Full-spectrum EEM end-member unmixing with statistical validation: A deep learning framework for organic pollution...

Accurately apportioning organic pollution sources in mixed land-use watersheds remains challenging due to the limited tracer capacity of conventional fluorescence-based approaches. Traditional methods condense multidimensional excitation-emission matrix (EEM) spectra into a few summarized indices (e.g., BIX, HIX, FI), creating a mismatch between available fluorescence tracers and the number of pollution sources, thereby limiting source discrimination when multiple sources overlap. This study presents an analytical framework for source apportionment integrating three elements. First, full-spectrum EEM images are used as source-specific end-members to preserve spectral information for source-pattern characterization. Second, a statistical assessment of end-member independence and separability performed using SSIM and ANOSIM. Third, a CNN-unmixing architecture was developed to learn source-discriminating patterns from full-spectrum EEM images and estimate source-associated contributions in mixed-source EEMs based on candidate end-member representations. An EEM dataset was constructed from pollution source samples collected in mixed land-use watersheds. Statistical assessment of end-member similarity and separability showed high within-group similarity (Pearson's r = 0.882-0.988; SSIM = 0.838-0.970) and significant between-group separation among the eight end-members in EEM spectral space (ANOSIM R = 0.876; PERMANOVA R² = 0.899). The framework was evaluated using ground-truth mixtures (2-5 components) and field stream samples. In controlled mixtures, CNN-unmixing showed R² > 0.7 and MAE <0.07 and generally reproduced dominant-source contributions and rankings across mixture complexities. The field application in the study watershed showed reasonable spatial consistency between the predicted dominant-source contributions and observed pollution-load distribution patterns, supporting the preliminary field-scale applicability of the framework for source-priority assessment in mixed land-use watersheds.