Probabilistic projections of temperature and rainfall for climate risk assessment in Vietnam

Climate change is a global phenomenon with widespread impacts, posing significant challenges to our planet nowadays (Arias et al. 2021). To cope with climate change, it is important to possess reliable future climate projections, typically derived from outputs of global climate models (GCMs). Both studies on the impacts of climate change and response policies heavily rely on information about potential climate extremes (Seneviratne et al. 2012; Supari et al. 2020). Nevertheless, it has been demonstrated that GCMs often underestimate extreme values (Kitoh & Endo 2016; Iles et al. 2020; Nishant et al. 2022). Furthermore, climate change assessments often employ projections directly from a set of GCMs, such as those utilized in the Fifth Phase of the Coupled Model Intercomparison Project (CMIP5; Taylor et al. 2012), or from their downscaling products, as practiced within the Coordinated Regional Climate Downscaling Experiment-Southeast Asia (CORDEX-SEA) framework (Tangang et al. 2020; Herrmann et al. 2022). These sets of model outputs typically involve a limited number of models under a constrained set of future greenhouse gas scenarios, leading to incomplete coverage of the entire range of future probabilities.

To address the aforementioned limitation, probabilistic approaches are applied (e.g., Rasmussen et al. 2016; Raftery et al. 2017; Vargas Zeppetello et al. 2022). Unlike relying solely on a limited set of GCM projections and specific scenarios, probabilistic methods systematically integrate models, scenarios, and related uncertainty quantifications into projection results. Thus, they can encompass all possible probabilities and consequences arising from diverse future scenarios, particularly accounting for rare yet highly impactful extreme events that are often missed by GCMs (Rasmussen et al. 2016). While numerous studies have focused on quantifying the probability distribution of future change on a large scale, such as the country-specific level or above (Raftery et al. 2017; Liu & Raftery 2021; Chen et al. 2023), there have also been works focusing on local changes (Schölzel & Hense 2011; Kopp & Rasmussen 2015; Rasmussen et al. 2016). For instance, Chen et al. (2023) employed pattern scaling to refine global climate change projections to local scales, enabling them to obtain projections of long-term temperature for any region of the world. However, their method has a relatively coarse resolution (2.5° × 2.5°), limiting its applicability to large areas exclusively. On the other hand, Kopp & Rasmussen (2015) used the surrogate mixed model ensemble (SMME) approach to estimate the probabilities of changes in both precipitation and temperature at the U.S. county-level, achieving a higher horizontal resolution of up to 1/8° × 1/8° (∼14 km). The SMME method relies on probabilistic projections from the Model for the Assessment of Greenhouse Gas Induced Climate Change (MAGICC) (Meinshausen et al. 2011) of global mean temperature change to evaluate the GCM outputs. It then constructs a surrogate model (SM) to capture the extreme portion of the probability distribution that GCMs cannot identify. The SMME has been applied to quantitatively analyze economic risks posed by climate change in the United States (Houser et al. 2015; Kopp & Rasmussen 2015). While other techniques, such as the Bayesian Model Averaging (Tebaldi et al. 2005) and the Weighted Interval Combination (Luo et al. 2019) approaches, also combine multiple model projections, the SMME has the advantage of being able to explicitly represent uncertainty contributors by integrating both parametric perturbation uncertainties from MAGICC and structural uncertainties through the use of SMs (Rasmussen et al. 2016). Thus, the SMME can provide better coverage of potential futures that could be missed by raw GCM ensembles.

With over 3,200 km of coastline and many cities located in low-lying areas, Vietnam is highly vulnerable to climate change (Dasgupta et al. 2007; MONRE 2020). According to the Climate Risk Index report of 2018 (David et al. 2018), Vietnam is among the countries most impacted by extreme events, ranking fifth in 2016 and eighth for the period 1997–2016. Climate change manifestations, such as increases in temperatures and altered precipitation patterns, significantly affect socioeconomic activities and impede growth (Espagne et al. 2021; World Bank 2022). The cumulative direct impact on various socioeconomic sectors in Vietnam leads to an average annual gross domestic product loss of 4.5, 6.7, and 10.8% for respective increases in global warming levels of 1.5, 2, and 3 °C relative to the pre-industrial period of 1851–1900 (Espagne et al. 2021).

In Vietnam, future climate projections primarily rely on downscaled GCM outputs using both statistical methods (MONRE 2009, 2012; Tran-Anh et al. 2022, 2023) and dynamical approaches (MONRE 2012, 2016, 2020; Ngo-Duc et al. 2014; Katzfey et al. 2016; Trinh-Tuan et al. 2019; Nguyen-Duy et al. 2023). However, it is important to note that these studies were constrained by the use of only a restricted number of CMIP GCMs. For example, in the national report on climate change and sea level rise in Vietnam (MONRE 2020), only 16 dynamical downscaling experiments based on 10 CMIP5 GCMs were employed. As discussed earlier, the existing projections for Vietnam cannot cover the entire range of future probabilities, and they may particularly miss rare yet highly impactful extreme events. It is crucial to emphasize that projected information on extremes is particularly vital for adaptation planning in a country highly vulnerable to climate change like Vietnam. Therefore, in this study, we made a first attempt to provide probabilistic projections of temperature and rainfall in Vietnam using the SMME approach. The results are expected to offer valuable inputs for studies assessing climate change and its impacts on socioeconomic activities during the 21st century, supporting adaptation planning to cope with climate change in Vietnam.

To perform the probabilistic projection for Vietnam, we used outputs from 31 CMIP5 GCMs (Table 1). These outputs include daily temperatures and rainfall for the baseline period 1986–2005, and projections for the future period 2006–2099, considering Representative Concentration Pathway (RCP) 4.5 and 8.5 scenarios (Moss et al. 2010). We chose RCP4.5 and RCP8.5 to align with the scenarios used in the latest national report on climate change and sea level rise in Vietnam (MONRE 2020). MONRE (2020) has prioritized these two scenarios: the moderate scenario RCP4.5, considering active international efforts to mitigate climate change, and the high greenhouse gas concentration scenario RCP8.5, enabling preparation for a potential worse-case future.

The daily temperatures (daily average, daily maximum, and daily minimum) and rainfall from the aforementioned 31 GCMs have been statistically downscaled to a 10-km resolution for Vietnam using the Bias Correction Spatial Disaggregation (BCSD) technique (Tran-Anh et al. 2022). This downscaled dataset is referred to as CMIP5-VN. CMIP5-VN has undergone validation, and the results demonstrate its overall good performance across different sub-climatic regions of Vietnam.

The observational gridded dataset, referred to as OBS, for rainfall and temperatures was described by Tran-Anh et al. (2022). OBS is used to downscale the SMME results (explained below) from the average monthly scale to the daily temporal resolution. OBS has a spatial resolution of 0.1° × 0.1° and covers the period 1986–2005. OBS's temperature was constructed using data from 147 stations and the Kriging interpolation method (Switzer 2014). Meanwhile, OBS's rainfall was constructed with the Spheremap method (Willmott et al. 1985), utilizing daily data from 481 rainfall stations across Vietnam.

We employed the surrogate mixed model ensemble (SMME) technique, as described by Kopp & Rasmussen (2015), to quantify the probability of temperature and rainfall changes in the 21st century in Vietnam. The SMME method utilizes probabilistic projections of global mean temperature change from the simple MAGICC (Meinshausen et al. 2011) to assign weights to GCM outputs. Subsequently, the SMME constructs model surrogates that capture the extreme tails of the MAGICC probability distribution. It is important to note that these extreme tails are generally not represented by the original GCM ensemble. The key steps of the SMME method are summarized below.

The MAGICC model (Meinshausen et al. 2011) is used to project global temperature changes under different greenhouse gas scenarios. Various versions of the MAGICC model have been widely employed by the climate community, with MAGICC6 specifically utilized to establish the probability distribution of global temperature increase in the Fifth Assessment Report (AR5) of the Intergovernmental Panel on Climate Change (IPCC) (Collins et al. 2013). For each RCP4.5 and RCP8.5 scenario (Moss et al. 2010), the probability distribution of global temperature increases is constructed using MAGICC6 in probabilistic mode with 600 simulations (Meinshausen et al. 2011), wherein each simulation incorporates distinct input parameters. The quantile ranges of MAGICC6 are then determined based on the outputs from these 600 runs. Accordingly, the 5th, 7th, 83rd, and 90th percentiles of these 600 runs correspond to temperature increases of 1.5, 1.6, 4.9, and 5.9 °C per CO₂ doubling, respectively (Kopp & Rasmussen 2015).

When considering a projection from either a CMIP5 GCM or a MAGICC6 experiment, we denote the variable (°C) representing the change in global temperature at a future time point relative to the period 1986–2005. In addition, represents the average for the far future period 2080–2099.

In the initial step, the values projected by MAGICC6 are utilized to establish 10 percentile bins with unequal widths: [0,8], [8,12], [12,20], [20,40], [40,60], [60,80], [80,88], [88,92], [92,98], and [98,100]. Notably, the low and high percentile bins are defined with narrower widths to enhance the SMME model's ability to identify extreme scenarios.

Subsequently, each CMIP5 GCM model is assigned to the corresponding percentile bin based on its projected value. It may happen that certain percentile bins from MAGICC6 lack representation by any CMIP5 model, such as the 1st, 2nd, 9th, and 10th bins in the RCP4.5 scenario, or may be represented by only one model, for example, the 8th percentile bin in the RCP4.5 scenario (Table 1). In such cases, SMs are constructed to ensure that each of these percentile bins is represented by two models in total. The CMIP5 GCMs selected to build the SM are those with values closest to the centered of MAGICC6 for the missing bin.

Next, the variables of SM-A are downscaled to a 10-km resolution for the Vietnam territory using the BCSD method, which was also applied to generate the CMIP5-VN dataset (Tran-Anh et al. 2022). The downscaled dataset derived from these SMs is referred to as SM-VN.

Thereafter, a probability weight is assigned to each model based on two factors: the width of the percentile bin and the number of models sharing that bin (Table 1, Supplemental Table S1). Models falling within the same bin are given equal weights. For instance, in Table 1, the fourth bin corresponds to the percentile range [20, 40], which includes three models: GISS-E2-R, GISS-E2-H-CC, and GISS-E2-H. Consequently, these three models are assigned equal weights (0.0667), calculated as the width of the bin (20%) divided by the total number of models (3).

Finally, the SMME model is constructed by synthesizing the patterns of the CMIP5 GCMs and SMs using their respective probability weights. It is worth noting that only monthly values at global scale are used in the above-mentioned steps.

We observe that ⁠, as represented by the GCMs, only covers the 12th–89th percentiles of the MAGICC6 results (corresponding to 1.13–2.9°C increases in global temperature at the end of the 21st century) for RCP4.5 (Table 1) and 12th–83rd percentiles (equivalent to 2.59–6.53°C increases) for RCP8.5 (Supplemental Table S1). Meanwhile, the SMME models, which include both the CMIP5 GCMs and SMs, can encompass the entire range of represented in MAGICC6 for both scenarios (Supplemental Figure S1). This demonstrates that the SMs could well capture the extreme intervals not represented by the CMIP5 GCMs, thus enhancing the coverage of future climate projections.

The local warming probability distribution for Vietnam is derived from the global SMME results, applying the pattern scaling technique (Santer et al. 1990; Tebaldi & Arblaster 2014). The local changes of a variable X from either a CMIP5-VN or an SM-VN experiment can be expressed using Equation (1) where i indicates a specific 10-km resolution grid cell in Vietnam and represents the estimation based on a 30-year moving average. Subsequently, the monthly SMME projections for each 10-km grid cell in Vietnam are obtained by aggregating the results from the component CMIP5-VN and SM-VN experiments, using weights received from the respective global models.

The monthly SMME projections, once generated, are further downscaled to daily temporal resolution using the random resampling technique of the BCSD approach outlined by Wood et al. (2002, 2004). This temporal-downscaling process involves selecting a random year from the historical period (1986–2005 in this study) as a reference for a given future year. Then, the monthly averages of the future year are disaggregated into daily data by adjusting the daily gridded OBS dataset. Temperature adjustments are made using addition, while rainfall adjustments employ multiplication, ensuring that the monthly mean of the future data remains unchanged. The final daily dataset is referred to as SMME-VN.

The interest in producing the daily SMME-VN dataset arises from the need to quantitatively assess future climate change risks and extremes (MONRE 2020), typically represented by a set of extreme indices, using daily information. For instance, the joint World Meteorological Organization Commission on Climatology (CCl) and the Climate Variability and Prediction (CLIVAR) Expert Team on Climate Change Detection and Indices (ETCCDI) has worked to define 27 core climate indices from daily temperature and rainfall data (Karl et al. 1999). Given that the advantage of probabilistic projections lies in the ability to account for extreme events, daily data becomes essential, among other reasons, for subsequently estimating associated extreme indices. These indices not only aid in quantifying future extremes but can also be used as inputs for further impact assessment studies.

We acknowledge that the BCSD approach applied in this study, along with its random resampling technique, does not maintain the daily correspondence of the coarse spatial resolution GCMs. Despite this limitation, BCSD has been extensively employed for temporal disaggregation of GCMs' projections in impact assessments worldwide (e.g., Zhang et al. 2017; Duan et al. 2021; Tran-Anh et al. 2022; Atiah et al. 2023; Michalek et al. 2024; Rabezanahary Tanteliniaina & Andrianarimanana 2024). Some prior studies have compared the performance of this temporal disaggregation method with traditional daily-preserving downscaling approaches (e.g., Maurer and Hidalgo 2008; Yang et al. 2019). For instance, Maurer and Hidalgo (2008) indicated that both methods exhibit some skill in reproducing daily variability of observed rainfall and temperature extremes, with a certain advantage for the latter approach. It is worth noting that recently developed downscaling approaches, such as the Quantile-Preserving Localized-Analog Downscaling (Gergel et al. 2024), allow for the efficient preservation of GCMs' daily variability. These approaches could be considered in our future studies.

In this study, we assume zero forced uncertainty (F = 0) and zero uncertainty of the scenarios (S = 0) at the beginning of the future period, with the unforced uncertainty representing 100% of the total model uncertainty.

It should be noted that despite RCP8.5 representing a scenario with higher radiative forcing, its divergence from RCP4.5 becomes clear only from the mid-21st century (Collins et al. 2013; Xin et al. 2013). Previous studies indicated that projection signals over a specific area, particularly in the near future period, typically exhibit uncertainties due to various factors such as projected greenhouse gas concentrations, model imperfections, and natural variability (Sorteberg & Kvamstø 2006; Deser et al. 2012; de Elía et al. 2013). Nguyen-Thuy et al. (2021) estimated the time of emergence (TOE), i.e. the time that the climate change signal exceeds the above uncertainties, for projected temperature and rainfall in Vietnam under RCP4.5 and RCP8.5. They found that the TOE of annual and seasonal average temperatures generally started after 2018 for both scenarios. This implies that the selected period for comparison with observational data in our present study, 2006–2018, predates the time when the TOE could be detected for both RCPs. As a consequence, the impact of the chosen RCP on temperature changes could be indistinguishable from the impact of other factors mentioned above. This elucidates why, despite RCP8.5 depicting an extreme future scenario, its median model can still exhibit a closer proximity to the observational data points compared to RCP4.5 during the early-century period (2006–2018).

This study has successfully developed the SMME-VN dataset for future probabilistic projections of climate change in Vietnam, with a specific focus on temperature and rainfall. The dataset was constructed using the SMME method and the outputs of 31 CMIP5 GCMs. Each GCM was assigned a probability weight based on its estimated global temperature anomaly relative to the MAGICC6 temperature probability distribution. Furthermore, SMs were established to capture temperature probabilities in the extreme tails that were not accurately represented by the GCMs. The pattern scaling method was then applied to build the probabilistic model for Vietnam, i.e., the SMME-VN dataset, based on the probability of climate change on a global scale, corresponding to the 5th, 17th, 25th, 50th, 75th, 83rd, and 95th percentiles.

The SMME models demonstrate satisfactory performance in representing temperature and rainfall changes for the period 2006–2018, under both scenarios RCP4.5 and RCP8.5, when compared to the baseline period 1986–2005. Across the seven randomly selected stations, the observed changes generally fall within the SMME ranges. However, it is noticeable that changes in rainfall at three out of the seven stations, under both RCP4.5 and RCP8.5 scenarios, fall outside the SMME 5th–95th percentile range. This highlights the importance of the SMME ensemble's ability to represent extreme values that may occur in the tail of the distribution.

The projected SMME temperature for Vietnam, with a 5% probability of occurrence (corresponding to the 95th percentile), exhibits an increase of 5.41–7.13°C during the period 2080–2099 under the RCP8.5 scenario compared to the baseline period 1986–2005. The northern region is projected to experience more warming compared to the central and southern regions. SMME-VN projects a slight increase in average rainfall across the country, with the largest increase occurring in the central region, reaching a maximum of 14.2% at the 95th percentile. Regarding the sources of uncertainty in future projections, unforced uncertainty dominates the total signal during the early part of the 21st century. Forced uncertainty remains relatively moderate but increases gradually over time, reaching a maximum of 20–25% by the period 2080–2099. As we approach the end of the century, the proportion of unforced uncertainty decreases, giving way to uncertainty associated with greenhouse gas scenarios, which accounts for 75–80% of the total signal.

The SMME-VN dataset is accessible online and offers significant potential for diverse applications. We highly recommend employing this dataset for further studies, specifically in quantitatively assessing climate change risks and their impacts on socioeconomic activities in Vietnam. By benefiting from the high-resolution and probabilistic nature of the dataset, policymakers, researchers, and other users can gain valuable insights into a wide range of potential climate outcomes. This, in turn, facilitates more informed decision-making and effective addressing of the challenges posed by future climate change in Vietnam.

QT-A processed data and performed the analysis. Both TN-D and QT-A discussed the results and wrote the manuscript.

This study was supported by the Vietnam National Foundation for Science and Technology Development (NAFOSTED) under Grant 105.06-2021.14. QT-A was supported by the GEMMES project funded by the French Development Agency.

All relevant data are available from an online repository or repositories: https://remosat.usth.edu.vn/Download/dat_GEMMES_WP1/SMME_CMIP5/.

The authors declare there is no conflict.