The fate of the Caspian Sea under projected climate change and water extraction during the 21st century

Close to 60 global climate models were included in CMIP5 (Taylor et al 2012) and about 120 models in CMIP6 (Eyring et al 2016). Analysis of their associated land-sea masks indicates that a considerable number of the CMIP5 and CMIP6 climate models either completely ignore, or do not accurately prescribe, CS area. Therefore, we set selection criteria based on (a) how well CS area is represented in the models, and (b) the availability of precipitation and evaporation fields for both representative concentration pathways (RCP) 4.5 and 8.5 (Meinshausen et al 2011) for CMIP5 and shared socioeconomic pathways (SSP) 245 and 585 (Riahi et al 2017) for CMIP6. RCP4.5 is an intermediate scenario, with 4.5 Wm⁻² radiative forcing by 2100, and RCP8.5 is an extreme climate change scenario, with 8.5 Wm⁻² radiative forcing by 2100. SSP245 and SSP585 represent similar (although not identical) intermediate and extreme scenarios, respectively, in CMIP6.

At the time this investigation was performed (September 2020), based on the above criteria, we selected in total 18 climate models, of which 11 are from CMIP5 and seven are from CMIP6. We only considered the first ensemble simulation (CMIP5: 'r1i1p1' and CMIP6: 'r1i1p1f1') if a model had multiple ensemble simulations. The list of the models used in this study and their land-sea masks are presented in figure 1. See also supplementary information figures S1 and S2 (available online at stacks.iop.org/ERL/16/094024/mmedia) for land-sea masks of models that were rejected from the main study due to poor CS representation and/or missing climate model fields (for model details see table S1 for CMIP6 models and table S2 for CMIP5 models). For comparison purposes, the model precipitation and evaporation fields were interpolated to the same resolution (6 arcminutes) by a first-order conservative interpolation method (remapcon), which works well for flux conservation (Jones 1999), using the Climate-Data-Operator software (Schulzweida 2019).

The hydrologic budget variation was assessed by comparing the mean 'precipitation minus evaporation' (P–E) field between the start and end of the 21st century from the selected CMIP5 (running between years 2006 and 2098) and CMIP6 (over years 2015–2100) models for the RCP and SSP scenarios. To estimate the CS level variation during the 21st century, we used a hydrologic model constructed for the CS by Koriche et al (2020a). The model is based on fluxes of runoff over the CS catchment, P–E over the CS, and WE for human use (equation (1)). Simulations of lake water level variation in a closed basin like the CS can be substantially affected by the variation of its water level, as this leads to changes in surface area, which would significantly impact the P–E over the sea at each time step. Therefore, for every time step, the CS surface area is updated based on the volume of the previous time step to be considered for the current time step water balance (P–E) estimation (equation (1)).

$$\begin{align}\Delta {\text{CS}}{{\text{V}}^t} & = \Big[ \left( {P_{{\text{land}}}^t - E_{{\text{land}}}^t} \right)A_{{\text{land}}}^{t - 1} + { }\left( {P_{{\text{sea}}}^t - E_{{\text{sea}}}^t} \right)A_{{\text{sea}}}^{t - 1} \nonumber\\ & \quad - \Delta {\text{W}}{{\text{E}}^t} \Big]\Delta t.\end{align} \tag{ 1 }$$

where: CSV is CS volume, P is precipitation, E is evaporation/evapotranspiration, and A is surface area, all over the land or sea part of the basin as denoted by their subscript, ΔWE is the increment in human extraction of water, Δt is the time step (in this case 1 month). We assume that groundwater contributions are small, based on previous studies (Zekster 1995, Golovanova 2015), and so we have not included a groundwater component.

Currently, the CS is fed by rivers from nine different countries whereas WE information is based on national-level data rather than on the CS catchment boundary. This makes it difficult to get appropriate estimations of water withdrawals solely from the rivers flowing to the CS. We have used records covering the period from 1940 up to 1995 (Shiklomanov 1981, Rodionov 1994, Golitsyn 1995), which are derived estimates of what the CS level would be with zero human WE (see figure 2(a), solid colour lines). These are indirect measurements that can then be used to infer the amount of water withdrawn from the rivers contributing to the CS when compared with the measured CS level observational record (figure 2(a), black dotted line). Following calculation of the yearly withdrawal volume, the estimated annual WEs demonstrate a roughly threefold increase between 1940 and 1990 (figure 2(b)). We note that we are referring to net WE (consumptive water use), which is the amount of water leaving the basin after accounting for the return of a proportion of the WE to the CS. Net WE can occur through several mechanisms, including evaporated water that precipitates outside the basin boundary or through export of water in irrigated crops, livestock, and other goods.

An alternative source of information relating to water withdrawal from the Caspian catchment was compiled by Demin (2007) from various economic and government sector reports for the years 1970–2003 (figure 2(b)). These data show a peak in water withdrawals around 1985–1990 before annual consumption decreases again, until it declines to 43 km³yr⁻¹ in 2003. These figures include consumption, as well as evaporation from reservoirs within the catchment. The reasons for the decline in WE include more efficient water consumption in domestic and industrial processes, changes to land-use, and changes to regional population (Demin 2007). However, this decline in WE was not sufficient to balance out the enhanced evaporation over the CS that occurred due to increased regional temperatures (Chen et al 2017a), and so CS level declined over the last few decades.

For this study, we created future WE values for idealised future projections between 2015 and 2100 in the CS basin based on these estimates (figure 2(b)). The first scenario (FWE1) is a constant extraction rate of 40 km³yr⁻¹, based on Demin (2007), and previously used in Kudekov (2006). A second scenario (FWE2) had constant annual withdrawal at 20 km³yr⁻¹ (figure 2(b), green dashed line). In a third scenario (FWE3) we used new country level population projections for nations within the catchment (Vollset et al 2020) to scale water withdrawal values. The regional population is projected to increase slightly up to mid-21st century before declining to below present-day levels by 2100 (figure 2(b); see also supplementary information table S3). In our simple translation we assume that the 2015 extraction rate is 40 km³yr⁻¹ and that projected changes in population can be linearly transformed to changes in water withdrawals (through domestic water use, agricultural activity, and industrial sector activity). These three WE scenarios bracket the large uncertainties in the compiled historical literature due to the difficulties in sourcing primary catchment level information (described above and shown in figure 2), as the modelled projections will likely be sensitive to the choice of extraction values.

We find a considerable spread in the annual mean water budget of the CS catchment (P–E) between climate models in both CMIP5 (figure 3, red symbols) and CMIP6 (figure 3, blue symbols), with some models displaying a positive net water balance and some a negative balance. When P–E is plotted against the prescribed CS lake area (figure 3) we find that models with larger CS surface areas tend to be drier, whereas models with smaller lake areas tend to be wetter (more positive P–E). The correlation of the modelled catchment water budget with the size of the prescribed CS is indicative of the importance of the magnitude of evaporation from the sea itself in controlling the overall balance. The larger the prescribed CS the larger the amount of evaporation, which tends to outweigh any resulting increase in precipitation and so produces a smaller overall P–E. It is also clear here that even though we have selected models that better represent CS surface area, some of the models (particularly in CMIP5) are up to 75% larger than the observed CS over the last century (figure 3, black symbol).

Zoom In Zoom Out Reset image size

One difficulty faced when attempting to compare and evaluate the modelled P–E with observational data is that the CS level has been increasingly influenced by human water withdrawals, which are not included in the model boundary conditions. A second potential issue is that there is large interannual to decadal-scale variability in the water budget due to modes of internal climate variability such as ENSO (e.g. Arpe et al 2000), and models do not necessarily reproduce the state of those modes at the correct historic time (nor would we expect them to do so). Therefore, comparison of the water balance to a short record or short reanalysis dataset (e.g. ECMWF reanalysis version-5 (ERA5) data, 1979–2005) will not be appropriate. Instead, we compare the model output with the CS level record, corrected to exclude any human water withdrawals (see figure 2(a)), and averaged over a much longer time period 1860–1995 (figure 3, black symbol). The observational data show that the CS has been precariously balanced, fluctuating between positive and negative over the last century. As can be seen in figure 3, models with a prescribed CS surface area closer to historical observations generally also produce a water budget closer to our observationally-derived estimate.

Figures 4(a) and (b) show the mean water budgets (P–E) of the CS basin by the end of the 21st century (2070–2100) as projected by CMIP6 and CMIP5 models for medium (SSP245/RCP4.5) and high (SSP585/RCP8.5) radiative forcing scenarios. Models that represent current CS area more accurately in CMIP5 (CMCC, MPI, CSIRO) tend to have a neutral or positive P–E by 2100 (figure 4(b)). In CMIP6, models with better prescribed CS area (MPI, AWI, EC-Earth3) tend to have a neutral or negative P–E by the end of the century (figure 4(a)). To evaluate the direction of future water budget change in the CS basin, we use the anomalies between the start and end of the 21st century for both modelling groups as presented in figures 4(c) and (d). In both scenarios the model P–E anomalies almost all show conditions getting drier (up to 40 mm yr⁻¹) by the end of the 21st century (figures 4(c) and (d)). The drying is generally more pronounced in the high radiative forcing scenario (SSP585/RCP8.5) than the medium scenario (SSP245/RCP4.5). This is the case for all CMIP6 models, and six out eleven CMIP5 models. The CMIP5 multi-model mean (MMM) does not show this trend as it is heavily weighted by the CSIRO-Mk3-6-0 model. This model displays considerable multi-decadal variability and a neutral long-term trend, and so the averaging periods are more affected by 'noise'. It is only the last two decades of the simulation that RCP8.5 anomalies become much wetter than RCP4.5, due to multi-decadal variability.

Zoom In Zoom Out Reset image size

The CMIP6 models tend to have a more negative (drier) water budget than the CMIP5 models, both historically and in the future projections. These models generally have higher spatial resolutions, better physics parameterizations, and more Earth system components (Eyring et al 2016) than CMIP5 models. Recent studies have also found that this generation of models also have higher equilibrium climate sensitivities (ECS) and warmer 21st century projections (Hausfather 2019, Tokarska et al 2020, Wyser et al 2020), which may play an important role here, given the importance of evaporation over the sea for the overall water budget of the CS. The treatment of the lake in the models (e.g. parameters relating to lake heat absorption and mixing) and coupling of the lake surface to the atmosphere will likewise be important in the variation between modelled water budgets. The magnitudes and patterns of the seasonal cycle of precipitation and evaporation over land are relatively consistent (see supplementary information figures S3(a) and (b) for CMIP6 and figures S4(a) and (b) for CMIP5) but highly variable between models over the sea (figures S3(c) and (d) and S4(c) and (d)). The timing of maximum evaporation varies between August and November and the minimum between February and May. Two CMIP6 models (EC-Earth3 and EC-Earth3-veg), which display highly negative water budgets, despite their CS areas being close to observed, have CS evaporation that remains relatively high even in winter compared to other models (figure S3(c)). These models have an ECS that is relatively high (>4 °C; Tokarska et al 2020). Conversely, the INM-CM5-0 CMIP6 model, which has a highly positive water budget, has a much lower maximum evaporation than other models and low ECS (<2 °C; Tokarska et al 2020). The same model seasonality characteristics are maintained through the future projections (figures S4–S8).

In this section, we explore the question of how increasing anthropogenic climate change and human water withdrawals will impact CS level using a water balance model driven by modelled climate projections and idealised extraction scenarios.

The first set of CS level simulations are driven by climate outputs from CMIP6 and CMIP5 models without considering WE (figure 5(a)). By 2100, up to 8 m decrease in the projected CS level is found using CMIP5 models under RCP4.5 and up to 10 m for RCP 8.5. In CMIP6 based simulations, our results show a decrease in CS level of up to 20 and 30 m for SSP245 and SSP585 scenarios respectively (figure 5(b)). The reasons for the larger decreases in CMIP6 CS level than CMIP5 CS level are partly explained in the last paragraph of section 3.1 (in particular, higher ECS). The declines in CS level are larger with models where larger CS is prescribed in the climate model, and those with higher projected evaporation (e.g. EC-Earth3 and EC-Earth3-Veg, which have up to 70% higher evaporation than other models). On the other hand, models where the prescribed CS in the climate model is smaller tend to display increases in the projected CS level (four CMIP5 models but only one CMIP6), since the P–E over the CS basin is positive (figures 5(a)–(b)). We observed that the projected CS level increases in models with cold bias and smaller ECS (e.g. INM-CM model families). We also find that the CS level projections for CESM-CAM5 (5–6 m) from CMIP5 are smaller than found by Nandini-Weiss et al (2020) of 9–18 m using the same model. Our water balance modelling results in a negative lake-level-evaporation feedback that is not represented in the other study. Here, as CS level declines the surface area shrinks, which reduces the evaporation component, slowing down the rate of desiccation. As a result, CS level decline is not as pronounced as in Nandini-Weiss et al (2020).

Zoom In Zoom Out Reset image size

Next, we incorporate the three idealised WE scenarios, as described in section 2.3. In our analysis of WE impacts we only consider the results from CMIP6 radiative forcing scenarios (SSP245 and SSP585) as they are based on latest versions of climate models with improved process representations. In all scenarios, all models display a decline in CS level but there is variation in the magnitude of this decline. Up to 7 m decline in CS level is observed due to WE. Under the SSP245 scenario the decline in the CS level ranges from 0.7 to 3.6 m for FWE1, 1.4 to 7.6 m for FWE2, and 1.2 to 7.3 m for FWE3 (figures S9(b)–(d)). Under SSP585 CS level ranges from 0.9 to 4.4 m in FWE1, 2.2–9 m in FW2, and 1.9–7.2 m in FWE3 (figures S9(b)–(d)).

We note that the relationship between CS level and CS surface area is highly non-linear. When CS level varies between −27 and −34 m the CS area changes are proportionately large compared with when CS level is below this. This occurs when the shallow northern part of the CS (average depth ∼ 6 m) comes into play. As a result, even when the long-term trend in CS level is seemingly relatively smooth, there is large interannual variability in the modelled CS surface area of up to 10% (figures 6(a), (b) and S10). This is most evident in those models that have smaller long-term trends in CS level. This interannual variability will result in larger seasonal variation in flooding of surrounding wetlands. The variability in CS area, particularly in the shallow northern CS, has implications for coastal communities and conservation of marginal environments at the edge of the lake.

Zoom In Zoom Out Reset image size

We considered a key indicator, or threshold, in the future of the CS to be the point when the shallowest northern section becomes completely desiccated (figure 6(d), grey shaded area). We used the CMIP6 MMM projections of the CS level under the two climate change scenarios and four idealised WE cases to calculate at what point in the 21st century this threshold occurs (if at all). In all scenarios, except for SSP245 with no WE (NoWE), this level of desiccation occurs at some point before the end of the century (figure 6(c)). For the extreme SSP585 scenario MMM and the population based FWE3 extractions the northern CS is desiccated by 2050 (and which is a point crossed by five out of the seven individual models). When considering this indicator of CS decline, the rate of WE is effectively as important as the climate change scenario in terms of the timing. The higher the extraction rate the less difference the climate change scenario makes, and vice versa. With FWE3 there is only ∼12 years difference between SWSP245 and SSP585, whereas with FWE1 there is ∼45 years difference (figure 6(c)). The timing of the desiccation among individual models are different (figures 6(a), (b) and S10).