Identification of Mixed Organic Pollutants in River Water Based on Graph Convolutional Networks

In the application of three-dimensional (3-D) fluorescence spectroscopy for river water pollutant detection, traditional feature extraction models often neglect wavelength information and complex intermolecular interrelationships. Moreover, experiments are frequently limited to single-substance classifications. This research addresses these limitations by proposing a novel model for organic pollutant recognition based on graph neural networks (GNNs). The proposed method begins by using the excitation-emission matrix and the wavelength information corresponding to each fluorescence intensity value as inputs. A convolutional neural network (CNN) is initially employed to extract preliminary features from the three-dimensional fluorescence spectra. Subsequently, the initial spectral features obtained during the pre-training stage are used to construct a graph neural network model. This model represents each category of organic pollutants as nodes within the graph, capturing the intrinsic connections between different pollution samples through node interactions. Finally, the model outputs a score vector for each category to classify the pollutants. To validate the proposed model, datasets comprising twelve common organic pollutants, both as single substances and mixtures, were collected. Comparative analysis with other methods demonstrates that the proposed approach achieves a high accuracy rate.