A pre-training enhanced deep learning framework for robust sparse unmixing in chemical imaging

Chemical imaging techniques, such as hyperspectral and Raman imaging, provide spatially resolved spectral insights for non-destructive analysis of complex mixtures. However, accurately decomposing mixed pixel spectra into constituent components remains challenging due to spectral overlaps and variability. While deep learning models have shown promise in jointly estimating endmember spectra and abundances, the increased complexity and parameters make them highly sensitive to initialization, often resulting in unstable performance. Therefore, developing accurate and robust unmixing algorithms is critical for reliable chemical imaging analysis.

We propose a versatile pre-training framework to enhance sparse unmixing (P4SU) for chemical imaging. P4SU leverages simulated spectra generated from spectral libraries to pre-train deep learning models, followed by fine-tuning on target chemical imaging data. To accommodate a range of mixing scenarios, we equip the framework with both linear and nonlinear decoder options. Evaluated across three chemically diverse datasets-pigment mixtures, sugar solutions, and pharmaceutical tablets-P4SU demonstrates superior accuracy and stability compared to conventional methods (e.g., SUnSAL, CLSUnSAL) and non-pre-trained models. Significantly, pre-trained models reduce the root mean square error (RMSE) by 15-32 % compared to non-pre-trained ones and lower the standard deviation of results by 90-98 % on the pigment dataset.

The method's robustness to spectral variability and noise demonstrates its potential for rapid, reliable chemical analysis in quality control and material identification. Implemented as an open-source Python toolkit (available at https://github.com/Ryan21wy/P4SU), P4SU offers a practical solution to streamline analytical workflows in chemical imaging analysis.