Modelling habitat requirements of bullhead (Cottus gobio) in Alpine streams

In the context of water resources planning and management, the prediction of fish distribution related to habitat characteristics is fundamental for the definition of environmental flows and habitat restoration measures. In particular, threatened and endemic fish species should be the targets of biodiversity safeguard and wildlife conservation actions. The recently developed meso-scale habitat model (MesoHABSIM) can provide solutions in this sense by using multivariate statistical techniques to predict fish species distribution and to define habitat suitability criteria. In this research, Random Forests (RF) and Logistic Regressions (LR) models were used to predict the distribution of bullhead (Cottus gobio) as a function of habitat conditions. In ten reference streams of the Alps (NW Italy), 95 mesohabitats were sampled for hydro-morphologic and biological parameters, and RF and LR were used to distinguish between absence/presence and presence/abundance of fish. The obtained models were compared on the basis of their performances (model accuracy, sensitivity, specificity, Cohen's kappa and area under ROC curve). Results indicate that RF outperformed LR, for both absence/presence (RF: 84 % accuracy, k = 0.58 and AUC = 0.88; LR: 78 % accuracy, k = 0.54 and AUC = 0.85) and presence/abundance models (RF: 79 % accuracy, k = 0.57 and AUC = 0.87; LR: 69 % accuracy, k = 0.43 and AUC = 0.81). The most important variables, selected in each model, are discussed and compared to the available literature. Lastly, results from models' application in regulated sites are presented to show the possible use of RF in predicting habitat availability for fish in Alpine streams