1- Department of Environmental Sciences, Faculty of Natural Resources, University of Guilan, P.O. Box 1144, Sowmeh Sara, Guilan, Iran
2- Department of Applied Ecology, Ghent University, J. Plateaustraat 22, B-9000 Gent
* Corresponding author?s E-mail: firstname.lastname@example.org
Classification trees (J48) were induced to predict the habitat requirements of tench (Tinca tinca). 306 datasets were used for the given fish during 8 years in the river basins in Flanders (Belgium). The input variables consisted of the structural-habitat (width, depth, gradient slope and distance from the source) and physic chemical (pH, dissolved oxygen, water temperature and electric conductivity), and the output ones were the abundance and presence/absence of tench. To find the best performance model, a three-fold cross validation was applied on the entire dataset. In order to evaluate the model stability, the dataset were remixed in 5 times, obtaining in total 15 different model training and validation events. The effect of pruning on the reliability and model complexity was tested in each subset. The performance evaluation was based on a combination of the number of Correctly Classified Instances (CCI) and Kappa statistic. The results showed that the predictive performance evaluation was suitable, confirming the reliability of classification trees methods. The overall average of CCI and Kappa for the prediction of tench was obtained 75.8% and 0.53. When analyzing the ecological relevance of classification trees, it seemed that the structural-habitat variables were important predictors compared to physic chemical variables.