Mount Lofty Ranges (MLR) Water Quality

An Improved Water Quality Model for the Onkaparinga Catchment

This component of the research focused on three key pollutants, namely total suspended sediment (TSS), total nitrogen (TN) and total phosphorous (TP). Using statistical models, the processes that drive hydrology and water quality in the Onkaparinga catchment were studied and applied to land-use change scenario modelling. Specifically, this research focused on:

• Applying a Bayesian calibration approach to calibrate the SIMHYD rainfall-runoff model for use in the Onkaparinga catchment and quantify potential sources of uncertainty in the hydrology
• Developing statistical models (site based models) for sites monitored in the Onkaparinga catchment in the MLR watershed for the purpose of quantifying constituent loads with an estimation of uncertainty
• Using statistical models to investigate three scenarios of land-use change and whether there are changes in loads and the uncertainty around loads

Statistical models employed to address the above points, consisted of generalised additive models (GAMs) and generalised additive mixed models (GAMMs) through the Loads Regression Estimator (LRE) package that was developed for the quantification of loads for the Great Barrier Reef catchments (by CSIRO). Site based models for the six sites studied in the Onkaparinga catchment used a variety of hydrological variables as covariates for understanding the variation in the data measured for each site. Specifically, these hydrological variables included flow, decomposed into baseflow and runoff as well as flow discounting terms that took into account past characteristics of the hydrograph. This could consist of a total accumulation of flow from the start of sampling to the short-term flow record prior to the current constituent sampled. Models were fit using the LRE package using the R statistical programming language.

Three scenarios that were explored as part of this research consisted of:

• Investigating the sale of SA Water land holdings in Scott Creek sub-catchment
• Quantifying the impact of continued expansion of perennial horticulture in the Cox Creek sub-catchment.
• Quantifying the impact on water quality of infill within township boundaries of Aldgate Creek Railway Station.

These scenarios were determined at meetings with SA Water and SA EPA and were structured around the statistical modelling approach used to evaluate each scenario. A Random Forests modelling approach was used to develop a spatio-temporal model for each constituent across the six sites of interest in the Onkaparinga catchment. The model is non-parametric and popular in the machine learning and is based on decision tree methodology. The approach can take a large number of potential covariates as predictors to develop an ensemble of decision trees on bootstrap samples of the data. Variable importance rankings can assist in identifying important variables.

Developing the dSedNet plugin

To improve the parameterisation and methods used in Source for modelling constituents, the SedNet dynamic modelling capability was extended. This was primarily a software development exercise, using the Source model already developed for the catchments of the MLR. Three key activities were undertaken:

• The development of a spatial parameterisation tool to enable the rapid set-up of dSedNet models
• The development of two component modules of the dSedNet plugin – the hillslope and gully sediment generation models
• Trialling of the dSedNet plugin in the Onkaparinga catchment (Houlgraves Weir).

The development of the dSedNet plugin and spatial parameterisation tool was the main activity within the project. Testing of the plugin was performed by running a series of small modelling trials. In addition to testing the robustness of the code, these trials provided an opportunity to consider the advantages of the dSedNet approach.

The dSedNet software is being developed as a Source plugin; the spatial parameterisation tool has been developed within the core of Source platform. Plugins are easily distributed between modellers and models, while changes to Source core are more difficult to manage as they require quality assurance by eWater and need to fit in with Source development and release cycles. This is advantageous from a maintenance perspective, but does reduce the control that the project team has over its release. At this point, a release date is unknown. The South Australian project team has a version of Source that contains the spatial parameterisation tool, and the dSedNet plugin, and a training workshop was held in May 2015.

The MLR Source model was modified to use the dSedNet plugin, with some parameters changed to match those used in other studies (such as that by Wilkinson et al. in 2014). This is referred as the baseline, noting that it has not been calibrated against observed water quality data.

Additionally, two calibrations (multi- and single-region) of the rainfall-runoff model (SIMHYD) were used, giving two sets of results. The multi-region simulation used separate SIMHYD parameters for each region, according to the calibration of its gauge. The single-region calibration treated all sub-catchments as a single region, i.e. they all had the same SIMHYD parameters. The need to do this emerged during the trials to more closely match the hydrology to other studies and demonstrates the importance of using good hydrological models for constituent modelling.

The Onkaparinga catchment was chosen for the trials, using observations at Houlgraves Weir. Three trials were conducted against previous studies: (i) SedNet sediment budgets Wilkinson et al. (2005), (ii) existing EMC/DWC Source model, and (iii) statistically derived loads.

Simulations were run for the two rainfall-runoff calibrations (multi- and single-region), resulting in predicted (potential) loads of 15.9 and 19.5 kTonnes/year. This compared well with the 15.0 kTonnes/year predicted using SedNet and as reported in the Wilkinson et al. (2005) study, noting that this model was also uncalibrated. The trial against the EMC/DWC approach looked at temporal patterns and the role of parameter settings and the fact that dSedNet has many adjustable parameters that are not available in an EMC/DWC model. The trial against the statistically derived loads highlighted the power of dSedNet being able to incorporate spatial and temporal variation into model inputs. Simulated dSedNet loads tended to follow observed loads.