Unified Model Compression Framework for Hyperspectral Image Super-Resolution

Abstract

Hyperspectral image super-resolution (HSI-SR) is vital for fine-grained Earth observation but remains impractical on resource-constrained airborne and spaceborne platforms. While existing methods prioritize reconstruction fidelity and degradation robustness, they largely neglect hardware efficient. To address this gap, we propose the first unified compression framework for HSI-SR, featuring two hardware-friendly quantizers. For inference, the blocked base-vector quantizer (BBVQ) exploits the strong local spatial–spectral similarity inherent in HSIs by representing each spatial block with a shared spectral base and low-bit residuals. This enables nearly integer-only arithmetic while preserving spectral fidelity. For backward propagation, the channel-aware scale-adaptive quantizer (CA-SAQ) reduces quantization error of gradients under aggressive bit-width compression by dynamically rescaling channels through integer bit-shifts, which provide a hardware-efficient alternative to fine-grained quantization. This design maintains gradient accuracy without introducing memory rearrangement. Both quantizers are plug-and-play and compatible with mainstream HSI-SR models. With optimized Field-Programmable Gate Array (FPGA) and graphics processing uni (GPU) kernels, our framework achieves 2.27× faster execution on GPU and 54.4% lower resource usage on FPGA at 8-bit precision, matching the spectral quality and degradation robustness of full-precision models, which will promote practical on-orbit deployment.