Accurate quantification of components with spectral imaging (e.g., hyperspectral and Raman) is fundamentally challenged by nonlinear mixing effects. To overcome the limitations of model-specific traditional algorithms, we propose P4NSU, a deep learning framework for projection-based pretraining of nonlinear sparse unmixing. P4NSU introduces a synergistic two-part strategy: hierarchical pretraining that distills large spectral libraries into compact, task-specific subsets and a learnable projection that maps spectra into a feature space where unmixing simplifies to a linear problem. Extensive evaluations on three synthetic data sets, generated using established three nonlinear models, demonstrate that P4NSU consistently outperforms state-of-the-art linear and nonlinear methods, achieving a 14-51% reduction in overall RMSE. Its practical utility is further validated on real-world data sets. For hyperspectral imaging of pigments, P4NSU attained the highest unmixing accuracy, particularly reducing the RMSE for the most spectrally ambiguous chalk by more than 40%. For Raman imaging of leukemia cells, P4NSU produced biochemically meaningful abundance maps that aligned with known subtypes; notably, utilizing these maps as features significantly enhanced automated classification performance, underscoring its dual strengths in accurate quantification and information distillation. Implemented as an open-source Python toolkit (https://github.com/Ryan21wy/P4NSU), P4NSU offers a robust and practical solution to advance quantitative analysis in spectral imaging across diverse fields.
Wang et al. (Mon,) studied this question.