Fast Decomposition of Single Excitation–Emission Matrix Fluorescence Spectrum via Encoder–Decoder Model

Three–dimensional excitation–emission matrix (3D–EEM) fluorescence spectroscopy is widely applied for the rapid characterization of dissolved organic matter (DOM) in aquatic environments. However, conventional decomposition based on parallel factor analysis (PARAFAC) requires multiple spectra and manual intervention, limiting its applicability for rapid analysis and future online implementation. The purpose of this study is to develop an efficient data–driven method capable of decomposing fluorescence components from a single 3D–EEM spectrum. We propose a conditional single–spectrum decomposition network (CSSD–Net) based on the encoder–decoder model. The encoder extracts fluorescence features from the input spectrum, while the decoder combines these features with conditional information on component count to generate up to five component maps. The component count can be automatically predicted by CSSD–Net or manually specified to support flexible application scenarios. CSSD–Net was trained using publicly available component spectra from the OpenFluor database without PARAFAC preprocessing. Validation on natural water samples demonstrates that the results obtained from CSSD–Net using a single sample are highly consistent with those from PARAFAC using multiple parallel samples, with a mean Tucker’s congruence coefficient (TCC) of 0.9615. These results show that CSSD–Net provides a fast and practical solution for decomposing single 3D–EEM spectra under constrained aquatic scenarios, and it has potential for future near–real–time and in situ applications.