4.4 Sensitivity Analysis for Surface Runoff Under the VIC Model
In this section, we conducted the sensitivity analysis for the surface runoff predictions using the three–layer hierarchical framework (i.e., GGES, GCMs, and the VIC model parameters). The Xinanjiang model was not implemented because it cannot provide the spatial distribution of surface runoff. Therefore, the model uncertainty does not exist in this segment of the research. The sensitivity indices for the annual and monthly surface runoff and annual peak surface runoff were calculated at each grid point over the study domain using Eq. (8).
4.4.1 Sensitivity Analysis for Surface Runoff at Annual and Monthly Scales
Figure 10 displays the annual time series (2021–2050) of the sensitivity indices of three uncertainty sources (GGESs, GCMs and VIC model parameters) to the annual average runoff projections. The first impression of Figure 10 is that the variation of sensitivity indices of GCMs and model parameters is relatively even, while the sensitivity indices of GGESs show a large variability during the study period. The hydrological parameters and GCMs are the main contributor of uncertainty in the surface runoff projections, accounting for 48.15% and 46.53% of the total uncertainty, respectively, which is similar to that for the discharge projection uncertainty. The GGESs are still the least important source of uncertainty (0.32%–29.68%), and the sensitivity indices tend to show a decreasing trend over the study period. The variability of total variance of annual runoff depth is shown in Figure 10d, the fluctuation of hydrological parameters is relatively low, whereas the fluctuation of GGESs and GCMs is strong.
Figure 11 shows the intra–annual variability of sensitivity indices of three uncertainty sources to runoff projections. Results demonstrate that the sensitivity indices caused by GCMs increase from January to June and decrease from July to December, with the largest uncertainty in June (86.6%) and lowest uncertainty in January (74.3%). In contrast, the opposite trend patterns are identified for the sensitivity indices of VIC model parameters. The largest contribution of the VIC model parameters to the runoff uncertainty is 25.0% in January, while the smallest contribution is 12.75% in June. The uncertainty due to GGESs contributes higher (lower) uncertainty in winter and spring (summer and autumn). The total variance shows the strong intra–annual variability, which is similar to that of discharge based on the four–layer hierarchical framework (Figure 7e).
The sensitivity indices for surface runoff depth predictions were calculated for all the grid cells over the study catchment and were evaluated at the seasonal scales (i.e., spring, summer, autumn and winter). Figure 12 displays the spatial distributions of the sensitivity indices of GCMs, GGESs, and VIC model parameters. As shown, the main impression is that the sensitivity indices of surface runoff projections are uneven at both spatial and temporal scales. The GCMs and GGESs still show the most and least important uncertainty contributions to the runoff depth predictions, respectively. Particularly, the contribution of GCMs is in the range of 76%–92% and tends to be larger during spring and summer than autumn and winter for most of the study areas, probably due to larger rainfall in spring and summer. The contribution of GGESs is generally large in spring and winter (0.28%–0.67%, especially for eastern and southern regions) and tends to be smaller in summer and autumn (0.18%–0.56%, especially for western regions). The uncertainty caused by the hydrologic parameter is larger in the northwest and northeast regions at all seasonal scales. Particularly for the central regions, the uncertainty contribution from the hydrologic parameters is comparatively lower in spring and summer (13%–16%) and higher in autumn and winter (19%–21%).