Modelling freshwater ecological systems under limited data and knowledge

Abstract

Freshwater ecological systems throughout the world are extensively altered, with threatening processes being physical, chemical and biological. Data is often lacking to describe the processes within the system and their interactions. This thesis explores how modelling can be used to help understand and manage such complex systems under limited data and knowledge, by examining two Australian case studies. The first case study explores how fish species are distributed in the Murray-Darling Basin (MDB), Australia's largest river system, and attempts to understand the variations in fish assemblages across the Basin. The second case study involves the development and evaluation of a habitat suitability model for an endangered species Astacopsis gouldi, the Giant Tasmanian Crayfish. The MDB fish case study adopted a suite of multivariate analysis techniques, including cluster analysis and non-metric multidimensional scaling, which were applied within the framework of a data-mining process, Knowledge Discovery from Data. This is the first known study to describe the large-scale patterns in fish assemblages in the Basin. The analysis identified regions in the Basin sharing similar fish assemblages; these patterns were generally consistent with the temperature, precipitation and elevation of the streams and the temperature and water quality preferences of the species. The delineation of the Basin into regions can be an important basis for future ecological studies, for example by providing context for finer-scale studies or a benchmark for studies examining temporal changes to the fish communities. Bayesian networks were applied to model A. gouldi habitat suitability in the second case study, with a focus on the evaluation process. A series of 18 models were built and tested, based on the same structure but with their parameters estimated from different combinations of expert opinion and training datasets. The data-based A. gouldi habitat suitability models achieved better performance accuracy than the expert models. The combined data- and expert-based models performed equally well as the data-based models but were considered more robust. The model evaluation process revealed interesting insights into the habitat suitability of the species, including that elevation seemed to have little influence on habitat suitability, contrary to other studies. It was also found that the species is not in equilibrium with its environment, suggesting that future models must consider temporal dynamics and avoid using species presence-absence as an indicator of habitat suitability. Both case studies highlighted the subjective nature of modelling and the importance of good modelling practice in producing meaningful and purposeful outputs. Good practice guidelines are provided for Bayesian network modelling, with emphasis on the need of thorough evaluation of the model and its results, and transparent reporting of the modelling process. For models built under limited data and knowledge, it is especially important to acknowledge and embrace the uncertainties. Modelling cannot overcome a lack of data, but can help to integrate information, develop and test hypotheses, and refine knowledge about the system and its processes. Modelling can also help guide future monitoring and research toward data and knowledge gaps most crucial to our understanding of the system. Show more