Deep learning-driven fine-grained classification and traceability of illicit pollution sources in urban stormwater networks using 3D-EEM images

Illicit wastewater discharge into stormwater pipes poses a severe ecological risk to receiving waters, yet conventional parallel factor analysis lacks the requisite precision to distinguish spectrally overlapping domestic wastewaters (e.g., office buildings, hotels) due to their intricate composition and overlapping spectral features. This study introduces an end-to-end deep learning framework integrating three-dimensional excitation-emission matrix fluorescence spectroscopy with a hybrid Swin Transformer-ResNet architecture. Through systematic comparison of eight convolutional neural network architectures and subsequent optimization, the refined model delivers superior classification performance with 98.21% accuracy. Gradient-weighted Class Activation Mapping reveals that the model focuses on diagnostically relevant spectral regions (Ex/Em ≈ 225-375/300-475 nm), suggesting potential for reducing the EEM scanning range, although instrument-level benefits require further validation. Within the studied region, the framework generalizes to complex multi-source mixtures without retraining, achieving exact, partial, and type match accuracies of 79.59%, 88.89%, and 98.41%, respectively. T-distributed stochastic neighbor embedding highlights overlapping category pairs, such as Convenience store and Supermarket, indicating cases requiring additional field verification. This work establishes a robust intelligent screening and progressive tracing tool for narrowing pollution sources in drainage networks, reducing blind field investigation and enabling precision rehabilitation of illicit connections to mitigate hazardous contaminant loads to aquatic ecosystems.