Use of Artificial Neural Networks (ANNs) to assess xenobiotics in a river catchment using macroinvertebrates as...

The Danube flows through various European regions, exposing its aquatic ecosystem to multiple stressors, including dams, canalization, and agricultural activities. Fertilizers, manures, pesticides, animal husbandry activities, irrigation practices, deforestation, and urbanisation (e.g., industrial effluents and domestic waste) are the primary drivers of environmental change in the Danube catchment area. This study demonstrates the advantages of applying cutting-edge Machine Learning (ML) models to the Joint Danube Survey 3 (JDS 3) for detecting xenobiotics using reliable biomarkers. Macroinvertebrate communities, recognised as key indicators by the Water Framework Directive, serve as sensitive proxies for chemical pollution through their varied responses to stressors. We employed ML models (4-Layer Perceptron, Long Short-Term Memory, and Transformer Neural Networks) to precisely assess river ecological condition based on biological and chemical parameters. Machine learning analysis revealed significant correlations between specific pesticides (2,4-Dinitrophenol, Chloroxuron, Bromacil, Fluoranthene, and Bentazone) and the composition of the macroinvertebrate community in the Danube River basin. Among the tested models, Artificial Neural Networks emerged as the most effective approach. The Long Short-Term Memory models best captured relationships between 2,4-Dinitrophenol and Bentazone and the macroinvertebrate communities. The 4-Layer Perceptron model showed superior performance for 2,4-Dinitrophenol and Fluoranthene predictions, whereas Transformer Neural Networks outperformed others in modeling Bromacil and Fluoranthene dynamics. These results demonstrate that Artificial Neural Network architectures can reliably link chemical stressors to biological indicators with transferability potential to other lotic systems through tailored biological parameter inputs.