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.