Catchment grouping and regional calibration for predictions in ungauged basins

Abstract

Runoff prediction in ungauged regions remains a challenging task in hydrology (Sivapalan et al. 2003). The process of finding suitable parameter values to model runoff in ungauged catchments, by inferring and learning from model calibrations in gauged catchments, is generally referred to as 'regionalisation'. An increasing number of studies now use a regional calibration approach (Vogel 2005; Vaze et al. 2011a; 2013). The regional calibration finds one set of parameter values to represent the entire 'hydrologically similar' region rather than considering each catchment independently. In the regional calibration, the model parameters are optimised to produce an overall best simulation for all the gauged catchments within the region. Vaze et al. (2011a) suggested that the regional calibration has capability to accommodate extra local information by, e.g., using different sets of parameters to represent catchments with different vegetation or land use types across the target region. Vaze et al. (2013) also showed that the regional calibration approach has an advantage as it can incorporate information from new data sources, like remotely sensed vegetation, evapotranspiration and soil moisture, to improve model characterisation, reanalysis and predictions. While applying a regional calibration to a large region such as south-eastern Australia, all catchments cannot be assumed to have similar hydrological behaviour, and hence a sub-grouping approach based on differences in physical catchment characteristics is required. This can be facilitated by using remote sensing data which provides a complete coverage of a range of catchment characteristics. There are many catchment characteristics that have been used in past studies to measure the hydrologic similarities between catchments for model regionalisation (Oudin et al. 2008; Zhang and Chiew, 2009). But there still remains a question about which catchment characteristics are more informative to measure the hydrological similarity for hydrological response in a specified region. This study attempts to answer this question by undertaking modeling experiments by sub-grouping catchments based on three physical catchment and climate characteristics - the fraction of vegetation coverage (fPAR), aridity index (AI) and rainfall distribution over seasons (Seasonality). The hydrological model is regionally calibrated for each of the sub-groups and the calibrated model parameters are used to simulate runoff over different independent periods (model validation). The results are compared with those obtained from using sub-groups of randomly selected catchments. The results have also been compared against classic regionalisation based on spatial proximity, where all the 196 catchments are calibrated individually and the calibrated parameter set from the geographically closest catchment is used to simulate streamflow for 'ungauged' catchments.The experiments are carried out using data from 196 unregulated and unimpaired medium sized catchments (50-2000 km2) (Vaze et al., 2010) across south-eastern Australia for the period 2000-2008. The hydrological model used in this study is the 4-parameter GR4J Model (Perrin et al. 2003). The results show that the classic local calibration and validation approach perform best, and the regional calibration against the whole catchment set (only one parameter set for all catchments) gives the worst results. The results also show that daily runoff predictions from the catchment characteristics-based sub-grouping approaches perform substantially better than that from random sub-grouping. The results also indicate that predictions from sub-grouping based on both aridity index and the seasonal rainfall distribution are able to offer comparable performance to local regionalisation based on spatial proximity. Although the standards for sub-grouping and the number of sub-groups may be subjective, the results suggest that regional calibration with catchments sub-grouping based on catchment physical properties has potential to improve hydrological predictions over large regions.