Habitat suitability modelling of rare species using Bayesian networks: Model evaluation under limited data
Health, Engineering and Science
School of Natural Sciences
Paucity of data on rare species is a common problem, preventing the use of most approaches to model development and evaluation. This study demonstrates how models can be developed and different forms of evaluation can be performed despite a lack of sufficient data, by presenting a habitat suitability model for the rare Astacopsis gouldi, the giant freshwater crayfish. We use a Bayesian network approach that readily incorporates incomplete data and allows for the evaluation of uncertainties. To supplement the limited field data on A. gouldi, expert knowledge was elicited through surveys designed to provide probability values that described the strength of relationships between the habitat suitability of the species and three variables - elevation, upstream riparian condition and geomorphic condition - and credible intervals around those values. A series of 18 alternative models were developed based on the same model structure but parameterised using different sources - expert judgement, field data or a combination of the two. The models were evaluated by estimating and comparing their performance accuracy and sensitivity analysis results, and in assessing the assumptions underpinning each of the models. Using performance accuracy as a measure, the data-based and combined expert- and data-based models performed better than the expert-based models. The sensitivity analysis results show that geomorphic condition was the most influential variable in the majority of models and that elevation had minimal influence on the occurrence of A. gouldi. Overall the models were found to have large predictive uncertainties, although the modelling process itself revealed insights into the habitat suitability of the species and identified key knowledge and data gaps for future monitoring, management and research.