Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
4377939 | Ecological Modelling | 2008 | 13 Pages |
The present study compared prediction ability, transferability and sensitivity of the three artificial intelligence (AI) based models of fuzzy neural network (FNN), fuzzy habitat preference model (FHPM), and patterns of preference level (PPL), and one conventional model of habitat suitability index (HSI) in predicting spatial distribution of Japanese medaka (Oryzias latipes) dwelling in agricultural canals in Japan. Through the analyses, characteristics and applicability of these models were clarified. Based on field surveys, two independent sets of data were prepared; one was for calibration and the other was for validation. The models were first developed and tested by using the calibration data and then the transferability was verified by the validation data. All the models except for HSI were developed under 50 different initial conditions. Subsequently, sensitivity analysis was carried out to all the models in order to compare the model structure and to show how different data sets may affect the result of habitat prediction among the models, in which different levels of perturbations were given to input data. As a result, all the AI-based models appeared to have better prediction ability than the conventional model of HSI but lacked transferability. Among all, FNN was found to have the best predictive power despite of the high sensitivity which reflects the differences in its model structure under different initial conditions. FHPM appeared to have the best description in habitat preference as appeared as good convergence in preference curves among the AI-based models. PPL showed better prediction than FHPM in calibration but worse in validation and larger variances in model structure, resulting in deviation in sensitivity. HSI represented qualitatively the same habitat preference as the other models of FHPM and PPL, of which uniform habitat preference curves in a certain habitat category appeared as low sensitivity. The present result supports the use of AI techniques and their hybridization in predicting spatial distribution of the fish. Consequently, it is also suggested to take into consideration the mathematical characteristics of the habitat preference models, since they play the most important role in habitat evaluation and prediction.