Spectral indicator development using excitation–emission matrix fluorescence and deep learning for quantifying organic pollution in mixed land-use watersheds

• Full-spectrum EEM images used to estimate source-specific pollution indicators in real-world watersheds. • Deep learning enables robust discrimination of overlapping organic pollution sources based on spectral complexity. • Model predictions align with spatial patterns and independent environmental data. • Framework supports scalable, data-driven water quality assessment, management and policymaking. Achieving reliable quantification of individual organic pollution sources remains a persistent challenge in mixed land-use watersheds, where multiple sources often co-occur and interact in complex, nonlinear ways. Conventional statistical approaches, which rely on a limited set of fluorescence indices or chemical tracers, are insufficient to resolve the spectral overlaps and intricate source mixing that characterize these environments. In this study, we present a novel, data-driven framework that leverages full high-dimensional information contained in Excitation–Emission Matrix (EEM) fluorescence images to directly estimate the proportional contributions of multiple organic pollution sources. River water samples and representative and representative source materials (e.g., soil, vegetation, livestock excreta) were collected from a mixed land-use watershed to construct a comprehensive EEM image dataset. By integrating these full-spectrum EEM images with a deep learning-based analytical system, we achieved robust classification and quantitative estimation of pollution source contributions in riverine samples. The model attained a weighted F1-score of 0.91 for source classification, which balances precision and recall across classes by accounting for sample size, and a mean absolute error of 5.62% for source contribution estimation. Predicted source contributions closely matched spatial patterns observed in the watershed, demonstrating the practical reliability of the approach for identifying major contributors. These deep learning-derived source contribution estimates serve as source-specific pollution indicators that capture nonlinear mixing patterns and address key limitations of conventional index- or tracer-based approaches. These indicators reflect the relative influence of diverse land-use origins on organic pollution, offering a scalable and interpretable decision-support tool for water quality assessment, management, and mitigation in heterogeneous watershed systems.