Forecasting fish distribution along stream networks: brown trout (<i><scp>S</scp>almo trutta</i>) in <scp>E</scp>urope

Abstract

Abstract Aim Species inhabiting fresh waters are severely affected by climate change and other anthropogenic stressors. Effective management and conservation plans require advances in the accuracy and reliability of species distribution forecasts. Here, we forecast distribution shifts of S almo trutta based on environmental predictors and examine the effect of using different statistical techniques and varying geographical extents on the performance and extrapolation of the models obtained. Location Watercourses of E bro, E lbe and D anube river basins ( c . 1,041,000 km 2 ; Mediterranean and temperate climates, Europe). Methods The occurrence of S . trutta and variables of climate, land cover and stream topography were assigned to stream reaches. Data obtained were used to build correlative species distribution models ( SDM s) and forecasts for future decades (2020s, 2050s and 2080s) under the A1b emissions scenario, using four statistical techniques ( generalised linear models, generalised additive models, random forest, and multivariate adaptive regression ). Results The SDM s showed an excellent performance. Climate was a better predictor than stream topography, while land cover characteristics were not necessary to improve performance. Forecasts predict the distribution of S . trutta to become increasingly restricted over time. The geographical extent of data had a weak impact on model performance and gain/loss values, but better species response curves were generated using data from all three basins collectively. By 2080, 64% of the stream reaches sampled will be unsuitable habitats for S . trutta , with E lbe basin being the most affected, and virtually no new habitats will be gained in any basin. Main conclusions More reliable predictions are obtained when the geographical data used for modelling approximate the environmental range where the species is present. Future research incorporating both correlative and mechanistic approaches may increase robustness and accuracy of predictions.