The Utility of Land-Surface Model Simulations to Provide Drought Information in a Water Management Context Using Global and Local Forcing Datasets

SASER (SAfran-Surfex-Eaudysee-Rapid) is a hydrometeorological modeling platform based on the SURFEX land-surface model. Usually, SASER uses SAFRAN as the forcing dataset (Section 3.1) but, in this paper, it can use any other suitable dataset. SURFEX (SURFace EXternalisée) is Météo France’s surface modeling platform (Masson et al. 2013). For natural soils, SURFEX uses the ISBA (Interaction Sol-Biosphère-Atmosphère) land surface scheme (Noilhan and Planton 1989; Noilhan and Mahfouf 1996), which has different versions. In this study, we have selected ISBA-3L (Boone et al. 1999) and ISBA-DIF (Boone 2000; Habets et al. 2003). The former considers a simple three-layer description of the soil by means of a force restore approach, while the latter uses a more complex multilayer approach. SURFEX lacks a routing scheme, which is the reason why SASER uses the RAPID routing model (David et al. 2011a, 2011b). The connection between SURFEX and RAPID is made by means of Eau-dysée.¹ An important limitation of SASER is that it does not represent groundwater processes.

In this study, SURFEX has been run on the same grid as SAFRAN, which has a resolution of 5km. Simulations have been performed using ISBA-DIF and ISBA-3L. Hereafter, the SURFEX simulations that use ISBA-DIF are labeled as DIF, and the SURFEX simulations that use ISBA-3L are labeled as 3L.

LEAF (Land-Ecosystem-Atmosphere Feedback) is the land-surface component of RAMS (Regional Atmosphere Modeling System) (Walko et al. 2000). Several modifications were made to LEAF to incorporate groundwater processes, resulting in LEAFHYDRO (Miguez-Macho et al. 2007). The water table depth is diagnosed based on the soil moisture when it lies within the resolved soil layers. When the depth is deeper, soil columns are extended to the dynamic water table below, thereby acting as saturation boundary conditions and affecting the soil water fluxes above. There is lateral groundwater flow among adjacent cells, leading to divergence from the high ground and convergence with low valleys at multiple scales. The water table, once recharged by rain events or fed by lateral flow convergence, relaxes into rivers within a grid cell through fluxes between groundwater and rivers. These fluxes can be 2-way depending on the hydraulic gradient, representing both loosing (i.e., leaking into groundwater) and gaining (i.e., receiving groundwater) streams. River discharge, which is fed by surface runoff and groundwater, is routed out to the ocean through the channel network using the kinematic wave method. The sea level is set as the head boundary condition for groundwater, which allows it to influence coastal drainage.

The LEAFHYDRO simulations for this study were performed on a grid with a 2.5 km resolution that is nested in the 5 km SAFRAN grid, which allows each cell in the SAFRAN grid to exactly encompass 4 cells of the LEAFHYDRO grid. Afterwards, the results were aggregated to 5 km to facilitate the comparisons with the SASER-based simulations. Hereafter, LEAFHYDRO is labeled as LHD.

We have calculated drought indices following the SPI. This index is calculated using monthly data or, alternatively, a time series of the accumulated precipitation for the previous n months, where n represents the scale of the index. A similar operation can be done for soil moisture (SSMI) and streamflow (SSI). We use a nonparametric methodology, which can be applied to different climatic variables, including precipitation, soil moisture and relative humidity, without having to assume representative parametric distributions (Farahmand and AghaKouchak 2015).

The methodology we used to study drought propagation is inspired by Barker et al. (2015): (1) the standardize soil moisture index (SSMI) is calculated with a temporal accumulation of 1 month, making the SSMI equivalent to SSMI-1; (2) for precipitation, SPI-n is computed for n in [1,2,...,24]; and (3) finally, for any given grid point, the n_x scale maximizing the Pearson correlation coefficient between SPI-n and SSMI is found. The resulting n_x can be interpreted as the scale at which drought propagates from precipitation anomalies to soil moisture anomalies. For streamflow, the methodology is mostly the same, with the difference being that the SPI is calculated with the areal mean of the basin precipitation corresponding to the gauging station of the studied streamflow.

In this section, we quantify the model result differences and assess the intermodel consistency for soil moisture drought. Table 2 shows the root mean square difference (RMSD) and Pearson correlation (r) calculated by comparing all simulations with each other. These values synthesize both the spatial and temporal aspects of the differences, as the comparisons include all data points (i.e., all grid points for all time steps). The table contains four subtables. The first row of subtables corresponds to the root zone, and the second row corresponds to the deep soil. Different color scales have been used for each score (RMSD and r), but the same color scale (where greener is better) is used for each score for both soil layers.

There is more uncertainty (i.e., redder) in the deep soil than the root zone results, which is expected, as the considered LSMs use different approaches to simulate deep soil water fluxes.

The disparity between simulations with the same model but different forcings is larger than that between simulations with the same forcing but different models, which implies that the forcing dataset plays an important role. For example, in simulations forced by SFR, the largest RMSD for the root zone soil moisture is 0.55, and the lowest correlation is 0.84 (both correspond to SFR-DIF vs. SFR-LHD). On the other hand, for all simulations using DIF, the largest RMSD is 0.61, and the lowest correlation is 0.80 (E2O-DIF vs. SFR-DIF). This result is identical if we compare E2O-LHD with SFR-LHD.

For the root zone soil moisture, the smallest difference between two simulations that use different forcing datasets and models is between E2O-DIF and SFR-3L (RMSD = 0.42 and r = 0.9). The largest difference is between E2O-LHD and SFR-DIF (RMSD = 0.74 and r = 0.71). To put these results into context, a RMSD of 0.74 is almost \(\frac {3}{4}\) of a standard deviation, which means that the drought status of a grid point at a given time probably changes category. The trend found in E2O increases the differences between simulations. If E2O is removed from the comparison, the simulations with the largest differences are MSW-LHD and SFR-DIF (RMSD = 0.65, r = 0.78), which are still high values. For deep soil moisture, the differences are larger; the most divergent simulations (taking E2O out of the comparison) are MSW-DIF and SFR-3L (RMSD = 0.87 and r = 0.6). When including E2O, the largest difference is between E2O-DIF and SFR-3L (RMSD = 0.96 and r = 0.51).

This analysis shows that drought studies performed in the same area with different models and forcing datasets should be compared carefully, as the uncertainties are still high. This is something that, for example, has also been found in a climate change setting by Marx et al. (2018).

If the drought status is important for a water manager, the time scales of drought propagation (from precipitation to soil moisture) are perhaps even more important, as their knowledge can be a tool for the forecasting of the soil moisture status in places where drought dynamics are relatively slow, which could benefit early warning systems. As stated in Section 4.2.2 we perform this analysis by finding the n_xscale maximizing the correlation between SPI-n and SSMI.

Figure 2 shows maps of n_x for the root-zone soil moisture. For instance, by taking SFR-DIF (a) and SFR-LHD (b), it is apparent that the spatial pattern of n_x is significantly different for each model. The spatial structure in DIF is overly homogeneous, which means that for this model, the SSMI correlates better with the SPI at scales of 2 months for almost all of the study area. In contrast, the LHD presents a richer pattern, with scales that extend from 3 to 12 months, as the dynamics are slower towards the south and faster towards the northwest. 3L is similar to DIF but with a NW-SE gradient that extends from scales of 1 month in the NW to 9 months in the SE. The conclusion of this first comparison is that the model formulations have a strong influence on the temporal scale at which the variability of precipitation affects the variability in soil moisture; thus, different models yield significantly different results. Another important question is how meteorological forcing affects n_x. The first column in Fig. 2 (panels a, d, f, and h), showing the n_x results for all simulations that use the DIF model, shows that the pattern is largely invariant. E2O-DIF and MSW-DIF have an area where n_x is one month longer than that in SFR-DIF, but this difference is significantly lower than that from the changing model. The results for the LHD model (second column; panels b, e and g) indicate that the spatial patterns of n_x for different forcings are more divergent than those in the case of DIF but still to a lesser degree than those of a different model. This suggests that, for some models, the forcing dataset has an impact on how precipitation variability propagates to soil moisture; however, this impact is still not as high as that of the model formulation itself.

A similar analysis can be performed for the deeper soil zone (below the root zone). In this case, (figures not shown), n_x is higher for all models except 3L, indicating that soil moisture variability in the deeper soil is generally related to more persistent precipitation anomalies, which is expected. For DIF, n_x is much larger in the deep soil than in the root zone when compared with the LHD results, where there is much less variation. This occurs perhaps because of the relatively fine resolution of the soil columns in the top 4 m of the soil and the coupling with the water table in this model.

These comparisons show that LSMs with different structures produce different drought dynamics in soil moisture. Unfortunately, we do not know what the reality is due to the lack of observations, but these large differences reveal that progress is needed in order to, first, improve the observational coverage of the soil moisture (particularly the deeper layers) and, second, bring models closer to the real behavior. Furthermore, the forcing dataset also introduces its own uncertainty.

Here, we evaluate the ability of the models to simulate streamflow drought (SSI). In this case, we compare the ability directly with that of the SMP-SMP, assuming that it provides a good estimate of natural flows. Figure 3 shows the RMSDs for the different simulations, with lower values indicating higher skills. SFR-DIF generally has the smallest RMSDs, ranging from 0.6 to 0.8, with maximum values between 0.8 and 1.0 in an area at the headwaters of the Tajo, Guadiana, and Júcar, which also includes the southernmost stations in the Ebro river basin. SFR-3L presents similar results, whereas the SFR-LHD exhibits larger RMSDs at most stations, with a maximum value of approximately 1.2. The aforementioned high baseflows in this model are likely the cause of these large RMSDs, which impact its capability to reproduce drought status. To put all of these results in the context of drought studies, an RMSD value of 1 represents one standard deviation of the index and, thus, even errors of 0.6, which is the best case for all simulations and may change the drought status of a given monthly flow (i.e., from normal to drought status). Concerning the effect of forcing, the maps show that the MSW has similar results to the SFR, while E2O, in general, yields somewhat larger RMSDs.

Figure 4 depicts the Pearson correlation coefficient (r) of the same simulations, with higher values representing better performances. The SFR-DIF is, again, the simulation with the best scores, followed by SFR-3L and SFR-LHD. In the last case, the correlations in some rivers are extremely low, especially in the southern half of the area of study, where summer flows are minimal and, thus, negatively affected by the high baseflow bias in the LHD. The impact of changing the forcing datasets remains low, as it was before.

The conclusion of this comparison is that, in terms of the SSI, model formulation is the most important uncertainty source compared to the forcing dataset.

Finally, we evaluate how the models simulate drought propagation from precipitation to streamflow. We have studied the correlation between the SPI-n and SSI for different values of n (not shown). We found that, in general, the SMP-SMP has a behavior closer to the DIF and 3L behaviors compared with the observations in many basins. This means that even a well-tested hydrological model, such as the SMP, has difficulties producing the right correlations between the SPI-n and SSI. For its part, the LHD presents a different behavior due to the connection of rivers to the groundwater in this model. In all cases, the simulations are closer to the other simulations when using the same model and not the same forcing dataset, which indicates that model formulation largely determines the drought propagation behavior, and the forcing dataset has a reduced impact.

We now generalize for the whole study area. Figure 5 shows the scale n_x, in months, of the SPI-n that better correlates with the SSI. The SFR-OBS (b) compares basin precipitation from the SFR and observed SSI at each station. There are many red dots in areas where the flow is highly intervened by water management, but other reasons cannot be excluded, such as karsts or groundwater influence, because some red dots correspond to rivers in the natural regime. The SMP-SMP (a) is the result of a well-tested hydrological model; however, we must keep in mind that it is not the ground truth, and it will not be able to correctly simulate the presence of large karst systems or other complex geological features; therefore, it does not have a perfect skill score, as discussed previously. The SMP-SMP shows some coherent geographical patterns, with low n_x values (1-3 months) in the north in most of the Ebro and in some areas of the Duero and Tajo basins. The highest values (reaching 1 year) are found in the Sistema Iberico mountain range, separating the Ebro and Duero basins. Moderate values occur in the SE and some areas of the Duero basin. With regard to the other models, The DIF and 3L present the lowest n_x, as in the case of the root-zone soil moisture, indicating the generally fast response of rivers to precipitation deficits. The comparison with the SMP shows that these n_x values are too low and too homogeneous in space. Forcing the DIF with the SLR instead of the SFR, MSW or E2O does not change the results significantly, suggesting that those results are mostly dependent on model formulation. Indeed, a model with a different approach, such as the LHD, presents a divergent behavior, with high values of n_x in many of the studied basins, which is likely due to the role of groundwater buffering the streamflow response to drought. Furthermore, it is more sensitive to forcing than the other models, as was the case for soil moisture.

Figure 6 shows the Pearson correlation between the SPI-n and SSI when n = n_x. If we compare the SFR-OBS with the SMP-SMP, we see that in the real world, which is complex and influenced by water management, the correlations between the SPI-n_x and SSI are lower than those in the non-human influenced and simplified world of a model. This result suggests that, in reality, there are many more factors driving streamflow variability, as defined by the SSI, than just precipitation. As expected, the DIF and 3L streamflow variabilities are too related to precipitation (i.e., high correlation), as they do not simulate management or groundwater processes; however, they are quite comparable to the SMP-SMP. The LHD, on the contrary, has much lower correlations, which means that the streamflow variability in this case is driven by other factors, such as groundwater processes.