Abstract
The richness of native fish is considered to be an indicator of aquatic ecosystem health, and improving richness is a key goal in the management of river ecosystems. An artificial neural network (ANN) model based on field data from 90 sample sites distributed throughout the Júcar River Basin District was developed to predict the native fish species richness (NFSR). The Levenberg-Marquardt learning algorithm was used for model training. When constructing the model, we tried different numbers of neurons (hidden layers), compared different transfer functions, and tried different k values (from 3 to 10) in the k-fold cross-validation method. This process and the final selection of key variables with relevant ecological meaning support the reliability and robustness of the final ANN model. The partial derivatives method was applied to determine the relative importance of input environmental variables. The final ANN model combined variables describing riparian quality, water quality, and physical habitat and helped identify the primary drivers of the NFSR patterns in Mediterranean rivers. In the second part of the study, the model was used to evaluate the effectiveness of two restoration actions in the Júcar River: the removal of two abandoned weirs and the progressive increase in the proportion of riffles. The model indicated that the combination of these actions produced a rise in NFSR, which ultimately reached the maximum values observed in the reference site of that river ecotype (sensu the European Water Framework Directive). The results demonstrate the importance of longitudinal connectivity and riffle proportion for improving NFSR and the power of ANNs to help decisions in the management and ecological restoration of Mediterranean rivers. Furthermore, this model at the basin scale is the first step for further research on the effects of water scarcity and global change on Mediterranean fish communities. © 2012 Elsevier B.V.
