A deep learning model based on multi-scale self-attention mechanism and 3D EEM fluorescence spectroscopy for water...

Assessing water quality is crucial for public health and environmental sustainability. While excitation-emission matrix (EEM) spectroscopy provides "fingerprint" of dissolved organic matter (DOM), its high-dimensional data challenges analysis and limits traditional pollution source apportionment. A deep learning model integrating 3D EEM spectroscopy with a multi-scale self-attention mechanism to classify five water sources: fish pond aquaculture effluent, groundwater, surface water, industrial wastewater, and simulated wastewater. the model was developed through systematic architecture optimization and multi-scale fluorescence regional integration (FRI) feature engineering (window sizes 25-100 nm). Notably, multi-scale analysis revealed a "dimension paradox" where the intermediate 50 nm feature scale consistently outperformed the higher-resolution 25 nm scale, challenging the conventional assumption that higher spectral resolution invariably leads to better model performance. The self-attention model using a 50 nm window achieved an optimal performance-complexity trade-off with a test accuracy of 88.89 %. Key findings include the identification of Ex = 300-350 nm/Em = 250-300 nm as the top-ranking characteristic for surface water, Ex = 300-350 nm/Em = 300-350 nm as the highest-weight region for groundwater, fish pond effluent, and industrial wastewater, while Ex = 450-500 nm/Em = 450-500 nm emerged as a distinctive feature for groundwater classification. This approach transforms the model from a "black box" into an interpretable scientific tool, offering an efficient method for water quality monitoring and pollution source apportionment.