Few-shot domain-adaptive hyperspectral image denoising via dynamic quantile-guided learning

Hyperspectral imaging (HSI) captures full spectra at each pixel, enabling applications in environmental monitoring, industrial inspection, and biomedical research. Its effectiveness, however, is often limited by complex noise, and existing deep learning denoising methods require large-scale paired datasets, which are scarce due to HSI's high dimensionality and susceptibility to corruption. Here, we present FT-DSES, a few-shot domain-adaptive hyperspectral denoising framework designed to address data scarcity and cross-system variability common in optical imaging. The method integrates a specifically designed spectral-enhanced spatial attention module with an adaptive dynamic quantile pooling mechanism, enabling noise-aware spatial encoding and precise preservation of spectral signatures under varying illumination and noise statistics. A two-stage adaptation scheme selectively recalibrates quantile parameters, allowing rapid transfer to new instruments and modalities using only 5-8 paired samples. Trained on synthetic complex photon-noise models and minimally fine-tuned in the target domain, FT-DSES achieves high-fidelity reconstruction (up to ∼36.47 dB in PSRN and ∼0.15 in SAM) with reduced computational burden (∼61.42% parameter and ∼57.86% time). Experiments across real-world remote sensing, reflectance imaging, and fluorescence microscopy demonstrate significant gains in PSNR and SAM, demonstrating robust cross-source, cross-domain, and cross-modal generalization in various photon-limited and data-limited regimes.