Changing the retention properties of catchments and their influence on runoff under climate change

The analytical form of Choudhury-Yang formulation of Budyko framework (equation (1)) used here contains a parameter n reflecting catchment water retention capability. Figure 1(a) shows the change in water retention capability (Δn) for 33 river basins during the period of 1981–2005. During this period, wet large catchments (i.e. Congo, Amazon, Mississippi, Yukon, Mackenzie, Volga and Yenisey basins) generally show decreases in n, indicating water retention capability in these basins has declined. That is the fraction of P lost by evapotranspiration decreased and instead runoff increased. In contrast, the values of n are found to increase dramatically in dry basins such as the Murray, Colorado, Yellow River and three river catchments in the southern Africa (figures 1(a) and S1 is available online at stacks.iop.org/ERL/13/094019/mmedia).

Zoom In Zoom Out Reset image size

Despite the simplicity of the Budyko framework, the observed magnitude and spatial variability of runoff changes can be well reproduced by this theoretical framework. Using measurement-based values of P, PET along with our fitted time dependent n values to estimate runoff, we obtain a small RMSE of 10.2 mm yr⁻¹ (figure 1(b), blue dots) against annual runoff measurements. This good fit of the sum of Budyko-estimated components (i.e. ∆Q_P + ∆Q_PET + ∆Q_n; equations (3a)–(3c)) against measured runoff is in part expected, as the components use the fitted n values. However, such linearization in to the separate components does allow an assessment of the individual contributions to total changes in runoff. If we consider ∆Q_P + ∆Q_PET only, we just capture 38% of spatial variability in observational changes in runoff (RMSE = 30.6 mm yr⁻¹; figure 1(b), orange dots). This illustrates the impact on runoff projections of neglecting changes in parameter-n, which some earlier studies have done. That is, we verify that this new part, ∆Q_n, is an important contributor to recent runoff changes. We analyze this further by a first-order linear Taylor series expansion of equation (1) to separate the direct impacts of P, PET and n changes on runoff changes in the recent past (ΔQ_P, ΔQ_PET and ΔQ_n respectively). Figure 2 illustrates that 11 out of 33 catchments (e.g. Fitzroy, Congo, Mississippi, St Lawrence, Rhine) experienced decreases in precipitation amount or increases in PET over the period 1981–2005, both acting to reduce runoff. However, the component data shown in figure 2 also illustrate that for these catchments, changes in water retention had the opposite effect to increase runoff markedly, making overall runoff changes positive (see also table S1). Figure 2 thus shows that the effect of decreased n is in general to mitigate the P- or PET-induced reduction in river runoff. Our result confirms the dominant contribution of water retention capability changes on runoff changes in 21 out of 33 major catchments (table S1). This also explains the large difference that can occur when this effect is prevented from changing in flow estimates (figure 1(b)).

Zoom In Zoom Out Reset image size

The inferred decreases in water retention capability may be partly related to plant physiological stomatal response to rising atmospheric CO₂ concentration and vapour pressure deficit, both leading to a higher fraction of P converted to runoff. In one study, there is strong detectable evidence that rising CO₂ concentration leads to reduced opening of stomata, which in turn is associated with lower leaf-level transpiration and higher runoff (Gedney et al 2006). In addition, land use change with the replacement of forest by short vegetation, which has less transpiration ability, may also explain some runoff increases diagnosed by the decrease of n. This may be expected to apply in the Congo Basin, where extensive deforestation has occurred (Duveiller et al 2008). In contrast to the tropical regions, decreased runoff in the temperate catchments such as the Yellow River may be partly attributed to recent afforestation, reflected by an increase of n. This will induce an increase in evapotranspiration, and in rainfall interception loss (Yang et al 2009, Feng et al 2016). Additionally, direct physical engineering of the terrestrial hydrological system (e.g. dam construction, reservoir building and river channeling) and agricultural management (e.g. irrigation and terraces) also lead to a significant change in n, and thus alter runoff (Jaramillo and Destouni 2015). Such water management can act to either decrease n and facilitate runoff (wetland suppression, river channeling), or to increase n and reduce runoff (evaporation from dams, and re-use of stream water for irrigation).

Here we apply the same Budyko-based attribution of runoff changes to P, PET and n changes. These are simulated by ten state-of-the-art ESMs from the CMIP5 database (Methods, section S3). The estimate of changes to historical runoff amounts, and using the Ensemble Mean of these CMIP5 models (CMIP5-EM), are shown in figure 3(a). Such estimates based on CMIP5-EM in general project lower values of runoff change (i.e ΔQ). We compare differences in the Taylor series-based individual components ΔQ_P, ΔQ_PET and ΔQ_n value, and between CMIP5-EM and measurements in figures 3(b)–(d). Values also in table 1 (standard deviations across catchments). Magnitudes of ΔQ_P for CMIP5-EM are in general smaller than those from measurements (figure 3(b)). This arises in part due to ESMs differing markedly from observations for precipitation estimates. This even includes the signs of these changes in some regions (Knutti and Sedláček 2012, Flato et al 2013), which will balance to suppress mean CMIP5-EM estimates. Other suppression of changes in CMIP5-EM is also due to decadal timescale precipitation fluctuations, which reflects natural variability caused by the chaotic nature of large-scale atmospheric circulations and unforced variance in the coupled land-ocean-atmosphere system (Deser et al 2012). Random internal variations in PET are smaller, and there is better agreement for ΔQ_PET between model and measurement-based values (figure 3(c)).

Zoom In Zoom Out Reset image size

For runoff changes attributable to retention, figure 3(d) shows a severe underestimation of ΔQ_n change magnitudes (both positive and negative) by ESMs, compared to measurements during the historical period. We conjecture that in the instances where ESMs fail to capture the observed decreases in ΔQ_n, it is a consequence of missing components of physical engineering and water management in the models. This is despite most ESMs having some representation of land use change. Besides physical changes, these alterations to the land surface may increase ET (for instance re-use of stream water in irrigation water and dam evaporation) which will also decrease ΔQ_n i.e. raise retention. In instances where ΔQ_n increases in data, but not in models (figure 2(d)), then this may be due to the omissions of channeling of small streams, wetland and riparian area suppression, and of some land management practices (e.g. no crop cover in winter). All these actions may increase runoff, i.e. raise ΔQ_n increase by lowering retention.

We formalize the impact of these missing ΔQ_n components in a simple statistic. We calculate an overall standard deviation across catchments (SD; mm yr⁻¹) with data values from table 1 as ${\sigma }_{D}\,=\sqrt{{\sigma }_{{\rm{\Delta }}{Q}_{P}}^{2}+{\sigma }_{{\rm{\Delta }}{Q}_{PET}}^{2}+{\sigma }_{{\rm{\Delta }}{Q}_{n}}^{2}}$ (this statistic assumes independence between components in the Budyko framework). In order to estimate the effect of model underestimation of inter-catchment variability in 'n-induced' runoff changes, we then recalculate an overall SD statistic, but instead replace ΔQ_n with the CMIP5-EM based value (replacing 33.9 with 9.3; table 1). This then gives a new overall SD (i.e. σ_D,n). The two values allow us to represent, as a percentage, the effect on runoff of missing water managements processes on modelled water retention changes, returning a value of underestimation as: $S=100\,\times (1-{\sigma }_{D,n}/{\sigma }_{D})=32 \% .$ This value quantifies the extent to which CMIP5 models lack the representation of physical engineering processes and water management impacts on observed spread of runoff changes observed across catchments. This is a severe limitation to the proper evaluation of models with observed runoff changes, except in catchments with minimum human intervention.

Although there are noted deficiencies of ESMs depiction of retention, compared to measurements, due to their non-representation of water management processes, we can still explore future water retention changes within the CMIP5 model ensemble. These changes in n reflect rising atmospheric CO₂ concentrations, changing imposed climate, varying vegetation distribution and its water use, along with land use change. We again do this via analysis of the ensemble mean of the CMIP5 models (CMIP5-EM), but here to analyze future runoff changes (we additionally provide individual model estimates in the SI). Figure 4(a) shows the estimated annual runoff change from CMIP5-EM by the end of this century under the RCP4.5 scenario, relative to the recent historical past (1981–2005). The increase in runoff mainly occurs in the Arctic region, northeastern China, South Asia, and North Africa, while the Amazon region, western and southern Africa, the United States, southern Europe, and central Asia experience decreased runoff (by −20 mm yr⁻¹ or more; figure 4(a)). Similar patterns for runoff changes occur under the more intensive RCP8.5 scenario across most regions (figure S2(a)). As for the analysis of historical measurements, the heterogeneity in future runoff change is successfully captured by the Budyko framework. The Taylor series expansion of the Choudhury-Yang equation allows the separation of runoff changes into precipitation amount, PET, and water retention capability driven components. These sum up to the total future runoff changes (similarities of figures 4(a) versus 4(b), and also S2(a) versus S2(b)).

Zoom In Zoom Out Reset image size

In figure 4(c), the future precipitation amount generally increases, leading to an increase in runoff. This is for most regions (more than 65.4% of land surface). Decreases in P-induced runoff only can be found in Central America, southeastern Amazon, southern Europe, Australia and southern Africa. Driven by rising temperatures, PET consistently increases across all the land surface (Scheff and Frierson 2014), resulting in a negative runoff change component (figure 4(d)). This cancels out the upward effect on runoff of increases in the precipitation amount in temperate North America, southern China, western Africa, and the northwestern Amazon region (figure 4(e)). For retention changes, represented in the 'n-parameter', we find in general reduced values of that parameter, contributing to increased runoff (i.e. ΔQ_n) values. This n-induced increase is projected over much of the globe, particularly in most parts of South America, large parts of central and southern Africa, the eastern part of North America, southeast China, and eastern Siberia (figure 4(f)). In these regions, the increase in runoff associated with decreased n is generally larger than +20 mm yr⁻¹ (figure 4(f)). However, there are regions where n-driven future runoff change is negative, including dry central Asia and northwest China. Critically, over much of the globe, future projections show that the changes in ΔQ_n are of similar order of magnitude to ΔQ_P and ΔQ_PET (figure 4). When assessing individual models, rather than the ensemble mean, we find that the inter-model contributions from changes in PET and water retention capability on runoff show relatively high consistency. However, models disagree on the contributions of precipitation amount (figures S3 and S4).

Overall the CMIP5-EM shows that the area projected to experience decreasing runoff under RCP4.5 scenario, when only considering the effects of precipitation amount and PET, is 47.6% of the land surface. However, when the additional effect of water retention change is included, i.e. with all factors changing, the area experiencing decreasing runoff reduces to 41.3% of land surface. At the global scale, mean runoff change associated only with changes in precipitation amount and PET is a decrease of −2.77 mm yr⁻¹. This is of opposite sign to the runoff change when all factors of change are included, which then gives an increase of +3.81 mm yr⁻¹. RCP8.5 scenario also shows a similar result. Under that scenario, the estimated global runoff associated with changes in precipitation amount and PET only is a decrease by −6.98 mm yr⁻¹. When all factors including water retention changes are considered, the estimated global runoff is an increase of +5.11 mm yr⁻¹.

The importance of retention changes in projecting runoff has been established. We now identify and explore a range of complex processes in ESMs that contribute to driving estimated future retention changes. We plot changes in n (i.e. Δn) during the 21st century for the RCP4.5 scenario (figure 5(a)). The Δn values exhibit substantial spatial heterogeneity of change, and as expected have similarities to ΔQ_n of figure 4(f). The increases in n associated with reduced ΔQ_n are found in some dry regions, while decreases in n appear in many other regions. The latter (i.e. lower retention n) dominates, covering more than 65.4% of land surface. These regions include wet tropical catchments, eastern North America, high-latitude Eurasia, and southeastern China. Similar spatial patterns of change in n are also found under RCP8.5 scenario (figure 5(b)). When forced with the more intensive warming scenario (i.e. RCP8.5), as expected the magnitude of Δn is larger compared to RCP4.5 scenario (figure 5(a)). Although there is general agreement, Δn diagnosed from individual ESMs shows a strong divergence in the high northern latitudes under RCP8.5 scenario (figure S6).

Zoom In Zoom Out Reset image size

Three ESM outputs are analyzed in relation to Δn: changes in leaf area index (ΔLAI); changes in transpiration per leaf area (Δ[Tr/LAI]); and changes in the daily intensity index ΔSDII (Hartmann et al 2013). SDII is designed to capture the situation where rainfall intensity is such that it exceeds infiltration capacity, increasing runoff. We find that in the dry mid-latitude regions, the water retention increases (Δn > 0) are mainly due to increased LAI for both RCP4.5 and RCP8.5 scenario (figures 5(c) and (d)). Higher atmospheric CO₂ levels lead to increased photosynthetic rate, thereby increased LAI, which raises transpiration and reduces runoff. A more complex situation emerges where Δn is negative, causing raised runoff. In figure 5(c), negative Δn change is mostly driven by precipitation intensity increases under RCP4.5 scenario (see also figure S8(d)). However, decreases in [Tr/LAI] instead dominate n decreases under RCP8.5 scenario (figure 5(d)), particularly in tropical regions, northeast China, and part of the central North America (see also figure S8(g)). This finding of a switch in dominant driver between emission scenarios for year 2100 also exists in many individual climate models (figures S9 and S10).

We further consider the climate scenario dependence for the dominant driving factors of Δn between RCP4.5 and RCP8.5 (figures 5(c) versus 5(d); figures S9 versus S10). It was previously reported (e.g. Gedney et al 2006, Milly and Dunne 2016, Swann et al 2016) that the reduction of stomata opening due to elevated atmospheric CO₂ concentration or changing climate (e.g. increased vapor pressure deficit) will impact transpiration and thereby alter runoff. However, our finding suggests that future precipitation intensity changes, rather than vegetation, play a major role in changing regional and global runoff over most land regions in the RCP4.5 (figure S4(c)). Based on both theoretical (O'Gorman and Schneider 2009) and modeling studies (Donat et al 2016), the intensity of precipitation, coincident with global warming, is expected to rise (figure S8). More intense precipitation in arid and semiarid environments where the precipitation rate is higher than the infiltration rate of the soil would result in higher runoff (Dunne et al 1991). Thus, for the same amount of annual precipitation, fewer but more heavy rainfall events will produce more runoff than many lighter ones, and result in parameter-n decreases. Under the RCP8.5 scenario, and consistent with previous results, parameter-n decreases are dominated by decreases in [Tr/LAI] (figure 5(d)), and this too will result in more water for runoff. Thus, high accuracy estimates of the expected increase in daily precipitation intensity and accurate parameterization of stomata closure in response to elevated CO₂ are important for ESM modeling community to achieve.

Furthermore, it is worth noting that the characteristics of rainfall seasonality including its magnitude, timing and duration components might affect plant activities and runoff generations, altering the local water retention capability of watersheds (Roderick and Farquhar 2011). Yet its influences are not explicitly quantified in this study. For example, an asymmetry where a decrease in P in the dry spell could lead to a small decrease in Q, while an increase in P in the wet spell could lead to a larger increase in Q. This might be of particular interest given evidence that in a warming world, wet seasons are expected to get wetter (partly due to increased water holding capacity in the atmosphere), whilst for drier regions, are projected to dry seasons get drier (Chou et al 2013).