Seeking functional homogeneity: A framework for definition and classification of fish assemblage types to support assessment tools on temperate reefs

•A method that uses fish-based metrics to define assemblage types is proposed.•Sites were clustered while taking into account natural and sampling variability.•Ensemble clustering was used to combine results from different metrics.•Three assemblage types were delimited and characterised.•A k-nearest neighbours classification model is proposed to classify new sites

Due to their important role in the ecosystem and high economic value, there is a need to assess the effect of anthropogenic impacts on marine fish assemblages. However, this can only be achieved if variations due to natural causes are known. Moreover, while most assessment tools rely on functional traits, bottom-up habitat classification frameworks tend to use species composition. The present study proposes an innovative framework to define fish assemblage types through metric pairwise constrained k-means (MPCK-means) clustering of sites based on functional guild categories and univariate metrics, an approach that takes into account within-site variability due to the sampling method and natural causes. This was followed by a label-based ensemble clustering approach, which finds patterns that minimise information loss when integrating clustering results from individual metrics. In order to test the method, fish assemblages on 14 nearshore rocky reefs along the Portuguese coast were sampled. The final typology configuration achieved through ensemble clustering consisted of three assemblage types and maintained an average normalised mutual information of 0.605 with the individual clustering results. Nested PERMANOVA found differences among types and the most variable metrics in the face of natural variation were identified. Ultimately, a k-nearest neighbours classifier is proposed to label new sites, based only on environmental variables that are unlikely to be directly affected by the presence of anthropogenic impacts. Optimal performance for the classification model was achieved with inverse distance-weighted voting of the 4 nearest neighbours with an average classification accuracy of 96.08%.