Comparison of CMIP5 and CMIP6 models for temperature and precipitation simulation over the Yarlung Tsangpo-Brahmaputra River Basin

Global warming has been recognized as the primary determinant of increased risk from global climate change (Su et al. 2021). In recent decades, the average global surface temperature has risen by about 0.85 °C (Zhang et al. 2019). Moreover, the temperature may increase by approximately 1.5 °C in the next 5 years (Pereira et al. 2021). Related studies have suggested that global warming will intensify the water cycle, increasing the intensity of extreme precipitation events and risk of floods (Papalexiou & Montanari 2019; Tabari 2020). Gründemann et al. (2022) reported that increasing global warming may increase rare daily extreme precipitation more than common extreme precipitation, increasing the difficulty of future extreme event defense efforts.

The Yarlung Tsangpo-Brahmaputra River (YBR) is an important international river in South Asia. Owing to its complex topography and climate, the characteristics of climate change are apparently different in space. Since the temperature on the Tibetan Plateau (TP) is rising twice as fast as the global average temperature (Wang et al. 2022), the Yarlung Tsangpo-Brahmaputra River Basin (YBRB), which is extremely sensitive to climate change, is severely impacting society and the economy in response to global warming, more significantly than the TP (Song et al. 2010). In YBRB, where climate heterogeneity is prominent and the ecological environment is extremely fragile, climate change has had a huge impact on the relationship ‘water-energy-food’ (Keskinen et al. 2016). In recent years, with the warming of the climate, the frequency of extreme precipitation events and floods in the humid areas in the east of YBRB and the downstream floodplain (FP) has increased (You et al. 2007; Debnath et al. 2023). It threatens the agriculture in YBRB, especially on the FP (Debnath et al. 2023). Meanwhile, as climate change becomes increasingly apparent, the ‘water wars based on climate change’ may further increase and lead to social unrest and national conflicts (Keskinen et al. 2016; Klare 2020). Temperature and precipitation assessments are the basis for studying distinguishing features and future changes in extreme events (Pereira et al. 2021). Therefore, studying the spatiotemporal features of temperature and precipitation in the YBRB can provide an important basis for further study regarding extreme events and theoretical support for natural disaster defense planning in the region.

Generally, studying the spatiotemporal changes in temperature and precipitation using station observation data is effective and reliable; however, the uneven distribution of stations and data limitations in many regions with complex topography (Lun et al. 2021), such as the YBRB, affect the feasibility and credibility of the studies. Recently, global climate models (GCMs) have been widely used to study climate changes and simulations, particularly for temperature and precipitation. The coupled model intercomparison project (CMIP), organized under the auspices of the World Climate Research Programme (WCRP), is now in phase 6 (CMIP6) and is dedicated to providing standard climate simulations and outputs (Taylor et al. 2007) to facilitate climate change and assessment studies. Compared to previous phases (coupled model intercomparison project phase 5, CMIP5), CMIP6 pays more attention to internal climate change, predictability, and uncertainty in various scenarios and focuses on climate change in the geosphere (Lin & Chen 2020; Touzé-Peiffer et al. 2020). However, existing studies have shown that GCMs still suffer from large systematic biases and uncertainties at the regional scale (Wang et al. 2022), which may be related to their regional climate applicability (Xin et al. 2020). The capability of each model differs in different regions for different variables (Zhu & Yang 2020), particularly for precipitation simulations (Jiang et al. 2016).

Whether the updated CMIP6 has reduced model uncertainties, improved simulation and prediction reliabilities at different scales, and is more powerful than CMIP5 are questions that need to be addressed (Chen et al. 2020). Wang et al. (2022) also showed that understanding the systematic bias of GCMs is a prerequisite for the accurate prediction of the hydrological cycle at the watershed scale. Therefore, it is crucial to evaluate and compare the reliability and suitability of CMIP6 and CMIP5 in climate simulations (Eyring et al. 2019). Since using GCMs with a single model may lead to errors in the results, most applications of GCMs use the multi-model ensemble (MME) method to fuse information from multiple models. The results of the MME on the TP are more reliable than those of a single model, as it effectively reduces the model's uncertainty (Luo et al. 2020; Lun et al. 2021). Among MME methods, the generalized three-cornered hat (TCH) method does not require known real observations for the ensemble model data (Premoli & Tavella 1993). However, few studies focused on the MME of GCMs using TCH; therefore, a comparative validation is required before using TCH to conduct relevant studies, which can also reduce the limitations of insufficient observational data.

Several studies have compared the simulation abilities of CMIP5 and CMIP6 on the TP (Zhu & Yang 2020; Lun et al. 2021; Hu et al. 2022); however, few studies have used CMIP6 models for the entire YBRB, leading to the cognitive limitations and uncertainties. Additionally, since the performances of the models and variables in different regions may be inconsistent, it is essential to evaluate and compare CMIP5 and CMIP6 data and reduce uncertainty using MME methods before the study implementation. So far, there are no comparisons between CMIP5 and CMIP6 in YBRB, and no MME and uncertainty analyses utilizing the TCH method based on literature review. Therefore, the purposes of this study are to compare and assess the simulation ability of 13 homologous models in CMIP5 and CMIP6 for temperature and precipitation in the YBRB and to determine whether the MME by TCH is more reliable than a single model for temperature and precipitation in the YBRB. First, TCH was used to calculate the temperature and precipitation uncertainties for CMIP6 and CMIP5 in the YBRB and to determine the MME results. Second, the bias, GCM uncertainty (GU), signal-to-noise ratio (SNR), standard deviation (SD), root mean square error (RMSE), anomaly correlation coefficient (ACC), and temporal correlation coefficient (TCC) were calculated for simulation capability, to compare CMIP6 with the corresponding CMIP5 for reproduction of spatiotemporal temperature and precipitation in the YBRB. Finally, Taylor diagrams were used for comprehensive assessment.

Affected by the southwest monsoon and cyclones in the Bay of Bengal, precipitation in the YBRB is spatiotemporally different. The distribution of annual precipitation over YBRB is extremely uneven, with 60–70% concentrated in July–September (Xu et al. 2019), and has apparent differences in dry and wet seasons (He 2021). The spatial distribution of precipitation in the basin varies significantly owing to the influence of topography and climate. A significant difference exists between maximum and minimum temperatures (Tasmax and Tasmin) during a year (up to 40 °C) and upstream and downstream region temperatures. Previous studies have shown that the YBRB is more susceptible to climate change than other regions (Pervez & Henebry 2015) and that the hydrological processes in the YBRB are more sensitive to changes in temperature and precipitation (Xu et al. 2019). Therefore, evaluating the reliability of models before estimating regional changes in temperature and precipitation is necessary.

Affected by the terrain and elevation, different climate types exist in different regions of the YBRB (Guo et al. 2022). In this study, the YBRB was divided into following three different geographical units that respond differently to climate change based on Immerzeel's division scheme (Figure 1(b)): TP (≥3,500 m, 44.4%), Himalayan belt (HB, ≥ 100 m and <3,500 m, 28.6%), and FP (<100 m, 27%) (Immerzeel 2008).

Considering the availability of model data and the frequency of their use by others (see Table S1 in the Supplementary Materials), 13 homologous models of CMIP5 and CMIP6 were selected under the initial conditions of r1i1p1 (CMIP5) and r1i1p1f1 (CMIP6). The variables included daily Tasmax, Tasmin and precipitation from 1961 to 2005. The model data were downloaded from https://www.wcrp-climate.org/wgcm-cmip. Table 1 shows the specific information. To facilitate evaluation and comparison, the model data were resampled. According to existing studies, GCM data were resampled through bilinear interpolation (Kim et al. 2021; Li et al. 2021; Allabakash & Lim 2022; Xie et al. 2022). Bilinear interpolation is a linear interpolation along the x- and y-axes that provide the value of the point to be interpolated (Allabakash & Lim 2022; Vishwakarma et al. 2022). Since the bilinear interpolation algorithm is simple and relatively easy to implement and has better interpolation quality and higher accuracy (Shi et al. 2014), it was used to resample each model's data to 0.25° × 0.25°, which was consistent with the spatial resolution of the reference data.

In this study, the gridded temperature dataset derived from PGF-V3 data (Princeton Global Forcings, https://hydrology.soton.ac.uk/data/pgf/v3/) was employed to assess the GCM temperature outputs, including Tasmax and Tasmin, and the corrected Asian Precipitation-Higher Resolved Observational Data Integration Towards Evaluation of Water Resources (APHRODITE) precipitation data (http://aphrodite.st.hirosaki-u.ac.jp/download/) were used to assess the GCM precipitation outputs. The PGF-V3 data were constructed by combining a global observation-based dataset with the NCEP/NCAR reanalysis dataset at a resolution of 0.25° × 0.25° (Sheffield et al. 2006). APHRODITE data are long-term daily gridded data of Asia obtained from ground-based rain-gauge observations on a continental scale since 1951, with a resolution of 0.25° × 0.25° (Sunilkumar et al. 2019). However, the original PGF-V3 temperature and APHRODITE precipitation data were biased in the YBRB and needed correction. Ji et al. (2020) use the delta method to correct the systematic bias in PGF-V3 temperature data and use linear scaling and cumulative distribution function combination (LS_CDF) methods to correct APHRODITE data, both of which yielded better performance. Over the YBRB, the precipitation bias changed from −29.69 to 4.6%, RMSE decreased from 13.17 to 11.97 mm, the correlation coefficient increased from 0.80 to 0.83, and the accuracy of the spatial pattern was 0.98, with higher reliability (Ji et al. 2020). Therefore, corrected PGF-V3 and APHRODITE data (Ji et al. 2020) were chosen as reference data to evaluate the performance of the GCM data in this study.

TCH is a method used to assess the uncertainty of multiple sets of data (≥ 3) and does not require known real observational data (Premoli & Tavella 1993). Currently, TCH is also used in data fusion; for example, Xu et al. (2020) applied TCH to an ensemble of global precipitation reanalysis data and showed that the ensemble was superior on average and the weighted precipitation data significantly reduced the random error. Furthermore, TCH has also been successfully used for uncertainty analysis and fusion of the Gravity Recovery and Climate Experiment (GRACE) products (Cui et al. 2022). In light of this, the present study has also endeavored to utilize the TCH method for analyzing the uncertainty of GCMs as well as for calculating MME.

To compare and evaluate the ability of CMIP5 and CMIP6 single models to predict Tasmax, Tasmin, and precipitation in the YBRB and to check whether there is higher confidence in the TCH-MME than the single model, SNR, bias, SD, RMSE, ACC, and TCC were calculated in this study, and where: denotes the observed value, denotes models, denote the means of observation and model data, = ⁠, ⁠, are the spatial averages of the observation and model data, ⁠, is the latitude in which the grid is located:

Regarding the GU of CMIP5 models, NorESM1-M had the largest GU in summer precipitation (9.15 mm) and ACCESS1-0 had the smallest GU in winter precipitation (1.31 mm). INM-CM-4 showed the highest degree of GU fluctuation (upper quartile minus lower quartile) for Tasmax in autumn (1.79 °C), Tasmin in summer (2.80 °C), and precipitation in summer (8.92 mm). MPI-ESM-LR for Tasmax in autumn and MRI-CGCM3 for Tasmin in summer showed the least fluctuations, both were 0.21 °C, in the range of 2.58–2.80 °C and 1.31–1.54 °C, respectively. For the SNR of CMIP5 models, the largest was noticed in INM-CM-4 for Tasmax in summer (16.46), while BCC-CSM1-1 for precipitation caused the smallest and negative SNR (−6.77), and the largest fluctuation was caused by MIROC5 for Tasmin in summer (19.90), and the simulation of precipitation in winter by ACCESS1-0 showed the smallest fluctuation (0.97 mm).

For CMIP6, BCC-CSM2-MR exhibited the largest GU (17.45 °C) for Tasmin in autumn, INM-CM5-0 showed the smallest GU (0.88 °C) for Tasmin in summer, GFDL-ESM4 showed the largest GU fluctuation (13.18 mm) for precipitation in summer, and INM-CM5-0 had the smallest GU fluctuation (0.15 °C) for Tasmax in spring. For SNR, INM-CM5-0 showed the largest SNR (23.50) for Tasmin in summer, BCC-CSM2-MR showed the smallest and negative SNR (−3.23) for Tasmin in autumn, and MPI-ESM1-2-LR showed the largest fluctuation (18.72) for Tasmin in summer, and FGOALS-G3 showed the smallest fluctuation (0.67) for precipitation in autumn.

Many CMIP6 single models showed improvements relative to the homologous CMIP5 models, but some differences exist among the models (Figure S1 and S2). For example, the GU and SNR of BCC-CSM2-MR for the Tasmax and Tasmin simulations in the YBRB were inferior to those of BCC-CSM1-1 in CMIP5; MIROC6 and GFDL-ESM4 for the precipitation simulation of precipitation were inferior to those of MIROC5 and GFDL-ESM2G; and INM-CM5-0 for simulations of annual mean Tasmax and seasonal mean Tasmax and Tasmin showed improvements. The reduction in GU for Tasmax and Tasmin ranged from 0.14 to 1.77 and 0.14 to 4.58, respectively, while the SNR improvement for Tasmax and Tasmin ranged from 0.35 to 2.77 and 0.85 to 11.12, respectively. For precipitation, except for average annual precipitation, NorESM2-LM improved significantly, with the reduction in precipitation GU ranging from 0.88 to 3.09.

In CMIP5, the ACC means of TCH-MME5 for Tasmax, Tasmin, and precipitation were 0.5, 0.29, and 0.25, respectively, which were almost all better than those of the individual models. Furthermore, the ACCs of Tasmax, Tasmin, and precipitation in the YBRB were all ≥0.1, except for INM-CM-4 (0.07). Among them, for Tasmax, MPI-ESM-LR (0.42) and ACCESS1-0 (0.41) were the best single models; for Tasmin, ACCESS1-0 was the best (0.3); and for precipitation, except for TCH-MME5, the ACC of MIROC5 was 0.24.

For CMIP6, the ACC of TCH-MME6 was significantly higher than that of the other individual models for Tasmax, Tasmin, and precipitation at 0.49, 0.30, and 0.26. For Tasmax, the ACC was >0.30 for most of the models, except for CanESM5 (0.14), ACCESS-CM2 (0.28), and GFDL-ESM4 (0.28). Additionally, the CMIP6 models were better for Tasmax in MPI-ESM1-2-LR, INM-CM5-0, EC-Earth3, and NorESM2-LM (all ≥ 0.37). For Tasmin, the ACC was ≥0.10 for most of the models, except for CanESM5 (-0.08), and INM-CM5-0 and MIROC6 were in agreement with TCH-MME6 (all 0.26). For precipitation, the ACC was ≥0.10 for most of the models, except for FGOALS-G3 (−0.13), and EC-Earth3 (0.28) was better, followed by TCH-MME6 (0.26).

In CMIP5 (Figure 4), FGOALS-g2 had the smallest bias for the summer Tasmax and annual Tasmin simulations on TP (0.43 and 0.85 °C, respectively). But the biases of BCC-CSM1-1 for winter Tasmax and INM-CM-4 for winter Tasmin on FP were more than 10 °C. For precipitation, TCH-MME5 had the smallest bias for annual precipitation on YBRB (0.01 mm), and IPSL-CM5A-LR had the largest bias for summer precipitation on HB (16.97 mm). In CMIP6 (Figure 5), FGOALS-G3 showed a minimum bias for winter Tasmax on FP (1.97 °C) and autumn Tasmin on FP (0.38 °C). But the bias of CanESM5 for the winter Tasmin on FP was more than 10 °C. Additionally, CanESM5 showed the smallest bias for autumn precipitation on TP (0.003 mm), but FGOALS-G3 showed the largest bias for summer precipitation on HB (18.39 mm).

Compared with CMIP5, CMIP6 had a more extensive dry bias in summer than in other seasons. Only a few grids had a dry bias in winter, whereas the rest showed a wet bias. For the annual and seasonal Tasmax and Tasmin, FGOALS-g2 and FGOALS-G3 showed a warm bias on TP and the entire region, respectively. BCC-CSM1-1 and CanESM2 exhibited a warm bias for summer Tasmax on the TP, whereas INM-CM5-0 and NorESM2-LM demonstrated a warm bias for both summer and autumn Tasmax on the TP. CanESM5 did not demonstrate improvement compared to CanESM2, except for autumn precipitation. Similarly, FGOALS-G3 did not show improvement, except for reducing the bias in the annual mean and summer-autumn's Tasmax and Tasmin on HB and FP.

Based on the comparison of different regions, the improvement in bias was more pronounced for the multi-model mean and single models on HB and FP. CMIP6 showed the most significant reduction in the bias of annual Tasmax and Tasmin on HB and FP. And INM-CM5-0 showed the most significant improvement in the bias of Tasmin in YBRB and in different subregions (>10 °C). Moreover, compared to IPSL-CM5A-LR, IPSL-CM6A-LR exhibited poorer bias simulations for Tasmin in various subregions within YBRB. While a few models in CMIP6 did not show any improvement in bias, the majority of them demonstrated significant improvements.

For CMIP5 (Figure 6), the TCC of TCH-MME5 was 0.85, 0.89, and 0.33 for Tasmax, Tasmin, and precipitation, respectively, over the basin. The TCC of TCH-MME5 reached 0.9 for Tasmax on TP and Tasmin on HB and FP. The spatial anomalies of Tasmax on the TP were also well predicted by other single models with a TCC >0.8. Except for INM-CM-4 (TCC of 0.71–0.78), the spatial prediction ability of Tasmin in YBRB and subregions was good, with TCC >0.8. However, the spatial precipitation anomalies in the YBRB were weak (all <0.5).

For CMIP6 (Figure 7), TCH-MME6 had the strongest spatial forecast capability for Tasmax on the TP with a TCC of 0.9, and the TCC of the remaining single models was >0.8. INM-CM5-0 and MPI-ESM1-2-LR showed better spatial forecast capability for Tasmin in YBRB and in different subregions (TCC was 0.86–0.88). However, the spatial forecasting ability of BCC-CSM2-MR for Tasmin in YBRB was weak (0.46 on FP), which is not reliable. While TCH-MME6 was also affected by BCC-CSM2-MR for Tasmin, and the TCC was between 0.65 and 0.73 in different regions, although it was >0.5 and already had forecasting significance, but the reliability of the results was also not very strong. For precipitation, the TCCs of CMIP6 were similar to CMIP5's, which was not meaningful for forecasting. In addition, GFDL-ESM4, INM-CM5-0, and MPI-ESM1-2-LR improved the spatial prediction of Tasmax and Tasmin in the whole basin or subregions; BCC-CSM2-MR, CanESM5, FGOALS-G3, and IPSL-CM6A-LR for precipitation in the YBRB and subregions showed improved spatial prediction ability (Figure S5).

Figure 8 showed that, for CMIP5, TCH-MME5 and FGOALS-g2 for Tasmax; ACCESS1-0, FGOALS-g2, and GFDL-ESM2G for Tasmin; and TCH-MME5, ACCESS1-0, EC-EARTH, FGOALS-g2, GFDL-ESM2G, and INM-CM-4 for precipitation simulated results closer to the observational data; for CMIP6, TCH-MME6, ACCESS-CM2, INM-CM5-0, and NorESM2-LM for Tasmax; ACCESS-CM2, FGOALS-G3, INM-CM5-0, and NorESM2-LM for Tasmin; and TCH-MME6, EC-Earth3, and FGOALS-G3 for precipitation showed simulations closer to the observation data.

Compared with homologous models in CMIP5 at an annual scale (Figure 8), INM-CM5-0 and NorESM2-LM in CMIP6 showed a significant improvement in Tasmax, while MPI-ESM1-2-LR was closer to the observed data; INM-CM5-0, MPI-ESM1-2-LR, MRI-ESM2-0, and NorESM2-LM in CMIP6 showed significant improvements in Tasmin; and EC-Earth3, FGOALS-G3, MPI-ESM1-2-LR, MRI-ESM2-0, and NorESM2-LM in CMIP6 showed some improved simulation results in precipitation. At the seasonal scale (Figure S6), TCH-MME6, FGOALS-G3, MPI-ESM1-2-LR, and NorESM2-LM showed significant improvements in Tasmax, Tasmin, and precipitation.

This study found that there are large differences in the simulation capabilities of different models in the same phase. For example, CanESM2 and INM-CM-4 in CMIP5 and CanESM5 in CMIP6 showed significantly larger biases for Tasmax, Tasmin, and precipitation at different time scales in the YBRB than those of the other models, while other models were spatially better than BCC-CSM2-MR in CMIP6 in Tasmin in YBRB. It may be related to the spatial resolution of models (Jiang et al. 2020), the terrain (Song et al. 2013), and the response of atmospheric circulation to climate change (Monerie et al. 2020).

Additionally, we found not all selected CMIP6 models had improved performance than CMIP5 in the YBRB. For example, the bias of BCC-CSM2-MR for precipitation in the YBRB was larger than that of the CMIP5 homologous model. Similar findings have been reported in previous studies. Luo et al. (2020) compared the simulation results of CMIP5 and CMIP6 at extreme temperatures in China and showed that some CMIP6 models were not significantly different from their homologous CMIP5 models. Additionally, Wang et al. (2022) compared the ability of CMIP5 and CMIP6 to simulate precipitation over the Han River and showed that CMIP6 models are not necessarily better than CMIP5. Hu et al. (2022) also found that the CMIP6 models have a more severe cold bias than the CMIP5 model for annual, spring, summer, and winter simulations, especially on the western part of TP, which is consistent with the results of the present study. Furthermore, Li et al. (2021) showed that CMIP6 does not have the advantage of simulating total precipitation and maximum number of consecutive dry days. Moreover, CMIP6 has been found to be slightly less sensitive to global hydrology than CMIP5 (Thackeray et al. 2022).

Meanwhile, TCH-MME5 and TCH-MME6 were not as good as some single models in CMIP5 and CMIP6 in reproducing the temperature and precipitation in the YBRB. As shown in the Taylor diagram, TCH-MME6 simulates Tasmin at different time scales in the YBRB less than other models except BCC-CSM2-MR and CanESM5, which is mainly influenced by the models with weak simulation ability. Zhu et al. (2020) also showed that the CMIP6 MME on the TP and northwest China had poorer temperature and precipitation simulations. Additionally, Zhu & Yang (2020) found that, compared to CMIP5, the MME of CMIP6 no better reproduces precipitation amplitudes in wet regions. However, when the strengths and weaknesses of the model simulation capabilities are unknown, TCH-MME can reduce the influence of poorer models and avoid errors caused by the inadvertent selection of poorer models for the study. Moreover, TCH-MME exhibits evident improvements in bias and other aspects; thus, it is advantageous to use TCH for the calculation of the MME.

However, there are some limitations in the study: (1) this study only compared and evaluated a selection of commonly used models for Tasmax, Tasmin, and precipitation for CMIP5 and CMIP6, without encompassing all models for different variables. If additional models are needed beyond those studied in this research, they should be assessed. (2) This study directly compared the original model of CMIP5 and CMIP6 without bias correction. The results reveal biases and limitations in the original CMIP5 and CMIP6 in YBRB. While the MME approach can mitigate biases, the MME results may still not meet accuracy requirements if the biases of the original data are too large. Therefore, bias correction is necessary when using CMIP5 or CMIP6 for climate change simulation and projection.

This study compared simulations from 13 CMIP5 and CMIP6 homologous models and their MMEs (TCH-MME5 and TCH-MME6) for temperature and precipitation over the YBRB, and the following conclusions were obtained:

• (1) Overall, the MMEs based on the TCH method yield better simulation results in terms of temperature and precipitation, with TCH-MME6 outperforming TCH-MME5, although biases still exist.

• (2) Among the CMIP6 models, INM-CM5-0, MPI-ESM1-2-LR, and NorESM2-LM exhibit the highest simulation abilities for Tasmax and Tasmin, while NorESM2-LM performed best for precipitation in the YBRB.

• (3) For a single model, while CMIP6 shows some improvements over CMIP5, not all models exhibit improvement. CMIP6 still exhibits cold bias over the YBRB, with high dry bias centers on HB and FP, and wet bias on TP.

• (4) Temporally, CMIP5 and CMIP6 perform well in simulating spring and autumn Tasmax, Tasmin, and precipitation in YBRB. Spatially, CMIP5 and CMIP6 demonstrate good forecasting capabilities for Tasmax on TP, Tasmin over YBRB, and precipitation on FP.

Generally, GCM's simulation capabilities for temperature and precipitation in the YBRB need to be enhanced, especially for precipitation, as the uncertainty and fluctuation of the models are large and the time and space correlations are relatively weak. Therefore, GCM's data are not recommended for direct use in simulating and predicting precipitation in the YBRB. Additionally, the TCH method is an effective MME mean method and is particularly suitable for areas lacking observations. It is expected to see broader application in the fusion and evaluation of climate model data going forward.

This research was funded by the National Natural Science Foundation of China (grant number 42061005), the Applied Basic Research Programs of Yunnan province (grant number 202101AT070110) and National College Student Innovation Training Program (grant number 202110673060).

All relevant data are available from an online repository or repositories. The CMIP5 and CMIP6 model data could be downloaded from https://www.wcrp-climate.org/wgcm-cmip. The raw gridded data for temperature reference could be downloaded from https://hydrology.soton.ac.uk/data/pgf/v3/. The raw gridded data for precipitation reference could be downloaded from http://aphrodite.st.hirosaki-u.ac.jp/download/.

The authors declare there is no conflict.