Combining machine learning with expert knowledge to classify anthropogenic pressures on Alpine rivers using benthic invertebrates

Streams and rivers are among the most threatened ecosystems globally, as they integrate processes and stressors across local and regional scales. Although research on multiple-stressor effects is growing rapidly, field assessments remain challenging due to complex interactions between natural and anthropogenic drivers. Combining expert knowledge with machine learning can support decision-making for classifying dynamic ecosystems, overcoming data-limitations. We aimed to assess how expert-based classifications of anthropogenic pressures match data-driven classification based on benthic macroinvertebrates taxonomic and functional composition by applying a Random Forest (RF) approach. We used data typically available at Environment Agencies in charge of implementing WFD requirements. Based on stressor-specific indices and local practitioners' expertise, 160 stream sites in Trentino (NE Italy) were classified according to hydrological, morphological, and chemical alterations, including pristine conditions. Most sites were a-priori classified as impacted by one or more stressors and 16% as unimpacted. RF classification matched the expert-based classification only partially, confirming macroinvertebrate sensitivity to hydro-morphological alterations. Functional traits were less informative than taxonomic features in discriminating between water quality alteration and pristine conditions, and in detecting pollution when additional hydrological stressors (e.g., reduced flow or hydropeaking) were present. However, both functional and taxonomic features reliably detected water pollution even where other hydro-morphological alterations occurred. These findings suggest that macroinvertebrate-based indicators commonly used to assess the ecological status of mountain streams capture overall waterbody stress but have limited ability to discriminate among different stressor types. The proposed approach can be replicated and upscaled to broader regions, offering valuable insights for the management of Alpine running waters. • Random Forest (RF) compares expert and data-driven stream alteration classifications. • RF used taxa, traits and environmental data as predictors as to classify stressors. • Six binary stressors comparisons reflect typical Alpine stream pressure conditions. • Model performance and key variables were assessed for each classification. • RF taxonomic models showed stronger effects than those based on functional traits.