Multimodal climate change prediction in a monsoon climate

Climate change assessment is essential to strategically plan the various resources under different climate regions. Climate change prediction in a monsoon type of climate is challenging for the following reasons: (i) the rainfall distribution over the year is very much skewed to monsoon; (ii) the rainfall is primarily influenced by the development of a climate in the surrounding ocean; and (iii) the temperature is an essential variable in the climate change. The problem becomes more complex with high variability in the output of different climate models. The reliability of a single model is debatable. The increase in global surface temperature in the 21st century is mainly attributed to anthropogenic greenhouse gas (GHG) emissions (Wilby et al. 2002; IPCC Climate Change 2014; IPCC 2021). In several parts of the world, extreme weather conditions with abrupt changes in the atmosphere are experienced (IPCC Climate Change 2014; IPCC 2021). An increase in the total precipitation on land, frequent climate extremes (such as heatwaves, droughts, floods, cyclones, and wildfires), the overall rise in surface temperature, shift in the rainfall pattern and its distribution, and mean sea level rise are some of the major issues which are being faced today, due to climate change.

The future climate mainly depends on the amount of anthropogenic GHG emissions and the response of the climate system to these emissions (Lowe & Jenkins 2003). Several organisations have come together to quantify this climate change based on the IPCC projection, known as global climate models (GCMs). GCMs are the primary tool for accurate prediction and demonstration of the possible future impacts of climate change to the policymakers due to increasing various GHGs and anthropogenic events without the need for expensive laboratory experimental setup (Fowler et al. 2007; Maraun et al. 2010; Dastagir 2015).

The Intergovernmental Panel on Climate Change (IPCC), in its sixth assessment (AR6) report, has projected eight possible emission scenarios based on a component of climate and its interaction and range of socioeconomic factors (such as demography and economic growth), namely, SSP119, SSP126, SSP435, SSP534OS, SSP245, SSP460, SSP370, and SSP585. SSP370 and SSP585 correspond to high and very high GHG emissions, with double CO₂ emissions from current levels by 2100 and 2050, respectively. SSP245 is an intermediate GHG emission scenario. SSP119 and SSP126 scenarios are based on varying net negative CO₂ emission levels. The AR6 report is the successor of the AR5 report and provides a better assessment of climate projection. Hamed et al. (2022) compared the output of CMIP6 to CMIP5 and concluded that the CMIP6 ensemble displays less uncertainty in the simulation than CMIP5.

The findings of the IPCC in its AR6 highlight several aspects of climate change all over the globe. The global averaged annual surface precipitation is likely to change by −0.2 to +4.7% during 2081–2011 relative to 1995–2014 in the low-emissions scenario SSP1–1.9 and 0.9–12.9% in the high-emissions scenario SSP585 (IPCC Climate Change 2021). Although the confidence in precipitation change averaged over the global land area is low due to insufficient data, a shift in the rainfall pattern and increased extreme weather conditions have been reported. Droughts with greater magnitude, severity and longer duration (since the beginning of the 20th century, rise in the number of heavy precipitation events (post-1950)) are some of the critical findings related to precipitation and temperature reported in IPCC reports with high confidence.

The spatial resolution of output from GCMs is in the range of 100–250 km, which is quite large for the regional assessment of the climate of a small region. Regional-scale studies require high-resolution climate projections. The regional climate change projection can only be assessed after downscaling the GCM output to the local region (Hay et al. 2000; Wilby et al. 2000; Lowe & Jenkins 2003; Wood et al. 2004; Maraun et al. 2010; Wetterhall et al. 2012). Apart from the regional and global scale mismatch of the climate variables, these models show biased representations of observed time series (Teutschbein & Seibert 2012). On the temporal scale, the GCM output is available at the hourly to annual level. The reliability of GCMs decreases with finer temporal scales (Anandhi et al. 2011). The uncertainty in the near-future projection is high due to natural variability, model uncertainty, and uncertainty in natural and anthropogenic aerosol forcing (IPCC 2021). The problem becomes more complex due to considerable scatter and differences in the sign of projected climate change from a set of models as well as the same model with different downscaling techniques (Mearns et al. 1999; Lowe & Jenkins 2003; Vavrus et al. 2015). Different models may differ greatly on the projected magnitude of the change, which raises questions about the reliability of a model. It is mainly due to the high variability in the interaction between different components of different climate models. The uncertainty in the projection can be due to several factors, such as the emission of various gases, transfer function relating these emissions to the abundance of forcing; climate sensitivity to external forcing; and ocean mixing (Wigley & Raper 2001).

Several researchers have advocated multimodal climate data to quantify uncertainty (Singh et al. 2016). The arithmetic average or multimodal mean among a set of model simulations is a crucial metric to quantify the uncertainty and increase the reliability of the climate projection (Diaz-Nieto & Wilby 2005; Chen et al. 2015; Vavrus et al. 2015). However, simple arithmetic averaging of the projection of different models may nullify the overall impact if there is a significant difference in the projection results. In the context of multiple GCMs where the variability of GCMs projection is high, some researchers have proposed a sign change method in which the projection is estimated by the dominant sign change from various models. However, these averaging methods do not provide a reliable projection. Presently, two methods are widely used in multimodal climate assessment: Bayesian model averaging (Hoeting et al. 1999) and reliability ensemble averaging (REA; Mearns et al. 1999; Giorgi & Mearns 2002, 2003). In the probabilistic approach, the weight of each model is determined based on the historical relationship between forecasts and observations, which enhances the performance of the ensemble compared to a simple average where each model is weighted equally (Krishnamurti et al. 2000). In the REA, the weightage for each model is calculated based on model bias in replicating the present-day climate and the variations in the projection of different models. Xu et al. (2010) extended the REA method by introducing the interannual coefficient of variation (CV) and spatial correlation to the original REA. The multimodal ensemble techniques reduce the individual model uncertainty significantly (Yang et al. 2018). The weightage of different GCMs can be tuned to have a reliable estimate of climate. Yang et al. (2018) applied an evolutionary multi-objective optimisation algorithm to optimise the weightage of each GCM such that the weightage ensemble mean is close to the observed data.

Multimodal climate studies have been conducted on various regions under various climatic conditions. However, ensemble averaging becomes difficult in a monsoon climate due to its skewed behaviour. India has a monsoon climate where most precipitation occurs in less than 3 months. The performance of GCMs under such climatic conditions is highly diverse. Several studies have been conducted to assess the climate change impact in a monsoon region. Kothawale & Kumar (2005) reported an increase in India's mean annual temperature by 0.22 °C/10 years between 1971 and 2003. There is a major shift in the seasonal asymmetry of temperature trends, with no region or season showing a cooling trend. The projections of monsoon precipitation over South Asia and East Asia show an increase in precipitation frequency and intensity (Jourdain et al. 2013; Chevuturi et al. 2018). Singh et al. (2016), in their work on the Indian regional (Roorkee) climate, projected an increase in precipitation intensity (7–96% depending upon the RCP scenario, return period, and climate model) for all the return periods and duration. The high variability in projected precipitation intensity for different models in a particular RCP and return period make it difficult for policymakers to make strategies. Iqbal et al. (2020) analysed 22 CMIP5 GCMs in the sub-Himalayan region of Pakistan and found significant heterogeneity in precipitation change for different future periods and RCP scenarios.

A comparative study conducted by Dutta et al. (2022) concludes that the latest CMIP6 models improve the model bias in the Indian subcontinent and are in better agreement with the observation. A multimodal climate study over the Tibetan plateau using 25 GCMs in the CMIP6 framework by Chen et al. (2022) predicts an overall increase in the precipitation by about 23% in autumn. Liu et al. (2022) studied the capability of the CMIP6 model on seasonal precipitation over Central Asia and concluded that most models have a consistent bias in extreme precipitation and have high variability in future projections. Pimonsree et al. (2023) conducted a study in the Southeast Asian region based on the CMIP6 framework. They concluded that the performance of most of the models is not good in the mountainous region. Anil & Raj (2022) applied multimodal averaging of selected CMIP6 GCM and found the performance of the REA is satisfactory in mimicking the observed precipitation in the Krishna basin of the Indian subcontinent. The study also shows an overall rise in precipitation with a higher impact in the far-end future in the region.

In this paper, a multimodal approach has been developed and adopted to handle the uncertainty arising from different climate models and, thus, predict changes in precipitation. In multimodal analysis, the statistical parameters of several models are considered, and a weightage is given to each model based on its performance. This performance is based on two parameters, namely, model bias and model convergence. The weightage average of each model is finally used to compute a multimodal average projection. Climate change prediction was carried out for precipitation based on IPCC's AR6 report. Twenty-six GCMs, with four Shared Socioeconomic Pathways (SSPs), namely, SSP126, SSP245, SSP370, and SSP585, are considered on a seasonal scale. The present study can be used for long-term planning based on the projected change in annual precipitation frequency and change in temperature.

In the present study, the AR6 of the IPCC is used to perform the climate change impact assessment. The climate model under four SSP scenarios, namely, SSP126, SSP245, SSP370, and SSP585, is considered. The work of Giorgi & Mearns (2003) on REA is used to bias correct and quantify the uncertainty in the climate projection of precipitation and temperature. The data and methodology adopted in the present work are discussed in the subsequent sections.

The 26 GCMs used in the present study for the projection of precipitation are listed in Table 1. The realisation of each of the models is r1i1p1 for the four selected scenarios, namely, SSP126, SSP245, SSP370, and SSP585. The future climate projection of each model is classified into two categories, near-end future (2021–2050) and far-end future (2051–2080), and the change is compared with the historical variables. The GCMs output was taken from different models at the nearest grid location available. These outputs were interpolated using the inverse distance method to the study area. The distances were calculated using the Haversine equation, considering the curvature of the Earth.

The upper and lower limit will be ⁠.

The data used, and the areal characteristics of all four cities, are shown in Table 2. Precipitation data from 1901 to 2010 obtained from IMD was first checked for any data errors and missing data. Any missing data were filled using interpolation and statistical correlation methods.

Table 2 shows the mean precipitation of the selected location between 1971 and 2000 and the natural variability in the precipitation from the 30-years moving average. From the precipitation data of 110 years for all four cities, the 30-years moving average series was prepared. The linear trend in the series was removed, if any. The detrending was done by removing the linear trend from the 30-years moving average data series. The slope and intercept for a linear equation, fitted to the 30-years moving average series, were calculated. Later, the values from the linear equation at each data point from the 30-years moving average were removed. The difference between the maximum and minimum values of the detrended data was then used to calculate the natural variability in the precipitation. It can be observed that most of the rainfall is received during monsoon and pre-monsoon seasons, except for Coimbatore, which also receives rainfall in the post-monsoon season. Udaipur, Rajkot, and Siliguri receive most of the rainfall during June to September, while in Coimbatore, most of the rainfall is received during September and November. Rajkot has higher monthly rainfall variability compared to other cities. Out of the four cities, Siliguri receives the highest annual rainfall.

The natural variability in the 30-years moving average for precipitation is calculated from 100-years monthly data after detrending the time series. The natural variability in Coimbatore is relatively high compared to that of other cities, which indicates a high variation in the precipitation pattern. On a seasonal scale, the natural variability in winter is comparatively higher than in other seasons due to the skewed nature of the precipitation pattern. Hence, it is likely that the performance of the model will be good in the winter season and for Coimbatore. Coimbatore has natural variability of 61.6 mm with a mean seasonal precipitation of 365 mm in post-monsoon, much less than in other seasons. For all the other cities, Udaipur, Rajkot, and Siliguri, the natural variability in the monsoon season is significantly less compared to other seasons.

The relative change in precipitation is estimated based on the difference in mean seasonal precipitation from the future projected rainfall to the baseline scenario, with 1971–2000 taken as the baseline period. The bias factor is in the range of 0–1, with 1 being no bias in the climate model. Hence, the higher the bias factor, the better the model. For example, if a climate model has a bias factor value of 0.6, the climate model has a 0.4 or 40% bias from the observed mean variable. Since the bias factor depends only on the performance of a model in simulating the present-day climate, another index, CV, has also been used for assessing the performance of each model. The prediction of precipitation for near-end and far-end futures has been carried out, and the results are compared with the baseline scenario.

The key criteria of GCM's performance are model bias. Any model which can replicate the present-day climate with the least systematic bias is generally considered a good GCM and likely to offer reliable future projections (Liu et al. 2022).

In Coimbatore, most of the models perform better in all seasons relative to other cities. AWI-CM-1-1-MR, EC-Earth3-Veg-LR, and IPSL-CM6A-LR are the best-performing models for all the seasons with no bias factor, as shown in Figure 2. CMCC-CM2-SR5, INM-CM4-8, and INM-CM5-0 models perform poorly in winter and post-monsoon seasons with bias factor values less than 0.3.

The performance of all the models except for IITESM (in winter) and CAS-ESM2-0 (in post-monsoon) shows no bias in winter and post-monsoon seasons in Udaipur. However, the bias factor is quite high in pre-monsoon and monsoon. For monsoon, except for IPSL-CM6A-LR and AWI-CM-1-1-MR, all the models have a significant bias with a bias factor of 0–0.8. The performance of models in Rajkot is similar to Udaipur, where in post-monsoon, except for CAMS-CSM1-0, all models have no bias. In pre-monsoon, the model performance is very poor, with relatively better performance in the monsoon season. AWI-CM-1-1-MR, BCC-CSM2-MR, CMCC-CM2-SR5, and CMCC-ESM2 models have no bias in simulating the present-day climate. The bias factor for Udaipur and Rajkot shows some similarity, mainly due to similar climatic conditions. The rainfall pattern of Siliguri is similar to Udaipur and Rajkot, but the amount of precipitation is quite high in all the seasons. Most of the models perform better in post-monsoon in Udaipur, Rajkot, and Siliguri. The model bias in winter and post-monsoon seasons is negligible in all the models and hence, performs well. However, the performance of models is very poor in simulating the pre-monsoon monsoon climate. Only two models, CMCC-CM2-SR5 and CMCC-ESM2, have no bias in any season for Siliguri.

In terms of overall model performance, it can be observed that the performance of most of the models is very poor in replicating monsoon precipitation, except for Coimbatore. The precipitation pattern for Coimbatore is comparatively less skewed, resulting in better performance of models in the monsoon season. Similarly, the model performance, in terms of CV, was also found to be poor in regions with high variability. Out of all the models tested here, GCMs, namely AWI-CM-1-1-MR, BCC-CSM2-MR, CMCC-CM2-SR5, and EC-Earth3-Veg-LR, show promising results in several seasons at certain regions considering both the aspects of precipitation, model bias and CV.

In Coimbatore, the pre-monsoon season has little impact on climate change, as shown in Figure 6(a). The pre-monsoon of Coimbatore is expected to have a marginal change in precipitation with a rise in the mean by 7–10 mm in SSP126 and SSP585 scenarios and a negative change in SSP245 and SSP585 scenarios. A similar trend is observed in the post-monsoon season as well, with an expected rise in mean rainfall by 10–20 mm. The monsoon rainfall is expected to rise by about 40 mm under all the scenarios in Coimbatore, with a high variability of 40–50 mm. The pattern in the change of precipitation for Udaipur and Rajkot is similar. The pre-monsoon impact is minimal for Udaipur and Rajkot, as shown in Figure 6(b) and 6(c), respectively. The REA precipitation in Udaipur and Rajkot is expected to rise by 40–50 mm. The rise is comparatively higher in the SSP126 scenario. It should also be noted that the variability in the change for Udaipur and Rajkot is quite high, in the range of 50–60 mm, while the present mean monsoon seasonal precipitation of Udaipur and Rajkot is 80 and 100 mm, respectively. The uncertainty range is also skewed on the positive side, suggesting a possible rise in monsoon precipitation. The post-monsoon projection shows a possible rise in mean precipitation by 10 mm under all the scenarios for both cities. In Siliguri, the pre-monsoon and monsoon seasons show a rise in precipitation, as shown in Figure 6(d). The monsoon rainfall is expected to rise by 20–30 mm under various scenarios, which is insignificant compared to its monsoon precipitation of 1,725 mm. The pre-monsoon rainfall is expected to rise, ranging from 0 mm under SSP126 to 50 mm in SSP585.

The seasonal mean of precipitation in each season in all the regions varies acutely. On a regional scale, the impact is higher for Coimbatore and Siliguri in winter, and for Udaipur and Rajkot, the impact is more in the post-monsoon season. In overall percentage, the winter precipitation in Coimbatore and Siliguri is likely to fall by 13%, and the post-monsoon precipitation in Udaipur and Rajkot is likely to rise by 24 and 26%, respectively. It should also be noted that the change in the monsoon rainfall, as shown in Figure 6, is higher than in any other season. Since the rainfall pattern is mostly skewed to the monsoon season, the change in precipitation in the non-monsoon period is very smaller compared to the monsoon rainfall. However, in terms of percentage change from the mean seasonal rainfall, the non-monsoon percentage change is quite high.

In Coimbatore, the mean seasonal rainfall in winter is likely to fall in the far-end future, with a maximum fall of 5–7 mm under SSP245 and SSP370 scenarios and a negligible impact under SSP126 and SSP585. In pre-monsoon, the mean precipitation is likely to rise by 10–25 mm with a maximum rise under the SSP585 scenario. A similar trend is observed in post-monsoon with a rise in the mean seasonal precipitation by 32–54 mm. In pre-monsoon and post-monsoon, the impact is minimum under SSP370 and maximum in SSP585. The monsoon precipitation of Coimbatore is likely to rise by 54–81 mm with a maximum rise under SSP585 and a minimum in the SSP126 scenario.

Udaipur and Rajkot show a rise in precipitation in the monsoon and post-monsoon seasons, as shown in Figure 7(b) and 7(c), respectively. The mean monsoon precipitation is likely to rise by 68–79 mm, and the post-monsoon precipitation may rise by 10–14 mm. In monsoon, the impact is minimum under the SSP370 scenario with a mean rise of 68 mm and maximum under scenario SSP245 with a mean rise of 79 mm. The mean precipitation rise in monsoon for Rajkot is likely to rise by 77–100 mm, maximum in the SSP585 scenario. Post-monsoon precipitation of Udaipur and Rajkot is likely to rise by 16 mm in the SSP585 scenario. The far-end change pattern of precipitation in Siliguri is similar to what was observed in the near-end, where SSP126 has the least impact and SSP585 has the maximum impact.

The seasonal rise in monsoon is expected to be about 80–185 mm, and the pre-monsoon impact is around 30–102 mm. At a regional level, the far-end impact is mainly in Udaipur and Rajkot, with about 30 and 40% rise in winter and post-monsoon precipitation. Rajkot is likely to be the severely impacted region with the rise in precipitation in all the seasons in the far-end future.

The above results indicate that the monsoon precipitation is likely to increase in all the cities. Hence, it can be expected that the southern peninsula is likely to receive more rainfall in the post-monsoon season. The eastern region of the Indian subcontinent is likely to receive more pre-monsoon rainfall. The projected precipitation indicates that Coimbatore, Udaipur, and Rajkot will show a rise in post-monsoon rainfall. This may indicate an early arrival of the monsoon in the eastern region and delayed rainfall in the southern and western regions. On a spatial scale, the impact is higher for Udaipur and Rajkot than Siliguri and Coimbatore, based on the expected rise and the current mean precipitation.

Model uncertainty is a key challenge in climate change modelling. Several GCMs show a considerable variation in the projected climatic variable, especially in a monsoon climatic condition. Hence, there is a need for the selection of a suitable GCM or ensemble of GCMs to address the uncertainty in various GCMs.

It is found that many of the GCMs may not work well for the skewed distribution of rainfall, and hence, the REA method using a weighted bias approach has been adopted for a reliable prediction of climate change. The study shows that the REA method minimises the uncertainty in various GCMs when it was tested for four stations located with high climate variability, based on the AR6 framework. Further, it was also found that the performance of any GCM is largely linked to the ratio of natural variability to the mean, with an R² value of 0.78. This index may be used as a potential tool to choose the number of models for further climate studies in other regions.

The uncertainty band from the multimodal analysis does suggest a considerable variation in the projected rainfall pattern, especially in the monsoon period, and is highly inconsistent spatially. Further studies involving different variables can provide a better understanding of model uncertainty in a monsoon climatic condition. In the present work, the model averaging was carried out by weighing each model based on the model bias from the observed mean and model variability from the average of all models. The work may be extended by considering other statistical parameters, such as standard deviation, correlation etc., in the model bias.

All relevant data are available from an online repository or repositories (https://esgf-node.llnl.gov/search/cmip6 and https://imdpune.gov.in/library/publication.html).

The authors declare there is no conflict.