Increase in Subdaily Precipitation Extremes in India Under 1.5 and 2.0 °C Warming Worlds

Abstract

An increase in short duration precipitation extremes poses challenges for storm water design in rapidly urbanizing India. The recent Paris Agreement aims to limit the global mean temperature (GMT) below 1.5 °C (and possibly below 2.0 °C) from the preindustrial level. However, the changes in subdaily precipitation extremes in India remain unrecognized under the 1.5- and 2.0-°C temperature targets. Here using observations and projections of the subdaily precipitation, we show that a majority (11 out of 15) of General Circulation Models underestimate 3-hourly precipitation extremes and overestimate the relationship between 3-hourly precipitation extremes and GMT (scaling) in India. A rise of 1.5 (2.0 °C) in GMT from the preindustrial level is projected to cause 20% (25%) increase in 3-hourly precipitation maxima at 100-year return period under the stationary condition, which can further rise by 10% under the nonstationary condition. Projected warming results in a much faster (almost twice) increase in 3-hourly precipitation maxima than 24-hourly 100-year precipitation maxima. Moreover, 3-hourly 100-year precipitation maxima are projected to increase significantly at 78 locations (out of 89) if GMT increases from 1.5 to 2.0 °C from the preindustrial level. Our findings have implications for urban storm water designs in India.

Plain Language Summary

Urban areas in India face frequent flooding due to increase in short-duration precipitation extremes. However, it remains unknown how short-duration precipitation extremes that are relevant to urban storm water designs change under the 1.5° and 2.0° warming worlds. Here using the data from CMIP5 GCMs and observations, we show that 3-hourly precipitation maxima at 100-year return interval are projected to increase by 20% (25%) if the global mean temperature rises above 1.5° (2.0°) from the preindustrial level. Under the nonstationary assumption, short duration precipitation extremes are projected to rise more than that of under stationary assumption. Our findings provide new insights for urban storm water designs in India.

1 Introduction

The continued anthropogenic emissions resulted in increased greenhouse gas concentration in the atmosphere leading to a significant climate warming (Wentz et al., 2007). The projected increase in global mean temperature (GMT) is likely to alter the hydrologic cycle, posing serious risks associated with hydroclimate extremes (Allen & Ingram, 2002; Held & Soden, 2006). The water holding capacity of the atmosphere increases by about 6–7% per 1-°C warming leading to more intense and frequent extreme precipitation events (Kharin et al., 2007; Trenberth et al., 2003). To prevent dangerous anthropogenic interference with the climate system, the Paris Agreement (2015) sets a long-term temperature goal of “holding the increase in GMT to well below 2°C above the pre-industrial levels (1850-1880) and pursuing efforts to limit the temperature increase to 1.5°C” (United Nations Framework Convention on Climate Change, 2015). Based on the post-Paris science agenda, recent studies estimated the climate impacts of these levels of warming on extreme precipitation (Fischer & Knutti, 2015; Schleussner et al., 2017) and reported a robust increase in daily precipitation extremes globally with an increase in GMT from 1.5 to 2.0 °C.

Increasing trends in daily extreme precipitation in India have been observed during the recent decades (Roxy et al., 2017; Vittal et al., 2013). Moreover, urban areas in India have witnessed eccentric precipitation extremes in the past, which have affected human lives and infrastructure (Ali & Mishra, 2017, 2018; Ali et al., 2014). The Clausius-Clapeyron (C-C) relationship can be used to estimate the changes in extreme precipitation under the warming climate if relative humidity remains constant and atmospheric circulation does not change considerably (Lenderink & Van Meijgaard, 2008, 2010). Precipitation extremes may increase with higher rates (more than 7%/K; superscaling) than C-C relationship depending on the nature and duration of precipitation extremes (Molnar et al., 2015; Wasko et al., 2015), temperature (Westra et al., 2014), season (Berg et al., 2009), and geographical location (Wasko & Sharma, 2017). For instance, subdaily precipitation extremes are more sensitive to warming than daily precipitation (Ali & Mishra, 2017, 2018; Barbero et al., 2017).

The higher sensitivity of subdaily precipitation extremes to warming climate has more significant implications for urban storm water designs (Fadhel et al., 2017). The stationary atmospheric conditions which were previously considered for storm water design values may lead to an underestimation of risks (Cheng & AghaKouchak, 2014; Hosseinzadehtalaei et al., 2017, 2018). Despite the implications of climate warming on subdaily precipitation extremes, it remains unclear how the urban storm water designs will be affected if GMT increases to 1.5 and 2.0 °C from the preindustrial level. Here using 3-hourly gridded precipitation from General Circulation Models (GCMs), we provide a first-ever assessment of the changes in subdaily precipitation extremes across 89 urban areas in India under the 1.5- and 2.0-°C warming world.

2 Data and Methods

We obtained 3-hourly precipitation data from Multi-Source Weighted-Ensemble Precipitation (MSWEP; 0.25°; 1979–2005; Barbero et al., 2017), Tropical Rainfall Measuring Mission 3B42v7 (TRMM; 0.25°; 1998–2005; Huffman et al., 2003), and Climate Prediction Center morphing technique (CMORPH; 8-km resolution; 1998–2005; Xie et al., 2017) for 89 urban areas that are planned to be developed as smart cities in India (Kumar et al., 2017; Figure S1 and Table S1 in the supporting information). We used these data sets as subdaily station data are unavailable, and these are more reliable than other available multisatellite precipitation products over India (Ali & Mishra, 2018; Prakash et al., 2016). The MSWEP data set is developed using the station, satellite (CMORPH, GSMaP-MVK, and TRMM)-based observations, and atmospheric reanalysis products (European Reanalysis (ERA)-Interim and Japanese Reanalysis (JRA)-55), which shows a good correlation against 125 FLUXNET towers estimates across the globe (Beck et al., 2017). TRMM data set (from 1998 onward) is prepared by adjusting bias from the station-based observations and passive microwave data where it is available and infrared elsewhere (Huffman et al., 2003). Similarly, CMORPH precipitation product is available at 8-km resolution from 1998 onward, which is developed using precipitation estimates from both passive microwave and infrared sensors. A detailed description of these data sets can be found in Ali and Mishra (2018).

We obtained monthly global surface temperature (GMT) anomaly from National Aeronautics and Space Administration's Goddard Institute for Space Studies Temperature Analysis (https://data.giss.nasa.gov/gistemp/.) for the period of 1969–2015 (Hansen et al., 2010). The GMT anomalies are estimated considering 1951–1980 as the reference period.

To estimate changes in subdaily precipitation extremes under the future climate, we used 3-hourly precipitation data from 15 GCMs (Table S2) that participated in phase 5 of the Coupled Model Intercomparison Project (CMIP5; Taylor et al., 2011, 2012). We used the first ensemble member (r1i1p1) for all the selected GCMs. The data were obtained for the historical (1969–2005) and future (2006–2100) periods. The projected future data are the combination of (a) present initial conditions and (b) future climate scenarios under representative concentration pathways (RCPs). RCPs are the scenarios which are targeted to achieve after implementing policy actions to reduce greenhouse gases emissions (Friedlingstein et al., 2008). The labels for the RCP denote a rough estimate of the radiative forcing (W/m²) at the end of 2100. There are significant improvements in CMIP5 over CMIP3 in simulating daily precipitation as CMIP5 GCMs have a greater response to natural forcing and aerosols (Sperber et al., 2013).

Comparing precipitation from the data sets at different spatial resolutions (e.g., GCMs and observations) is not appropriate as precipitation intensity varies with the area (or grid size). To convert precipitation from the spatial scale of a GCM (more than 1°, Table S2) to 0.25°, areal reduction factor (ARF) is applied. ARF is the ratio of precipitation intensity from data sets at different spatial scale for the same duration (Asquith, 1999). We used ARF based on U.S. Weather Bureau 1975 method (TP-29) which showed the best performance for urban areas in India as discussed in Ali and Mishra (2017). The ARF is expressed as

(1)

where is the annual maximum precipitation (AMP) for the year j obtained using gridded data of different spatial resolution (here GCMs; please see Table S2 for more details on spatial resolution of GCMs), P_j is the AMP for year j obtained using gridded precipitation products (MSWEP, TRMM, and CMORPH) at 0.25°, and n is the number of years.

We estimated regression slopes (scaling hereafter) using quantile regression which is a more robust and flexible method than binning technique (Ali & Mishra, 2017; Wasko & Sharma, 2014). Since GMT anomaly is available at monthly time scale, we matched 3-hourly precipitation to corresponding GMT anomaly for the same month (see Table S3 for more details). Then, for a set of data pairs (P, T), the quantile regression for 95th percentile is expressed as (Wasko & Sharma, 2014)

(2)

where log(P) is logarithmically transformed precipitation, T is corresponding GMT anomaly, and regression slope (scaling, dP95(%)/K) is estimated using an exponential transformation of regression coefficient β₁ (Ali & Mishra, 2017; Wasko & Sharma, 2014)

(3)

The performance of 15 GCMs (ALL-GCMs) was evaluated using bias (difference) in the 95th percentile of precipitation (P95) and scaling (from quantile regression) from 3-hourly precipitation data sets (MSWEP, CMORPH, and TRMM). The 95th percentile refers to the top 5% of the distribution of 3-hourly precipitation greater than or equal to 1 mm (P > =1 mm). Bias (%) is defined as

(4)

where P95_GCM is P95 from GCMs and P95_TRMM is P95 from TRMM (or from the other precipitation products).

The ensemble of the five best GCMs (BEST-GCMS), which show the lowest bias in P95 and scaling, is used to estimate stationary and nonstationary return levels (3-hr 25- to 100-year precipitation maxima; design storm hereafter). Design storm is the magnitude of 3-hourly precipitation at a given return period, which is often used for hydrologic designs. We estimated design storm for all the selected locations using Generalized Extreme Value (GEV) distribution based on annual block maxima (ABM) approach. Previously, Ali and Mishra (2017) reported that ABM and peak over a threshold approaches result in similar estimates of precipitation maxima; therefore, we limit our analysis based on the ABM approach.

Using the ABM approach, we fitted annual time series of the highest precipitation in a year (AMP) and corresponding monthly GMT anomaly to the GEV distribution using maximum likelihood estimates. For the stationary return levels, only AMP was used to estimate GEV distribution parameters: location (μ), scale (σ), and shape (k) parameter (Katz et al., 2002). Then, to account for the role of global warming, we estimated nonstationary return levels considering GMT anomaly as a covariate (Agilan & Umamahesh, 2017). We allowed μ to change linearly with GMT anomaly and kept the other parameters (σ and k) constant (Ali & Mishra, 2017; Mondal & Mujumdar, 2016). The detailed mathematical description of these steps is provided in the supporting information. We also evaluated improvements in the nonstationary GEV model over the simple stationary GEV model based on Deviance Statistic (D), which can be estimated using the following equation:

(5)

where l₁(M₁) is a maximum log likelihood estimate (maximum likelihood estimate) for nonstationary GEV model (based on the maximization of GEV log likelihood function; see supporting information) and l_o(M_o) is a maximum log likelihood estimate for stationary GEV model. If D > 3.84 (i.e., chi-square test at 5% significance level), then use of nonstationary GEV model and covariate is justified (Table S7; Ali & Mishra, 2017). We evaluated the significance of shape parameter using the Deviance statistic (Kharin & Zwiers, 2005). The D value for shape parameter was found to be lower than 3.84 indicating that the shape parameter does not contribute significantly to the nonstationary GEV model (please see Table S7 for more details). Therefore, our nonstationary model is based on the location parameter only.

We estimated the median change in design storm with the change in GMT from 0.5 to 2.5 °C from the preindustrial period (1861–1890) using the method described in King et al. (2017; see Table S4). To do so, first, we divided GMT (2006–2100) into 30-year moving windows. Then, we estimated the change (ΔT) in GMT for each 30-year time window against the preindustrial period (1861–1890). We distributed these 30-year time windows into different levels of warming (ΔT = 0.5–2.5 °C). For each 30-year time window for a given ΔT, we estimated 3-hr 100-year precipitation maxima (design storm) considering stationary GEV model using data from GCMs (best: BEST-GCMs; poor: POOR-GCMs, and all GCMs: ALL-GCMs). We further estimated the change (%) in the estimated design storm from the reference period (1976–2005). Finally, median change (%) in design storm was estimated using all combinations (GCMs and RCPs) for a given change in GMT (ΔT) considering the stationary GEV model. The same procedure was applied to the nonstationary GEV model using GMT as a covariate for the location parameter.

3 Results and Discussion

3.1 Evaluation of CMIP5 Models Against Observed Subdaily Precipitation Extremes

We start our analysis by evaluating 15 GCMs for 3-hourly extreme precipitation exceeding the 95th percentile threshold (P95) against MSWEP observations (Figures 1 and S2). We find that a majority (11) of the GCMs underestimate P95 (median bias up to 50%) as compared to MSWEP observations (Figure 1a) for the 1979–2005 period. The five GCMs that show a relatively higher (median bias 32–50%) dry bias in P95 are HadGEM2-ES, GDFL-ESM 2G, GDFL-ESM 2M, CMCC-CMS, and INM-CM4 (hereafter POOR-GCMs). The best five GCMs (BEST-GCMs) that show relatively less bias in P95 (median bias 1–15%) are MIROC5, MIROC-ESM, IPSL-CM5A-MR, ACCESS1-0, and CNRM-CM5. We also evaluated bias for other extreme precipitation thresholds (90–99th percentiles) and found that a dry bias in the GCMs is consistently present regardless of the threshold of extreme precipitation (Figure S3). Moreover, GCMs show a dry bias in simulating subdaily precipitation extremes against the other precipitation products (CMORPH and TRMM; Figures S4–S7). Our selection of the BEST-GCMs is consistent with Mishra et al. (2014) who evaluated the performance of GCMs for daily extreme precipitation in India and reported that CNRM-CM5 and MIROC5 are better performing models.

Figure 1

(a) Pooled IQR (25th, 50th, and 75th percentiles) in absolute bias (%) in 95th percentile of 3-hourly precipitation (P95) in GCMs against MSWEP for the period of 1979–2005. (b) Pooled IQR in scaling (%/K) obtained using 3-hourly P95 from MSWEP/GCMs with GMT using quantile regression for 1979–2005. (c) Scaling (%/K) obtained using 3-hourly P95 from MSWEP with GMT. (d) Same as (c) but for the BEST-GCMs. (e) Pooled distributions of scaling (%/K) using 3-hourly P95 from MSWEP (red) and BEST-GCMs (blue). (f) Relationship between absolute bias (median) in P95 and bias in GMT from BEST-GCMs (blue), POOR-GCMs (red), and other GCMs (green). (g) Same as (f) but for median scaling (%/K) obtained using 3-hourly P95 from GCMs with GMT. (h) Same as (f) but for median scaling with bias in P95. Boxplots in (a) and (b) show 25th, 50th, and 75th percentiles of values. Statistical significance (at 0.05 significance level) was estimated using the Kolmogorov-Smirnov (KS) tests for distributions of the scaling in (e) and probability distribution functions we obtained using kernel smoothing function (ksdensity). IQR = interquantile ranges; GCM = General Circulation Model; MSWEP = Multi-Source Weighted-Ensemble Precipitation; GMT = global mean temperature.

After the evaluation of GCMs for the subdaily extreme precipitation, we estimated the relationship between P95 and GMT anomaly (referred as scaling hereafter) using quantile regression (see section 2 for details). We find that 3-hourly P95 from MSWEP shows a median scaling of 9%/K (higher than C-C scaling of 7%/K) for all 89 locations (Figure 1b). Our results show that all the GCMs exhibit a higher (than observed) scaling for 3-hourly P95 against observed GMT. Similar to observations from MSWEP, the BEST-GCMs show scaling of 10–11%/K between 3-hourly P95 and GMT, indicating a better performance for bias and scaling. On the other hand, the POOR-GCMs display even higher scaling (scaling: 12–14%/K than the BEST-GCMs) between 3-hourly P95 and GMT, which was also confirmed by the comparison of scaling of MIROC5 (BEST-GCM) and INM-CM4 (POOR-GCM; Figure S8).

We find that the ensemble median of scaling from the BEST-GCMs is similar to the scaling obtained using MSWEP at 89 locations (Figure 1d). Most of these 89 locations show a super C-C scaling relationship between P95 and GMT for the observed MSWEP data set (Figure 1c). Moreover, both median and distribution of scaling from MSWEP and BEST-GCMs were not found significantly different (p > 0.05; Figure 1e). Bias in the scaling relationship between 3-hourly P95 and GMT is consistent with the other precipitation products (e.g., CMORPH and TRMM) indicating the robustness of the results (Figure S9). Consistent with previous studies (Kharin et al., 2013; Wang et al., 2017), the super scaling relationship between 3-hourly P95 and GMT increases under the projected future climate (Figures S10 and S11), which indicates a higher sensitivity of subdaily precipitation extremes. A higher scaling in the future climate can be attributed to warming, which can result due to the higher atmospheric water holding capacity (Asadieh & Krakauer, 2015; Shaw et al., 2011).

We find that a high bias in GCMs to simulate 3-hourly P95 is associated with bias in GMT anomaly (Figures 1f and S12). For instance, the POOR-GCMs show a higher negative (median −0.016 K) bias in GMT anomaly and also display a high dry (median 38%) bias in 3-hourly P95 (Table S5). We find a strong relationship (correlation = −0.89) between bias in P95 and bias in GMT. However, the correlation decreases when all the values for 89 locations were selected (Figure S12). The reduction in correlation considering all the locations is essentially driven by the differences in bias in P95 instead of bias in GMT (as it remains the same for all the locations). We, therefore, consider median (of 89 locations) as a robust measure to show the negative relationship between bias in P95 and bias in GMT.

The cold bias in GMT anomaly in the majority of GCMs may be due to bias in sea surface temperature (SST). The GCMs with a larger cold bias in SST underestimate extreme precipitation due to the weakening of moisture fluxes from the Arabian Sea over India (Levine et al., 2013; Sandeep & Ajayamohan, 2014). Dai (2006) also reported that the deficiency of GCMs in measuring tropical rainfall is correlated with SST bias. These cold and dry biases in GMT in GCMs also affect the scaling relationship between subdaily precipitation extremes. For instance, the POOR-GCMs show a higher scaling (median 13%/K) between 3-hourly P95 and GMT. Therefore, improving the cold bias in GMT in the GCMs can enhance their ability to better simulate subdaily extreme precipitation and the scaling relationship. However, there may be other important physical processes and spatial resolution responsible for bias in P95 in GCMs. Kharin et al. (2007) argued that the large intermodel variation in precipitation over the tropics may be linked to the underrepresentation of physical processes that are responsible for extreme precipitation. Moreover, the sensitivity of precipitation to small changes in SST can lead to a large uncertainty in precipitation projections (Brown et al., 2015).

3.2 Storm Water Design Under the Warming Climate

After evaluating the GCMs for precipitation extremes and scaling, we estimated projected changes in 3-hourly storm water design estimates at 100-year return interval under the 1.5-, 2.0-, and 2.5-°C warming worlds using the data from BEST-GCMs (Figure 2). As the GCMs show a super C-C scaling between 3-hourly P95 and GMT (Figure 1d), we find that increase in GMT from the preindustrial level will lead to a substantial increase in 3-hourly 100-year precipitation maxima (Figures 2a–2c). We find that 3-hr 100-year precipitation maxima increase by 10% (median) in response to moderate warming of 0.5 °C from the preindustrial level (Figure 2g). A rise of 1.5 (2.0 °C) or above in GMT will lead to more than 20% (25%) increase in 3-hourly 100-year precipitation maxima. Under the stationary assumption, we find a substantial (10–40%) increase in 3-hourly precipitation maxima at 100-year return interval at 1.5- and 2.0-°C warming world (Figures 2a–2c). Similar results were obtained for different time windows and model combinations (Figure S13). Under the nonstationary condition, the increase in 3-hourly precipitation maxima at 100-year return period is projected to rise further (from stationary assumption) significantly (p < 0.05) by about 10% (Figures 2d–2g and S14a–S14e). Similar increases in 3-hourly precipitation maxima under the stationary and nonstationary conditions were found for the return periods of 25 and 50 years (Figures S15 and S16). Our results show that the projected increase in 3-hourly precipitation maxima is higher in the POOR-GCMs (median 37%) than that of the BEST-GCMs (median 28%) for 1.5-°C warming (Figure S17 and Table S5). Since subdaily extreme precipitation is more sensitive to warming than daily precipitation extremes, we estimated the projected change in daily precipitation maxima at 100-year return period for a 0.5° to a 2.5° warming world (Figure 3). We find that projected increase in 3-hourly precipitation maxima at 100-year return period is almost twice of the increase in 24-hr precipitation maxima (Figures 3a–3d). For instance, a 1.5-°C rise in GMT from preindustrial is projected to result in 30% and 18% (median) increase in 3- and 24-hr precipitation maxima at 100-year return period (Figure 3). This higher increase of 3-hr precipitation maxima can be partially attributed to the higher scaling shown by 3-hourly precipitation extremes with the increase in GMT (Figure 3e). We find that scaling between 3-hourly P95 is about 20–40% higher than the scaling for daily precipitation extremes (Figure 3e).

Figure 2

(a–c) Median change (%) in 3-hr 100-year precipitation maxima (design storm) with increase in GMT (ΔT) by 1.5 to 2.5 °C from the BEST-GCMs assuming stationary (S) conditions. (d–f) Same as (a–c) but difference (nonstationary-stationary, %) in change in design storm considering stationary (S) and nonstationary (NS) conditions with GMT as covariate. (g) Pooled median change (%) in design storm assuming stationary (S; red) and nonstationary (NS; blue) with change in global mean temperature from preindustrial level. (h) Pooled median (%) change in stationary design storm using BEST-GCMs (blue), POOR-GCMs (red), and remaining GCMs (green) with pooled median bias (%) in 3-hourly P95 in GCMs from MSWEP (1979–2005). (i) Same as (h) but with scaling using 3-hourly P95 from GCMs with GMT (1979–2005). Error bars show a range of one standard deviation from the median. The median percentage change in 3-hr 100-year precipitation maxima for all windows was estimated considering 1976–2005 as the reference period. Percentage change in 3-hr 100-year precipitation maxima was estimated using the stationary (S) and nonstationary (NS) conditions with GMT as a covariate. GMT = global mean temperature; GCM = General Circulation Model; MSWEP = Multi-Source Weighted-Ensemble Precipitation.

Figure 3

(a–c) Median change (%) in 24-hr 100-year precipitation maxima with increase in GMT (ΔT) by 1.5 to 2.5 °C using BEST-GCMs assuming nonstationary (NS) conditions. (d) Pooled median change (%) in 3-hr 100-year (blue) and 24-hr 100-year (red) precipitation maxima assuming nonstationary (NS) conditions, with global warming. (e) Difference (%) in scaling using precipitation from 3-hourly and daily BEST-GCMs (1979–2005). (f) Pooled median (%) change in nonstationary design storm using BEST-GCMs (blue), POOR-GCMs (red), and remaining GCMs (green) with scaling (%/K) using daily P95 from GCMs with GMT (1979–2005). Error bars show a range of one standard deviation from the median. GMT = global mean temperature; GCM = General Circulation Model.

A higher scaling relationship for subdaily precipitation extremes has been reported in previous studies (Ali & Mishra, 2017, 2018; Lenderink et al., 2017) and attributed to the convective nature of short-duration precipitation extremes. However, other factors related to dynamic scaling that is controlled by the large-scale atmospheric processes (wind velocity and atmospheric circulation) can also partially contribute to a larger increase in subdaily precipitation extremes (Ali & Mishra, 2018; Pfahl et al., 2017). We find that the scaling has a significant relationship with the projected increase in subdaily (correlation = 0.89) and daily (correlation = 0.87) extreme precipitation under climate warming (Figures 2h, 3f, and S12 and Tables S5 and S6). For instance, POOR-GCMs that show a higher scaling in daily (median 10.14%) and subdaily precipitation (median 13%/K) also project a larger increase in daily (median 26%) and subdaily (median 38%) precipitation extremes, respectively, under the 1.5-°C warming climate (Figures 2h and 3f). Our results highlight the need of reducing the bias in extreme precipitation simulated by GCMs, which is associated with the bias in GMT in the models along with the other factors such as convective precipitation (O'Gorman, 2015), spatial resolution (Gutowski Jr et al., 2010), and other physical processes (Kharin et al., 2007).

Since the Paris Agreement aims to limit the rise in GMT below 2.0 °C and with a more ambitious target of 1.5 °C, we evaluate the sensitivity of rising in GMT on subdaily precipitation maxima relevant to storm water design (Figure 4). We find that, under the nonstationary assumption, an increase in GMT from 1.5 to 2.0 °C, there will be an additional increase of 12.8% in 3-hourly precipitation maxima at 100-year return period. The 3-hourly 100-year precipitation maxima are projected to increase significantly at 78 locations (out of 89) if GMT increases from 1.5 to 2.0 °C from the preindustrial level. Moreover, if we let the GMT rise from 1.5 to 2.5 °C by the end of the 21st century, 87 out of 89 locations are projected to experience a significant rise in 3-hourly precipitation maxima at 100-year return period (Figure 4b). This 1-°C rise from the more ambitious target of 1.5 °C will result in a median increase of 23.4% in 3-hourly 100-year precipitation maxima (Figure 4b). This increase in 3-hourly precipitation maxima is consistent with the stationary design estimates (Figure S18). Similar to our findings, Wang et al. (2017) and Schleussner et al. (2017) argued based on observations that a half degree rise in GMT can lead to substantial increase in precipitation extremes. O'Gorman (2015), however, argued that the influence of mesoscale convection and precipitation duration needs to be better understood to reduce the uncertainty in precipitation extremes over the tropics.

Figure 4

(a) Change (%) in design storm with increment of warming from 1.5 to 2.0 °C based on nonstationary conditions. (b) Same as (a) but increment of warming from 1.50 to 2.5 °C. (c) Pooled distributions of increase in design from reference period (1976–2005) for 1.5-°C warming (blue) and 2.0-°C warming from preindustrial period (1860–1890). (d) Same as (c) but for 1.5-°C warming (blue) and 2.5-°C warming (red). The red circle shows increasing change; blue circle shows decreasing change. The unfilled circle shows insignificant change; filled circle shows significant change. Statistical significance was estimated using the Ranksum and KS tests for mean and distributions of increase in 3-hr 100-year precipitation maxima.

4 Conclusions

Based on our findings, the following conclusions can be made.

1. The majority of GCMs (11 out of 15) underestimate 3-hourly precipitation extremes for the 89 locations in India. GCMs that have less bias in GMT anomaly better simulate subdaily precipitation extremes and the scaling relationship over India in comparison to the GCMs that have a relatively high bias in GMT anomaly.
2. We find that 3-hourly 100-year precipitation maxima under the stationary assumption are projected to increase by about 20% (25%) with a rise of 1.5 (2.0 °C) or above in GMT from preindustrial level. However, if we consider the nonstationary condition, there will be an additional increase of 10% in 3-hourly precipitation maxima in comparison to the stationary condition. Under the similar warming, we find that projected increase in 3-hourly precipitation maxima at 100-year return period is almost twice of the increase in 24-hr precipitation maxima for the same return period.
3. Under the nonstationary assumption, a rise in GMT from 1.5 to 2.0 °C will result in an additional increase of 12.8% in 3-hourly precipitation maxima at 100-year return period. Moreover, 3-hourly 100-year precipitation maxima will significantly increase at 78 locations (out of 89) if GMT increases from 1.5 to 2.0 °C from the preindustrial level.

Acknowledgments

The authors acknowledge data availability from CMIP5, CMORPH, MSWEP, TRMM, and GISTEMP. All the data sets used in this study are publicly available. The first author acknowledges the funding from the Ministry of Human Resources Development (MHRD). The work was partially supported by the BELMONT forum grant to the second author.