Earth has been experiencing unprecedented climate change since the 1950s . Changes in the climate system are expected to lead to regionally divergent alterations in the hydrological cycle . In particular, with the CO2-induced increase in radiative forcing, global runoff is expected to increase . Yet the change is also highly heterogenous across different regions . For example, while large runoff increases are expected in moist tropics and high latitudes, dry tropical regions are likely to experience a decrease in runoff . Moreover, obvious uncertainties also exist for projected changes in regional and global runoff. Coupled state-of-the-art global climate models (GCMs) and global hydrological models (GHMs) are increasingly used for assessments of changes in the hydrological cycle . Different GCMs use distinct representations of the climate system, leading to “climate model structural uncertainty” . Furthermore, differences in GHM structures could also result in large uncertainties in modelled runoffs. In particular, GHMs are modelled on a global scale, and most GHMs are not calibrated. It is common that the performance of GHMs tends to vary with regional location and catchment size . Because simulated river runoffs can guide policy decisions regarding regional water resource management and climate change adaption , assessing model performance and reducing uncertainties in modelling results are especially desired at regional scales .

The Mekong River (MR) is an important international river running from the Tibetan Plateau through China and the countries of mainland Southeast Asia (i.e. Myanmar, Laos, Thailand, Cambodia, and Vietnam) before emptying through southern Vietnam into the South China Sea . The upper reach of the MR, located in China, is called the Lancang River and the lower reaches, as they pass through each country, are known as follows: in Laos it is Mènam Khong, in Thailand it is Mae Nam Khong, in Cambodia it is Mékôngk, and in Vietnam it is Sông Tiên Giang (https://www.britannica.com/place/Mekong-River, last access: 23 February 2023). The agricultural production and life of the residents along the river are directly affected by the changes in the water volume of the MR. Seasonally, the water from the Mekong flood pulses when the river backs up and floods the Tonlé Sap in Cambodia, which is responsible for local fish raising that provides up to 70 % of Cambodians' animal protein intake, and also allows for the growing of floating rice, which feeds the communities in central Cambodia . Furthermore, the fish that originate in the Tonlé Sap leave the lake as the water level decreases and restock a large part of the Lower Mekong reach , providing an important source of income for the population of this area. The Mekong also provides the freshwater necessary for growing rice in the delta in Vietnam, which is considered the “rice bowl” of Vietnam .

The main stream of the MR extends over 4800 km, with a drainage area of about 795 000 km2 . The average annual runoff at the outlet is 14 500 m3s-1, making it the 10th largest river in the world in terms of water discharge . However, the performance of GHMs for the Mekong River basin (MRB) has rarely been reported , which could impede improved predictions of future runoff. Importantly, while a number of models have been used to simulate the runoff of the MR , these studies focus on the simulation and analysis of the MR runoff under different climate models by using a single hydrological model and without comparing the performances of different hydrological models. On the other hand, assess the applicability of 10 hydrological models in the MR using one set of meteorological forcing data from the Global Soil Wetness Project Phase 3 (GSWP3) under the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP). Their study shows that the calibrated GHMs have the best performances during the historical scenario period. However, these studies do not systematically analyse the runoff simulation results of long-term historical periods (including the historical period of historical scenarios and the real-time period of Representative Concentration Pathway (RCP) scenarios, i.e. from the start simulation year of the RCPs to 2020, for which observed runoff data are available) under different GCM–GHM combinations. In the context of high uncertainty in runoff projections under the RCPs, the use of real-world observations to evaluate future projections during the real-time period can increase the reliability of the simulation for more distant future periods. Such an analysis is meaningful and urgent to potentially assess and reduce the uncertainty or bias of runoff simulations introduced by both GCMs and GHMs to achieve more reliable future projections . As one of the longest rivers in the world and one that is the major water source for 65 million people in the five countries of the Lower Mekong region, comprehensive model evaluation for the MR runoff is critical in order to understand the limitations and strengths of, and also further improve, the global hydrology models for wide application and for better policy decisions for the region.

The goal of this study is to understand the temporal and spatial variation characteristics of the MR runoff, with a focus on the future runoff changes under different RCP scenarios. To this end, we perform the following analyses: (1) we first perform a trend analysis and significance test on the historical observed runoff during the period 1971–2020 at the eight gauging stations located along the Lower Mekong main stream (Fig. ). (2) We then evaluate the runoff simulations of different GCM–GHM combinations in ISIMIP over historical scenario periods and real-time periods of RCPs against observed runoff from the above gauging stations and identify the best GCM–GHM combination for predicting future runoff changes. (3) Finally, we comprehensively analyse the future runoff pattern changes (including annual runoff and seasonal runoff) in the upper, middle, and lower reaches of the MRB under future RCP scenarios based on the best GCM–GHM combination.

This study systematically analyses the performance and uncertainty in the runoff simulations from five GHMs driven by four GCMs within the MRB during historical periods. An interesting finding is that the variability introduced by the GCMs was similar to or even greater than that introduced by the GHMs on the runoff simulation (Figs.  and ). For example, in Fig. a, the median R2 of different GHMs under the same GCM driver can differ by 0.20, but the median R2 of the same GHM under different GCMs drivers can differ by more than 0.20. To reduce the variability in the runoff simulation, on the one hand, we can obtain a well-performing GHM through comprehensive evaluation. In this study, three performance indicators were combined under eight hydrological stations, and WaterGAP2 (i.e. GHM) was found to have the best performance (highest R2 and NSE and lowest Pbias) under four GCM drivers in the MRB. In addition, even a good GHM has high uncertainty for future runoff projection under different GCM drivers. A feasible approach at this time should be to combine the ensemble average of runoff results from the GHM driven by different GCMs, which can help reduce the uncertainty from climate models in future projections. At the same time, the standard deviation of runoff results from the GHM driven by different GCMs can be used to quantify the uncertainty in future runoff projections. This approach gives equal weight to each GCM, often referred to as “model democracy”, and has been widely used in climate impact assessments . Another approach that can potentially reduce uncertainty is a weighting scheme that considers the performance of the GCMs . The GCMs are weighted by different statistics in the past or present, and the weighting coefficients are applied to the future projections of the GCMs. However, there is a risk that a GCM that performs poorly in the current climate may perform better when environmental conditions are beyond the contemporary range of change . It is worth mentioning that a novel and promising approach to constrain uncertainties is the emergent constraints (ECs). The EC approach consists of statistical (emergent) relationships between an observable quantity (X) in the past or present climate and a quantity (Y) related to the future climate across GCMs . Combining emergent relationships with observations can potentially reduce uncertainty in future projections, and several published ECs have shown us positive effects . We encourage further experimentation with various approaches, including those described above, to overcome the uncertainty among GCMs in the MRB.

Another point is that under different RCPs, the interannual runoff of the three representative sites has a significant (p<0.05) increasing trend, which is consistent with the previous relevant studies suggesting that MRB runoff would increase in the future due to climate change . A novel finding is that the upstream, midstream, and downstream stations in the MRB show different patterns of runoff change under three RCP scenarios. The increase in runoff at the upstream station N2 increased sequentially as the scenarios changed from RCP2.6 and RCP6.0 to RCP8.5. The difference is that the downstream station N8 has the highest runoff increase under the RCP6.0 scenario, while this is not the case under the RCP8.5 scenario. This behaviour is closely related to the combined effects of temperature and precipitation on runoff under different RCP scenarios. Specifically, at upstream stations, the synergistic effect of increased glacial meltwater and increased precipitation caused by warming under different scenarios is greater than the effect of increased evapotranspiration caused by warming. This results in the highest runoff increase under RCP8.5. At downstream stations, the proportion of glacier meltwater to total water volume decreased, suggesting that its impact on total runoff was also lower. In addition, the increase in evapotranspiration due to warming increases with decreasing downstream latitude. Under the combined effect of these factors, the runoff increases under the RCP6.0 scenario (16.36 ± 12.44 m3s-1a-1) and is higher than that under the RCP8.5 scenario (13.28 ± 12.20 m3s-1a-1). This means that the risk of future flooding in the middle and lower reaches of the MRB is still likely to remain a high level, even if we try to manage to stay on a moderate emissions path (i.e. RCP6.0). The novel change patterns of the upper, middle, and lower reaches explored in the study may be able to provide a scientific basis for the future implementation of local water resource management schemes in each reach of the MRB.

Furthermore, in the far future period (2070–2099), the distribution of seasonal runoff within the MRB is more complex. Despite the apparent increase in interannual runoff, water stress in the dry season would not decrease or become more severe. Under all RCP scenarios, runoff will decrease in future dry-season months (e.g. May). Even under the RCP8.5 scenario, the percentage of runoff reduction at representative sites in May was above 15 %, reaching a maximum of 24 %. This can exacerbate water shortages during the dry season and will have particular impact on Cambodia, which, as noted above, relies on the Mekong to fill the Tonlé Sap, and on the Mekong Delta in Vietnam. The increase in interannual runoff is mainly reflected in the rainy months. For example, under the RCP6.0 scenario, the midstream representative station (N5) will have a runoff increase of 25 % in July. This behaviour will increase flood events in the basin, affecting human safety, normal livelihood, and economic activities. Although studies have shown that rational reservoir operation can mitigate hydrological extremes in the basin, the management of such transboundary rivers requires closer cooperation among all the countries in the MRB. It is worth mentioning that our current study focuses on the impacts of future climate change on runoff in the MRB without quantifying dam- and/or reservoir-induced changes in runoff, due to the limited availability of dam and/or reservoir data. In future work, we expect to be able to quantitatively analyse the impacts of human activities and climate change on runoff by acquiring or collecting dam and/or reservoir data to explore the potential of reservoir operations to mitigate the extreme hydrological events under complex future runoff change scenarios (e.g. droughts and floods).

From the 1970s to the present, there has been no significant (p>0.05) change in the runoff of the MRB. In the early operation stages of the reservoirs built in the 1990s, the annual runoff decreased obviously. However, the impact of the reservoir on the annual runoff after the completion of water storage is small. The ensemble-averaged results of GCMs can reduce the uncertainty in the runoff simulations by different climate models. Moreover, WaterGAP2 performs the best runoff simulation at each station, with the average R2, NSE, and Pbias of the stations being 0.78 %, 0.68 %, and 5.5 %, respectively. Based on these evaluation results, the WaterGAP2 runoff simulation has been used in the MRB to analyse runoff changes under future scenarios.

Under the RCP scenarios, the future interannual runoff of the MRB increases significantly (p<0.05). Notably, the upper and lower reaches of the MRB show different patterns of runoff change under three RCP scenarios, which are associated with the combined effects of temperature and precipitation on runoff for each reach. Under the RCP6.0 scenario, the MRB has the highest increase in interannual runoff. Seasonal changes in annual runoff in the MRB under a future climate are more complex. Under the RCP2.6 and RCP6.0 scenarios, the runoff of the MR during the rainy season will increase, and the increase in the RCP6.0 scenario is higher than that in the RCP2.6 scenario. The changes in runoff in the dry season are relatively stable under the two scenarios. However, the seasonal runoff changes in the MRB under the RCP8.5 scenario are extremely complex. The specific performance of the ensemble average of the GCMs and the WaterGAP2 combination suggests that the dry season will become drier, the rainy season wetter, and the distribution of water resources over the year more uneven. Overall, this study provides novel insights for future runoff projections from a whole river system perspective and may be able to offer a scientific basis for the future implementation of water resource management schemes in the MRB.