strategy and a band-group optimization scheme are introduced to synergistically stabilize high-dimensional training and refine bit allocation. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of CFSP-Net, especially on structurally complex HSI scenes and low-to-medium bitrate conditions. Notably, on the Chikusei dataset, the proposed method achieves a peak signal-to-noise ratio (PSNR) of 52.69 dB and a spectral angle mapper (SAM) error of 0.0300 at 0.283 bpppb, improving the PSNR by approximately 0.45 dB over the strongest competing method at a comparable bitrate while maintaining superior spectral fidelity.