Satellite Remote Sensing-Implemented Nontargeted Screening of Emerging Contaminant Fingerprints in a River-to-Ocean Continuum through Interpretable Machine Learning: The Pivotal Intermediary Role of Dissolved Organic Matter

Abstract

Emerging contaminants (ECs) can exert irreversible health impacts on humans, even at trace concentrations. Currently, nontargeted screening of ECs has been developed for their assessment, which requires sophisticated instrumentation. Although satellite remote sensing is a cost-effective technology for water quality assessment, accurately measuring ECs in a river-to-ocean continuum remains a significant challenge due to their trace levels. To address this challenge, we innovate a strategy utilizing satellite remote sensing to achieve high-resolution nontargeted EC screening. By employing DOM as an intermediary variable, bridging the gap between satellite remote sensing and ECs in river-to-ocean continua. DOM, including the total sum of ECs, reflects their distribution and spectral sensitivity, enabling satellite sensing to capture their unique fingerprints. In this study, this strategy has enhanced the accuracy of nontargeted EC screening from 32.2 to 95.7% using machine learning. Interpretable machine learning causal inference and SHAP models reveal that shortwave infrared (SWIR) S2-B11 is crucial for EC screening while emphasizing the importance of avoiding multicollinearity with similar SWIR band S2-B12. Additionally, the band reflectance is influenced by the proportion of polarity-related heterogeneity in the ECs. Furthermore, we developed a real-time remote sensing surveillance system featuring interactive maps for nontargeted screening of ECs and GPT-based contamination interpretation.