Assessment of river condition at a large spatial scale using predictive models

1. RIVPACS-type predictive models were developed at a relatively large spatial scale for the Australian state called New South Wales (NSW, 801 428 km²). Aquatic macroinvertebrate samples and physical and chemical data were collected from 250 reference sites (little affected by human activities) and 23 test sites (with known human impacts) throughout NSW in autumn and spring 1995 and identified mostly to family level. Reference sites were grouped based on their macroinvertebrate data using classification (UPGMA) and ordination techniques. Relationships between macroinvertebrate and environmental data were established using principal axis correlations and stepwise multiple discriminant function analysis. models for predicting invertebrate assemblages were developed separately for edge and riffle habitats for autumn and spring data sets and for combined autumn and spring data sets. 2. Sites in the lowland sections of the western flowing rivers were characterized by low taxonomic richness and were distinct from the sites in the eastern part of the state. Sites on the western slopes of the Great Dividing Range in southern and northern NSW mostly fell into separate groups. In eastern NSW, site groups did not represent a north, central and south division. Sites on highland streams, coastal fringe streams and large rivers mostly formed distinct groups, but most of the sites on east-flowing rivers fell into large site groups that did not have clear geographic boundaries. 3. Environmental variables that were strongly correlated with ordinations of macroinvertebrate presence/absence at sites were water temperature, altitude, longitude and maximum distance from source. The predictor variables determined by DFA for the six models created included alkalinity, altitude, location (longitude and/or latitude), stream size and substratum composition. These are generally in common with the variables determined for other large geographic areas in Australia and the United Kingdom. 4. Model outputs from reference sites suggest that, among the six models, the riffle model combining autumn and spring is likely to give the most reliable predictions. The combined edge model also performed well but refinements are needed for single season models to provide reliable outputs. 5. Combined season models both for riffles and for edges detected biological impairment at all but one of the test sites. Single season riffle models also detected impairment while single season edge models characterized sites as unimpaired despite other models’ indications of impaired fauna. Riffle models may be more sensitive than edge models but the sampling of riffles is often limited by flow. Edge habitats are available at most sites but there may be few riffles in floodplain rivers. Available resources, desired model sensitivity, and river type should be considered jointly to determine the most useful habitat to sample.