An approach to model selection when predicting water quality in NSW using geospatial predictors

The NSW Office of Water, within the Department of Trade and Investment, Regional Infrastructure and Services is developing water quality guidelines for regions within New South Wales as part of the state’s implementation of the National Water Quality Management Strategy (NWQMS). The new guidelines will be tailored more closely to specific catchments and regions than the current default guidelines in NSW, the Australian Water quality Guidelines for Fresh and Marine Waters, National Water Quality Management Strategy (ANZECC & ARMCANZ, 2000). They will also be used to inform various national and state natural resource management targets. Water quality guidelines typically include reference values that indicate what the “best case” water quality values are for a region: these reference values are usually derived from reference sites, which have been purposefully chosen because they provide the best example of undisturbed conditions within a catchment. Reference sites were not available for this project, hence the Office of Water has used a predictive modeling approach to underpin the development of water quality guidelines by developing statistical estimates of reference condition. The predictive models have drawn on the last 10 years of water quality and in-stream flow records, and associated geospatial information from catchments across NSW. The project has concentrated on building predictive models for five common water quality variables: turbidity, electrical conductivity, water temperature, total nitrogen and total phosphorus. Water quality is affected by both natural and anthropogenic factors. That is, rainfall and any landscape features not influenced by human behavior versus all land use types, distance to upstream dams and all variables related to vegetation type or cover. Flow variables were categorised separately because they are influenced by both natural and anthropogenic activities. It is possible that current targets for some sites may never be able to be met due to the impact of natural factors. One of the aims of this research is to identify if that is ever possible and, if so, under what conditions. This research will also assist in determining which regions will most benefit by targeted activities to reduce the impact of human behavior on waterways. An earlier pilot study established that natural and discrete groupings could be formed based on different water quality characteristics alone and that sets of geospatial factors associated with a water quality monitoring station’s drainage can be used to explain the water quality characteristics of that station. The current research has built on the pilot study, refining quality control procedures and increasing the scope and number of monitoring stations, whilst the water quality variables of interest remain as total nitrogen, turbidity, total phosphorus, electrical conductivity and water temperature. The range of geospatial variables, although fine-tuned, has still remained a significant number, viz. 72. This paper will discuss the approach taken to reduce the possible number of geospatial predictors to a number acceptable for a robust prediction of a data series, how the time series of each water quality variable at each water quality monitoring station was summarised to allow investigation of the impact of the geospatial predictor variables, how these predictors were then used to build separate models to predict each of the water quality variables at each water quality monitoring station based on the subset of important geospatial predictors for the corresponding water quality variable and how the models were validated. The resulting predictive models and the estimation of undisturbed water quality values are not discussed in this paper, but will be addressed in future publications.