Determination of Fluorescent Dissolved Organic Matter Using Three-Dimensional Fluorescence Spectroscopy and Convolutional Neural Networks

Fluorescent dissolved organic matter (FDOM) — particularly tryptophan (Trp), tyrosine (Tyr), and humic acid (HA) — serves as a crucial indicator in environmental monitoring. This study introduced a novel quantitative analysis approach for analyzing three-dimensional excitation–emission matrix spectra (3D-EEMs) of FDOM using convolutional neural networks (CNNs). The performance of the CNN model was evaluated and compared with the self-weighting alternating trilinear decomposition (SWATLD) algorithm. Results revealed that the proposed model significantly outperforms the SWATLD algorithm. Specifically, the CNN model achieved R 2 , RMSE, and MAPE values of 0.964, 0.047, and 14.950%, respectively, while for the SWATLD algorithm, these values were 0.944, 0.062, and 17.439%. Augmentation of the original spectral dataset did not yield a substantial improvement in the performance of the SWATLD algorithm, but it significantly enhanced the prediction ability of the CNN model. This enhancement was evident in the improved R 2 , RMSE, and MAPE values of 0.989, 0.030, and 12.837%, highlighting the critical role of data augmentation in boosting the performance of the CNN model, especially when dealing with a limited dataset. Application of the CNN model to water samples from Laizhou Bay yielded satisfactory results, enabling a simple and rapid analysis of FDOM in seawater. Therefore, an accurate and convenient analytical model was developed based on EEMs and CNNs, which can swiftly determine the concentration of FDOM in the environment and provide valuable references for environmental monitoring and early warning.