Assessment of CMIP5 and CORDEX-SA experiments in representing multiscale temperature climatology over central India

The major factor that is responsible for drought is rainfall, but temperature also plays a significant role to find the severity of drought over a region (Barzkar et al. 2022). It is imperative to perform climate change studies of meteorological variables such as precipitation and temperature over the central India region, Bundelkhand, due to frequent drought events observed in the past decades (Vishwakarma et al. 2020). The latest-generation climate models are the key platform to predict the future where global climate models (GCMs) under the Coupled Model Intercomparison Project Phase 5 (CMIP5) have more accuracy in predicting future climate as compared to the older GCMs. Coordinated Regional Climate Downscaling Experiment (CORDEX) is also an upgraded platform that provides the regional climate data in terms of regional climate models (RCMs; dynamically downscaled GCMs) to make adequate climate predictions used in the impact studies. Limited ranked-based research has been performed to judge the best models applicable to the climate of central India. Mishra et al. (2014) checked the reliability of global and regional climate models for precipitation extremes over India. The study found that GCMs present more reliable weather than the RCMs of CORDEX over most parts of India. Some Host-GCMs (EC-EARTH and GFDL-ESM2M) and COREDEX RCMs (RegCM4-LMDZ and SMHI-RCA4) provided better pictures of the future climate over India. Before using CORDEX RCMs data, Mishra et al. (2014), Singh et al. (2017) and Mishra et al. (2018) also recommended bias-correcting them.

Singh & Goyal (2016) utilized the ESM-2M CMIP5 model for three RCP scenarios over the North Sikkim Eastern Himalayas. The study downscaled the precipitation using the ESM-2M model and found the patterns of extreme indices and lapse rate of precipitation over the study region. Singh et al. (2017) also analyzed the accuracy of nine CORDEX RCMs with their host CMIP5 GCMs for rainfall during the summer monsoon season in India from June to September. For all of the attributes of the Indian monsoon, the study revealed no consistent development in the RCM simulations concerning their host GCMs. It was also noticed that the ACCESS 1.0 and NorESM1-M GCMs simulated considerably acceptable monsoon climatology than the forced RCMs. Choudhary et al. (2018) similarly listed the important five CORDEX RCMs (GFDL-ESM2M-IITM-RegCM4, ICHEC-EC-EARTHSMHI-RCA4, LMDz-IITM-LMDz, LMDz-IITM-RegCM4 and NorESM1-M-CSIRO-CCAM) for India and two (GFDL-ESM2M-IITM-RegCM4 and ICHEC-EC-EARTHSMHI-RCA4) are the same models that were found as best in Mishra et al. (2014) study. Kundu et al. (2018) assessed the climate change and land-use effects on future evapotranspiration (ET) using a CMIP5 model, ACCESS 1.0, for the Narmada river basin in central India. A decrease in ET was seen in the future over the study area. Sinha et al. (2020) collected the future climate data from the five CMIP5 GCMs, ACCESS 1.0, CCSM4, CNRM-CM5, MPI-ESM-LR and NorESM1-M, under the CRODEX project for RCP 4.5 and 8.5 emission scenarios with daily time steps. This study analyzed climate variability and land-use effects on sediment yield and streamflow in a tropical mountainous river basin in South India's Western Ghats. Poonia et al. (2021a) also utilized the four best GCMs (ACCESS1-0, CCSM4, CNRM-CM5 and MPI-ESM-LR) under the CORDEX-SA experiments to find the effect of climatic variation on crop and irrigation water necessities over the eastern Himalayan zone. Besides climate model studies, Mehta & Yadav (2022) assessed the temporal trend of the climate of Jalore, Rajasthan, by utilizing the non-parametric trend approach using historical observation.

Hence, based on previous studies to inspect the precision of the models available for the South Asia region (especially for India), they are still unable to show the better accuracy of climate models for central India in signifying the temperature. Hence, five CORDEX RCMs (CCAM-CSIRO-ACCESS1-0, CCAM-CSIRO-CCSM4, GFDL-ESM2M-IITM-RegCM4, ICHEC-EC-EARTH-SMHI-RCA4 and MPI-M-MPI-ESM-MR-IITM-RegCM4) and their driving CMIP5 GCMs found best for India has been chosen for this study to apply for the central India region, i.e., Bundelkhand region. Various performance indicators such as average absolute relative error (AARE), agreement index (AI), coefficient of correlation (r²), mean absolute percentage of error (MAPE), Nash-Sutcliffe efficiency (NSE), root mean square error (RMSE) and scatter index (SI) are available to check the applicability of climate models (Najafzadeh et al. 2018). The present study utilized the worldwide-used NSE and r² indicators. One modified indicator, skill score (SS), has also been used to assess the model's performance accurately.

Daily gridded data of temperatures (maximum and minimum) for 24 years (1982–2005) have been collected from the NASA-POWER platform for 82 rain gauge stations of Bundelkhand. CMIP5 GCM datasets are accumulated from the Earth System Grid Federation (ESGF; https://esgf-index1.ceda.ac.uk/search/cmip5-ceda) and CORDEX South Asia (CORDEX-SA) RCM datasets from the Indian Institute of Tropical Meteorology (IITM; http://cccr.tropmet.res.in/home/data_portals.jsp). The descriptions of CORDEX-RCMs and their driving CMIP5-GCMs are shown in Table 1.

It is necessary to make all the observed daily data as well as the CORDEX RCMs and their forcing CMIP5 GCMs data on a common grid platform by using specific interpolation methods before using them. Climate Data Operators (CDO), a program developed by the Max Planck Institute, have been used to regrid all of the data. The scientific community utilized the CDO operator, which collects several algorithms for interpolation. Performance indicators, SS, r², and NSE have been used to assess the accuracy of the models.

The methodology involved in achieving the research objectives is described under the following steps;

There are five CORDEX RCMs and their driving GCMs (total of ten models) and two RCP scenarios as, RCP 4.5 and RCP 8.5, have been taken to prepare future data. Each climate data includes daily maximum and minimum temperatures for the periods from 1982 to 2005 (historical) and 2006 to 2100 (projected). The climate data includes a variety of grid resolutions. As a result, it is essential to prepare the data and resolve mistakes using regridding and bias corrections procedures before utilizing the projected data. Models data are first checked to ensure there are neither technical nor numerical bugs and to validate metadata integrity. Then, regridding or remapping is done by the CDO. Because the grid resolutions of CORDEX RCMs and CMIP5 GCMs differ, spatial interpolation of model data is required to arrange them on a reference grid. Numerous grid interpolation methods are available; the most appropriate method for the intended task should be chosen. Thus, daily maximum and minimum temperatures data of all ten-climate models are remapped to 0.5° × 0.5° resolution by utilizing the bilinear interpolation method.

When statistically compared to climatic observations, climate models have the ability to forecast future climate with systematic biases. Model simulations are calibrated using bias-correction methods to ensure that their statistical characteristics are equivalent to those of the corresponding actual results (observed historical data). Delta mapping, distribution mapping, linear scaling and quantile mapping methods are among the many ways of correcting climate model results. The empirical quantile mapping (QM) approach has been applied to bias-correct the climate model data, which effectively corrects systematic distributional biases.

The three performance indicators, r², NSE and SS have been used to assess the performance of climate models.

The efficacy of the five selected CORDEX-RCMs and their forcing CMIP5 GCMs for the study region have been analyzed using daily maximum and minimum temperature records gathered for 24 years (1982–2005). The efficacy of maximum temperature was assessed using the mean annual maximum temperature (MAMT), mean seasonal maximum temperature (MSMT) for monsoon months, June, July, August and September (JJAS), and mean non-seasonal maximum temperature (MNSMT). Similarly, the performance of minimum temperature has been examined using the mean annual minimum temperature (MAMIT), mean seasonal minimum temperature (MSMIT) for monsoon months (JJAS), and mean non-seasonal minimum temperature (MNSMIT). To find the ensemble, the arithmetic and weighted averages of RCMs and GCMs were assessed. The accuracy of the models was also assessed using their weighted average, calculated using the weightage from the SS test.

While testing using the r², all models showed the best correlation with observed data on all scales. On the annual time scale, NSE fulfilled all of the models with the best correlation value. Out of all the models, only GCM-CCSM4 and GCM-GFDL-ESM-2M have shown adequate correlation on the NSE scale for annual and seasonal timeframes. The NSE results also revealed that the bias-corrected GCMs performed good than the CORDEX RCMs, as shown in Table 3. In comparison, GCM-CCSM4 and GCM-GFDL-ESM-2M lie nearest to the average observed data.

Overall, In MAMT findings, there is the least fluctuation found in climate data from the observed ones. The majority of models have a high correlation with the mean maximum temperature measured. To claim the best model appropriate to the Bundelkhand, it is required to find the best one out of a group of models. The model's performance has also been examined using the SS method. SS has been applied on MAMT, MSMT and MNSMT scales of bias-corrected RCMs and driving GCMs. GCM models, CCSM4 and GFDL-ESM-2M have been found the best on all the time scales (MAMT, MSMT and MNSMT) for maximum temperature.

MAMIT findings are also observed to be appropriate for all models for future climate forecasts based on the NSE and r² reported in Table 4. All models have validated the best correlation with the observed data on all time scales in signifying minimum temperature. Especially, GCM-EC-EARTH, GCM-CCSM4, GCM-GFDL-ESM-2M, and GCM-MPI-ESM-MR have shown a good correlation on the NSE scale for annual and seasonal timescales. In comparison to CORDEX RCMs, bias-corrected GCMs performed well.

Overall, the MAMIT findings have shown minimal variation. The majority of models have been found closer to the average observed minimum temperature. To claim the best model appropriate for the Bundelkhand region for minimum temperature, it is required to identify the best one out of the selected models. Hence, SS has been used for all RCMs and their forcing GCMs on MAMIT, MSMIT and MNSMIT scales; as a consequence of the maximum temperature analysis, similar results have been obtained. As a result, bias-corrected GCMs CCSM4 and GFDL-ESM-2M have found the best models for minimum temperature on all time scales. The combined SS indicated a broad range of values across all of the models for all of the time scales. Table 5 shows the best maximum and minimum temperature models based on combined SS ranks reported on different time scales.

SS test concluded the best models as GCM-CCSM4, GCM-ESM-2M and GCM-ESM-MR based on the best ranks presented in Table 5 on various time scales. Ensembles of both the categories (arithmetic and weighted average) could not clear the efficacy of models for both GCMs and RCMs for central India.

There were minimal variations found in mean annual maximum as well as minimum temperatures of observed and modelled data. The majority of models have been found close to the observed maximum and minimum temperatures on an annual mean scale. Based on the NSE performance indicator, most of the bias-corrected GCMs have shown satisfactory performance over the CORDEX RCMs for both maximum and minimum temperatures. The performance of models has also been examined by the SS method on multi scales (annual, seasonal and non-seasonal). GCM-CCSM4, GCM-ESM-2M and GCM-ESM-MR have claimed the best ranks of SS for all the time scales. Arithmetic and the weighted ensemble of the models have also shown a poor representation of climate signifying temperature. The study revealed that the GCM models, CCSM4, GFDL-ESM-2M and GFDL-ESM-MR could predict maximum and minimum temperatures with greater accuracy for the Bundelkhand region.

The study also concluded that the performance of the models could be best judged by their spatial representation on multiscale as well by NSE and SS performance indicators. It was also revealed that bias-corrected GCMs outperformed CORDEX RCMs in predicting maximum and minimum temperatures across the Bundelkhand region, India. Moreover, the study observed the better accuracy of climate models on mean monsoon observations. The results of best-suited latest generation climate models would also help policymakers to forecast the frequency and severity of future drought or flood and future water requirements for agriculture. The study has opted for latest-generation climate models under the CMIP5 and CORDEX-SA experiments; those are the improved version of earlier CMIPs and RCMs, respectively. Hence, the study will be beneficial for the effective and accurate prediction of the climate of central India.

Some other performance measurement methods such as BIAS, mean absolute percentage of error (MAPE) and root mean square error (RMSE) could also be applied to proposed models to find their adequacy (Najafzadeh & Saberi-Movahed 2019; Najafzadeh & Oliveto 2020). Based on some recent investigations, the projected climate data could also be utilized against the analysis of drought risk management and flash drought analyses (Poonia et al. 2022) over the study region. The blockchain technology-based framework could be used to improve the drought risk managing system (Poonia et al. 2021b). Most recent copula-based analysis (Poonia et al. 2021c) could also be utilized to characterize the drought duration and severity over the study region.

The authors acknowledge the Indian Meteorological Department (IMD) for providing the necessary climate data to conduct this research work.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.