5 Discussion
In this study, we developed an advanced hierarchical sensitivity analysis framework for quantifying the relative contribution of the uncertainty sources to the hydrological projections at both temporal and spatial scales. The four–layer hierarchical framework considering four uncertainty sources (i.e., GGES, GCMs, hydrological models, and model parameters) and the three–layer hierarchical framework considering three uncertainty sources (i.e., GGES, GCMs, and hydrological model parameters) were comprehensively tested in a humid subtropical basin in southern China. Compared with the qualitative comparison method (Chen et al., 2011b, 2013; Dobler et al., 2012) and the ANOVA method (Aryal et al., 2019; Bosshard et al., 2013; Vetter et al., 2015, 2017), the framework presented in this study is capable of grouping different model uncertainty sources and considering the dependence relationships among uncertainty inputs as well as the spatio–temporal variations of uncertainty sources. More importantly, this framework can theoretically be applied to the quantitative analysis of n (n ≥ 2) kinds of uncertain sources in the context of climate change impact studies.
The results highlighted strong temporal and spatial variability of general sources of uncertainty in hydrological predictions, and indicated that GCM structure is one of the largest uncertainty sources, consistent with the previous findings by the CMIP3 and CMIP5 models (Bosshard et al., 2013; Chen et al., 2011b; Déqué et al., 2007; Hattermann et al., 2018; Kay et al., 2009; Prudhomme and Davies, 2009; Su et al., 2017; Wilby and Harris, 2006). In addition, we notice that the contribution of hydrological parameters to uncertainty is also significant and can be larger than that of GCMs at the interannual scale (more than 50%, Figures 5 and 7). This suggests that the influences of the hydrological parameters on the hydrological projections should be considered, especially for the long–time projections. Interestingly, the importance of uncertainty caused by hydrological model parameters was comparably low reported by Chen et al. (2011b) due to the small difference between parameter sets based on 10 times of calibration results. In contrast, the 20 parameter sets were considered in this study to randomly sample within a bounded sensitive parameter space, resulting in comparably large uncertainty of hydrological parameters, especially for the output of annual peak discharge (Figure 9). As expected, the GGESs is the smallest contributor of hydrological projections, but the uncertainty of GGESs tends to show large variability over the projection periods (e.g., Figure 13), a consistent finding with the projections of extreme precipitation events (Wada et al., 2013; Xu et al., 2019). The contribution of GGESs to uncertainty varies significantly in different river basins (Kay et al., 2009; Vetter et al., 2017). Kay et al. (2009) compared the effect of different sources of uncertainty on flood frequency in two catchments of England, which found that uncertainty due to emissions is very low for one catchment, but more important for another catchment. And Vetter et al. (2017) compared the contributions of multiple sources of uncertainty (emission scenarios, GCMs, and hydrological models) in five basins and conducted the conclusion about the large differences of the uncertainty caused by emission scenarios between basins. We also found that the sensitivity indices for discharge and surface runoff on monthly scale tend to be periodical (Figure 7 and Figure 11). The uncertainty caused by GCMs tends to be higher in summer than in winter, whereas the uncertainty due to GGESs, hydrological model and parameters is higher in winter than in summer. In terms of future work, the intra–annual variability of various uncertainty sources as well as their dependence relationships in hydrological projections needs to be further explored to better explain this phenomenon.
It should be noted that the results shown are subject to a few limitations. First, only one downscaling technique was applied in this study. It is widely accepted that the uncertainty caused by multi–downscaling methods is potentially large (Bosshard et al., 2013). Second, the sensitive hydrological model parameters were only randomly sampling, although the effects of insensitive parameters to model outputs are minimal. Third, the number of parametric samples may be insufficient, although the use of the LHS method can effectively reduce the number of samples. Therefore, more sources of uncertainty (e.g., downscaling technique), abundant model parameters and sufficient samples of simulations need to be considered within the hierarchical sensitivity analysis framework to better understand the spatio–temporal variations of general sources of uncertainty in hydrological predictions.