Isolating the water quality responses of multiple land uses from stream monitoring data through model calibration

Abstract

Summary Quantifying relationships between stream water quality and catchment land uses is a major goal of many water quality monitoring programs. This is a challenging task that is rarely achieved through simple analysis of raw data alone. Multiple regression analysis provides one approach, which despite significant limitations, can be successful when very large data sets are available and only annual estimates are required. However, regression techniques have limited application to sub-annual data sets. We present a new method for isolating the water quality responses of different land uses from monitoring data through hydrological model calibration, using a process of simultaneous calibration at several monitoring sites. In addition to model parameters, model algorithm complexity and the number of land-attribute groups are also used as calibration ‘parameters’. This helps increase model parameter uniqueness and model predictive certainty. We applied the technique to water quality data from the Johnstone River catchment (1602 km 2 ) in north-east Australia, using the HSPF model. The data comprised >4000 samples from over five years of monitoring at 16 sites, which drained sub-catchments of differing land area and proportions of each land use. Monitoring occurred at flow gauging sites during high stream flows, and regularly at all sites during non-event periods. Variables modelled included discharge, suspended sediment, and various forms of nitrogen and phosphorus. The calibration process aimed to maximise both goodness-of-fit and parameter sensitivity. We achieved a substantial simplification of HSPF algorithms without appreciable reduction in goodness-of-fit, by a combination of: fixing parameters, tying parameters, and introducing new, simpler equations. Two key calibration tools were reducing the number of land-use groups (by combining land uses) and tying parameters between the three flow paths modelled (surface flow, interflow and base flow). These in turn removed insensitive parameters, increased the sensitivity of remaining parameters, and increased model predictive power. For calibration of suspended sediment, we found our conceptual model was too complex for the information content of the data, so it was necessary to group certain land uses and simplify algorithms in HSPF. In contrast, for nitrate, our conceptual model was too simple, with successful calibration requiring a finer resolution of land use in the catchment. A major advantage of our technique is that it can isolate land-use effects from stream monitoring data on a modelling time step that is unachievable by regression analysis, with a sensitivity that enables detection of relatively small differences in time and/or space. The technique also enhances confidence in estimates, it identifies ‘sensitive’ and ‘insensitive’ parameters and processes, and it provides a robust calibration by reducing the parameter solution space. Further, by using a hydrological model the technique can account for spatial and temporal variation in land use and rainfall. It is also predictive, so it can be used to assess impacts of land-use-change scenarios within the catchment. Additionally, it can provide insights into the type and amount of data required to isolate land-use effects on different water quality parameters, which can be used to enhance the effectiveness and cost-efficiency of the monitoring program design.