Evaluation of precipitation extremes over the Asian domain: observation and modelling studies

2.1 The seven reference datasets

In this study, we used gridded daily precipitation datasets with a high-resolution. These datasets are roughly grouped into three categories: rain-gauge-based (APHRODITE, CPC-UNI), satellite-based (TRMM, GPCP1DD), and reanalysis (ERA-Interim, MERRA, and JRA55). To ensure a common period among datasets, we selected data for period from 1998 to 2007. A brief summary of the datasets is given in Table 1. Detailed description of the datasets follows.

2.2 CMIP5 model data

We used the Atmospheric Model Intercomparison Project (AMIP) simulations from 29 CMIP5 AGCMs for the 28-year period from January 1980 to December 2007 (Table 2). The models are grouped according to their spatial resolution: high-resolution (grid spacing of < 1°), medium-resolution (grid spacing between 1° and 2°), and low-resolution (grid spacing of > 2°). Each model uses different physics schemes. All the above described datasets are produced by different numerical models or assimilation systems, thus can be considered independent of each other.

2.3 Analysis region

Our region of interest is the Asian summer monsoon region (15°S–45°N, 70–150°E: Fig. 1a). This region consists of South Asia (the Indian sub-continent), East Asia (China, Korea and Japan) and Southeast Asia (Myanmar, Thailand through to the Indonesian islands and the maritime continent). Considering the spatial features associated with Asian summer monsoon and the number of stations, for some analysis we separated the entire domain into three sub-regions: East Asian region (20–40°N, 110–150°E), South Asian region (5–25°N, 70–90°E), and Southeast Asian region (10–20°N, 100–110°E). Maritime continent also belongs to one of the Asian summer monsoon sub-domains, but are excluded when averaging over the Southeast Asian region.

Open image in new window

2.4 Methodology

According to Yang et al. (2014), upscaling of fine resolution data into a coarse grid through averaging reduces signals/variances. Consequently, the mismatch between process and analysis scales often leads to difficulties in producing reliable statistics for the aggregated data. In this study, we apply to interpolate the datasets from coarse to fine resolution to compare the characteristics among the multiple datatsets.

Rainfall is a point process with large spatial and temporal discontinuities ranging from very weak to strong events within small temporal and spatial scales (Malik et al. 2012; Wulf et al. 2012). To define daily precipitation extremes, we used the percentile approach (Diffenbaugh et al. 2005; Malik et al. 2012; Kopparla et al. 2013; Singh et al. 2013). In particular, for a percentile threshold of p, we find the pth percentile of the distribution of daily rainfall over all the grids over the Asian domain. We then examine the spatio-temporal characteristics of rainfall exceeding this percentile, the absolute value of which will differ between different datasets. The advantage of this approach is that it removes the effects of bias in precipitation amounts between the different datasets, while retaining reliable information about precipitation pattern and behavior (Kendon et al. 2012). Here we calculate the 75th, 95th and 99th percentile of daily precipitation probability distribution function (PDF) at each grid point for June–August (JJA) of each year. The frequency of heavy rainfall is defined to be more than 30 mm/day (Bhowmik et al. 2007; Kusumoki et al. 2012; Agnihotri et al. 2015). In addition, we analyze the distribution of the right tail of the precipitation distribution, defined as the range between the 99th and 90th percentile (Scoccimarro et al. 2013, 2014), similar to Interquartile range (IQR), which is calculated as the 75–25th percentile. Furthermore, we evaluate the extreme precipitation fraction as a ratio of the 95th percentile precipitation to the precipitation climatology.

3.1 Climatological features over the asian domain

Changes in precipitation extremes are primarily influenced by changes in large-scale circulation over the tropics (Maredith et al. 2015). Most of the reanalysis datasets reasonably reproduce the climatological circulation features. Hence, Fig. 1b displays the vector winds at the lower tropospheric level (850 hPa) derived from the ERA-Interim dataset. This figure clearly illustrates the summer monsoon circulation flow pattern. Over the South Asian region, the strong southwesterly flow over the Arabian Sea (0–20°N, 40–70°E) transports large amounts of moisture from the Indian Ocean towards the Indian sub-continent. This flow further gathers moisture from the Bay of Bengal and transports towards northeastern parts of India and towards the Myanmar–Thailand region. The most dominant feature over East Asia is the North Pacific Subtropical High. The southeasterly/southerly/southwesterly flow along the western edge of this high (Figs. 1b, 10°S–30°N, 110–130°E) transports moisture from the West Pacific Ocean towards China, Korea and Japan. Moisture from the South China Sea is also transported towards South China and the neighboring regions of Southeast Asia in particular Philippines and Vietnam. Maritime continent over the Indonesian region gets moisture from the South Pacific Ocean brought by the southeasterly winds. Thus the main moisture sources for the precipitation can be inferred from Fig. 1b. Next, we present the seasonal summer monsoon rainfall patterns as depicted by the seven datasets (Fig. 2).

Open image in new window

Over South Asia, large amounts of rainfall are located over the west coast of India (orographic effects), central and northeast India. Over South-east Asia, the Arakan and Myanmar coast (orographic effects) are regions of large amounts of rainfall. Over East Asia, regions of southeast China extended until the Korea–Japan peninsula display large amounts of rainfall. However, CPC-UNI does not capture rain over the Myanmar and the Arakan coast. Interestingly, ERA-Interim correctly reproduces very less precipitation over the southeast Indian peninsula (Fig. 2f). This region gets major proportion of annual rainfall during the October–December period (Kripalani and Kumar 2004). This same dataset also reasonably reproduces the climatology and Global Monsoon precipitation with highest skill (Lin et al. 2014). MERRA and JRA55 tend to overestimate rainfall over southeast Asia for boreal summer. More details on South and East Asian Monsoons are available in a recent paper (Preethi et al. 2017) and over Southeast Asia in earlier papers (Kripalani and Kulkarni 1997, 1998). Perfect reference datasets do not exist and there are large uncertainties in extreme rainfall among various datasets (e.g., Sunyer et al. 2013; Turco et al. 2013). Many previous studies have suggested use of rain-gauge datasets for reference, but they too have shortcomings, as the sites do not cover the required domain. On this basis, it is necessary to investigate the differences in extreme precipitation, if any, across the seven datasets before assessing the climate models.

3.2 Extremes based on percentiles

Figure 3 shows extreme rainfall for boreal summer in individual datasets and their ensemble mean in terms of the 95th percentile values in the daily precipitation PDF. Although the ensemble of our seven datasets is affected by smoothing, it still captures well precipitation in mountainous regions compared to each dataset. Over individual datasets, the distributions of extreme values are similar to the daily mean precipitation in terms of spatial characteristics, but show larger differences in magnitude. In particular, GPCP and TRMM (i.e., datasets based on satellite) display higher extreme precipitation intensity over the Asian monsoon region compared with the remaining five datasets. CPC-UNI exhibites lowest precipitation over Myanmar and the neighboring regions. This is similar to the findings of Rana et al. (2015), who showed that CPC-UNI poorly simulates seasonal mean precipitation over the same region. ERA-Interim also tends to show lower values over India and the maritime continent. As noted earlier ERA-Interim is the only dataset which displays subdued monsoon activity (Fig. 3f) over the southeast Indian peninsula. As pointed out in the preceding section, most of the annual rain over this region occurs during the October–December period (Kripalani and Kumar 2004). JRA55 is similar to the ensemble mean among the reanalysis datasets, but relatively shows higher rainfall values over Southeast Asia and Southeastern Asia. MERRA reproduces less rainfall over Korea and Japan.

Open image in new window

In summary, the main inferences drawn from Fig. 3 are interesting. The highest resolution dataset (TRMM 0.25°) clearly depicts the heavy orographic rainfall zones over South Asia, in particular the west coast of India and hilly region of northeast India and adjoining Bangladesh and the Arakan/Myanmar coast. This may be due to the low-pressure systems that form over the Bay of Bengal and transport moisture over these regions. This high-resolution dataset also brings out zones of heavy rain events over South China and the adjoining Thailand region over Southeast Asia. These heavy rain events may be due to the tropical storms over the South China Sea striking the south China and the adjoining coasts. A recent study (Zhan et al. 2017) reported that western China including the Tibetan Plateau has experienced a significant change in the extreme events over the past decades. The GPCP data at 1° resolutions also displays these regions of heavy rain events, to a certain extent. The remaining five datasets depict heavy rain events in the 20–30 mm/day range. In general, regions of heavy rain events are better displayed by high-resolution datasets. Thus the spatial patterns for the same variable (here the 95th percentile precipitation) show some differences among the various datasets considered here. However, the spatial distributions of the 95th percentile scores resemble the patterns obtained for the mean seasonal rainfall pattern (Compare Figs. 2, 3). Such inferences have been documented earlier (Boers et al. 2016).

Another distinctive feature of extreme event is the shapes of right tails in the precipitation PDF for JJA in the seven datasets. This is illustrated by the differences between the 99th and the 90th percentile values (Fig. 4), GPCP and JRA55 are similar to ensemble of all the datasets. TRMM shows a particularly intense distribution of the right tail between the 90th and 99th percentiles, while APHRODITE, ERA and MERRA show a relatively even smaller value of the right tail than the others (Fig. 4).

Open image in new window

Over South Asia, western parts of India through central India up to northeast India; over East Asia major parts of China and Korea–Japan and over Southeast Asia Vietnam, northern parts of Philippines heavy rain events with differences exceeding 35 mm/day are clearly depicted by the TRMM dataset. Similar patterns are displayed by the GPCP, CPC, and JRA55 datasets but with differences in the 25–35 mm/day range. MERRA shows less variation over the entire domain. In fact ERA and MERRA hardly show any differences over the maritime continent. The differences between the 99th and 90th percentile displayed by ERA and MERRA are considerably less compared to other reference datasets. It means that the range of extreme value is narrow.

3.3 Frequency of heavy precipitation days

In terms of the frequency of heavy rainfall (Fig. 5), GPCP and TRMM display high frequent events of heavy rainfall (≥ 30 mm/day) over South, East as well as Southeast Asia. The frequency of heavy rain case displayed by ERA-Interim is near zero over the Indian subcontinent. Although, both CPC-UNI and APHRODITE are rain gauge based datasets, they appear to have large differences over maritime continents and India (Fig. 5). Furthermore, frequency of occurrence of heavy rain displayed by APHRODITE, ERA and MERRA over southeast Asia appear to be considerably less compared with the other four datasets. In addition, frequency of heavy rain events appear to be well captured by the TRMM dataset over west coast of India and northeast India compared to GPCP, CPC-UNI, and ERA-Interim. Regions surrounding the South China Sea i.e. South China, Vietnam, Borneo, Philippines and the Indonesian islands also are well captured. Even regions receiving moisture from the West Pacific i.e. southeast China and the Korea-Japan peninsula show regions of heavy rain events. While the APHRODITE dataset captures the heavy rain events over South Asia, ERA dataset does not display any heavy rain events. CPC captures these events over Southeast Asia. Hence, there are differences among the datasets in displaying these heavy rain events, in particular over Southeast Asia and the maritime continent.

Open image in new window

3.4 Ratio of extremes to climatology

To alleviate the possible uncertainties in rainfall characteristics among the various datasets, since each dataset have some difference in capturing the spatial distribution of mean and extreme rainfall, we presume higher similarities may be exhibited by the ratio of extreme precipitation to climatology. As shown in Fig. 6, over most of the region the ratios are greater than six. However, over the maritime continent the ratios are below four. In fact the ratios are below two as depicted by the ERA and MERRA datasets. Maximum ratios exceeding 12 are displayed over central and western parts of India by most of the datasets. Satellite and rain-gauge based sources tend to show increased agreement, as compared with the ensemble data; however, ERA and MERRA show systematic biases compared with distribution of the ensemble. Even though CPC-UNI is a rain-gauge dataset, the scaled value was relatively higher in areas between latitude 30°N and 40°N. The patterns for the ratios between the 95th percentile and the climatological mean rainfall intensity are similar between APHORODITE and JRA55 and also similar between ERA and MERRA. In addition, GPCP and TRMM resemble the ratio of extremes with higher values. Extremely high ratios north of 30°N are displayed by all the data sets, in particular for the CPC data due to considerably less seasonal rain intensity over these regions.

Open image in new window

Tables 3 and 4 show quantitative assessments of high percentile precipitation across the seven datasets for sub domains defined in Sect. 2.3. The consistency of high percentile of rainfall over the East Asia region is greater than that for India and Southeast Asia (Indochina). Over the East Asia GPCP, TRMM, and APHRODITE are in much better agreement with the ensemble as per the results of pattern correlation, standard deviation, RMSE, and skill score. For CPC-UNI, ERA-Interim, and MERRA, there is lower similarity of spatial distribution as compared with the ensemble. In particular, CPC-UNI has the largest inconsistency of high percentile precipitation and skill score (95th percentile: 0.215, 0.128), (99th percentile: 0.350, 0.161) over East Asia and Southeast Asia. Despite CPC-UNI (0.5° × 0.5°) having double the horizontal resolution of GPCP (1.0° × 1.0°), GPCP is more similar to APRODITE. This could reflect the discrepancy in number of stations, with APHRODITE using a more extensive gauge network than CPC-UNI (Gebregiorgis and Hossain 2015; Rana et al. 2015).

In the scaled values of higher percentile precipitation, the consistency among all datasets improves over East Asia, India, and Southeast (Indochina) according to pattern correlation, standard deviation, RMSE, and skill score. Notably, CPC-UNI has a low skill score while in scaled value the skill score increased (95th percentile: 0.849, 0.415) (99th percentile: 0.843, 0.443). Moreover, over Southeast Asia, all datasets have lower coherency on extreme values than over East Asia, and India, especially CPC-UNI, JRA55. However, when scaled, the performance of each dataset suggests improvement. Although each dataset may contain numerous systematic biases regarding the amount of higher percentile precipitation, our analyses show that they can be better for capturing the characteristics in terms of the ratio with respect to climatology. Exceptionally, scaled values over Southeast Asia for JRA55 bring worse outcomes in pattern correlations (95th percentile: 0.726, 99th percentile: 0.755) than original values (95th percentile: 0.191, 99th percentile: 0.532) .

3.5 Signal-to-noise (SNR) ratio

Figure 7 shows uncertainty ranges of climatology, daily mean for boreal summer, 95th percentile precipitation, the right tail of precipitation (99–90th percentile precipitation), and scaled precipitation across the datasets through the SNR distributions. The SNR spatial pattern for climatology distribution appears similar to that for the mean precipitation in JJA (Fig. 7a, b), while the SNR for the extreme value (95th percentile value) appears similar to the 99th–90th percentile pattern (Fig. 7 c, d). The SNR for the differences between the 90th and 99th percentile precipitation appear smaller than the 95^th percentile precipitation (compare Fig. 7c, d). In other words, the width of the right tail of the precipitation PDF may have larger uncertainties. The SNR values for the 95th percentile appear relatively higher than those of the 95th percentile precipitation over South and East Asia (compare Fig. 7c, e). Consequently, using a scaled value, systematic biases for regions with large uncertainties over heavy precipitation regions appear largely reduced. In summary the signal to noise spatial patterns, show higher ratios over East Asia compared to South and Southeast Asia. This may imply that precipitation data over East Asia are consistent among the datasets.

Open image in new window

3.6 Interannual variability

To examine uncertainty in the interannual variability of extreme precipitation across the datasets, the standard deviation of the 95th percentile precipitation for JJA have been calculated (Fig. 8). Higher uncertainties in their interannual variability of extreme value exist among the datasets, and especially strong for heavy rain regions. TRMM shows relatively strong interannual variability, while ERA-Interim, MERRA show relatively weak variability. GPCP, TRMM, APHRODITE, CPC-UNI, and JRA55 reproduced larger interannual variability over complex topography. But, ERA shows higher variability over northeast India and neighborhood. In summary, TRMM clearly shows the largest interannual variability over South Asia as well as East Asia. The spatial pattern for GPCP is similar, however with reduced variability compared with TRMM. Similar patterns are seen for CPC and JRA55 except for maritime continents and Myanmar.

Open image in new window

3.7 Decadal trends

Kim and Park (2016) suggested that the linear trend in precipitation may have larger uncertainty than in the mean and in the interannual variability among the reference datasets. Figure 9 shows the decadal trends of extreme rainfall (i.e., 95th percentile precipitation for JJA) in the datasets over the period 1998–2007 calculated through a linear regression coefficient. Except for APHRODITE, ERA-Interim, and JRA55, the rest shows a remarkable increasing trend over central India, which is consistent with the finding of Rana et al. (2015) and Prakash et al. (2014). Singh et al. (2013) found that APHRODITE has opposite trend signs for extreme precipitation intensity from the 2000s onwards when compared with the IMD dataset (2140 rain-gauge station).

Open image in new window

Decadal trends in the 95th percentile values indicate interesting results (Fig. 9). Increasing trends are observed over the central Indian region from the west to east; on the other hand, decreasing trends are observed just north over the foothills of the Himalaya. This is consistent with results shown by Preethi et al. (2017). Four of the datasets (GPCP, TRMM, APHRODITE and MERRA) are able to capture these features over the Indian subcontinent. While CPC data could capture only the increasing trend over the Indian region, ERA does not capture any of these trends. A threefold rise in extreme precipitation events over central India has been recently reported (Roxy et al. 2017; Goswami et al. 2006), Over East Asia, a wave-like pattern with increasing, decreasing, and increasing trends from north to south China are observed. Except the ERA dataset the remaining five datasets display this feature. This feature is also consistent with earlier studies (e.g. Preethi et al. 2017). Over Southeast Asia no trends are exhibited by any of these datasets. Distinctively, JRA55 has negative trends over Asian summer monsoon domain unlike the other datasets. Next we examine the outputs from the CMIP5 datasets.

4.1 Precipitation biases

The composite precipitation biases are illustrated in Fig. 13, the upper panels for the 95th percentile, the central panels for the 99th–90th percentile and the lower panels for the frequency of heavy rain events. The negative biases over northern and central parts of India amplify as we examine the high-resolution models through to the low-resolution models for the 95th percentile (Compare Fig. 13b–d). For the same statistic, similar inferences can be drawn over China. For the 99th–90th percentile the positive biases depicted by the high-resolution models change to negative biases as depicted by the low-resolution models (Compare Fig. 13f–h). Similar inferences can be drawn for the heavy rain events (Fig. 13j–l).

Open image in new window

A known source of error in the simulation of precipitation in current models is the convective parameterization schemes. Also the increase of model resolution influences simulating mean and extreme precipitation. This is due to improved representation of complex topography and land surface processes (Kendon et al. 2012). A comparison of the results for the three resolutions in Fig. 13 confirms the results shown in Kendon et al. (2012).

4.2 Comparison of low and high percentile

Figure 14 illustrates the performances of individual models using Taylor diagrams of the 75th, 95th, and 99th percentile values against the reference. Over East Asia, most models in the high-resolution group perform better in terms of 75th, 95th, and 99th percentile precipitation than the low-resolution group, and especially for the 99th percentile. These results are consistent with those of Jiang et al. (2015), who shows that models with higher resolution produce relatively smaller errors in extreme precipitation. The low-resolution group reasonably well reproduces low percentile (i.e., 75th percentile) precipitation, as shown by the correlation, standard deviation, and RMSE. The spatial correlation in all models over Southeast Asia (Indochina) is relatively high (0.5–0.9) despite having larger ratios of variance and RMSE. Over India, there exists a relatively large spread and a lower skill of spatial correlation for all models, and especially for variability of 99th percentile precipitation for models in the low- and high-resolution groups. In contrast, the medium-resolution group shows a relatively better performance for the 99th percentile precipitation. Over Southeast Asia (Indochina) there exists outliers in 75th, 95th, and 99th percentile precipitation belonging to the low-resolution group. Among all of the models considered, the best skill scores are for CMCC-CM, MRI-AGCM3.2H, and CAM5 (all in the high-resolution group), and the worst were for FGOALS-g2 and IPSL-CM5B-LR (all in the low-resolution group).

Open image in new window

To confirm whether similar characteristics exist for different GCMs, the distribution of mean and extreme value is evaluated statistically through pattern correlation, the spatial standardized deviation ratio, RSME, and Taylor skill score calculates on the daily mean and the 95th percentile precipitation. Extreme precipitation simulated by CMIP5 models shows higher pattern correlation than does the mean over East Asia, India, and South East Asia, but shows lower skill according to the spatial standardized deviation ratio, RSME, and Taylor skill score. Over India and South East Asia, the skill score of extreme values for the low-resolution group is noticeably lower, indicating that extreme values have a higher model resolution dependency compared with mean values, which is consistent with the findings of Volosciuk et al. (2015). Scaling the precipitation, the skill over India and Southeast Asia for the low-resolution group is remarkably improved as compared with the original value, which could imply that models with low-resolution fail to reproduce extreme rainfall owing to limitations of resolution or parameterization, but are able to moderately well capture the ratio of heavy rainfall with respect to climatology.

4.3 SNR distribution

The SNR distribution associated with mean and extreme value across the inter-models (Fig. 15) roughly appears similar to the datasets examined earlier. For daily mean and extreme rainfall over monsoon rainfall regions (i.e. Central India, East Asia, and Indochina), the SNR value for the high-resolution group is relatively higher than for the other groups, implying smaller inter-model spread. The relatively lower SNR for the medium- and low-resolution groups suggests that such models include large uncertainties. We find larger uncertainties in the right tail distribution of precipitation than we do for the daily mean and the 95th percentile precipitation over Central India, Southeastern China, Indochina, and the maritime continents. In scaled extreme precipitation, the uncertainties for all three groups are noticeably reduced over East Asia, India, and Southeast Asia as compared with the original 95th percentile precipitation. Most notably, the high-resolution group shows the largest consistency across inter-models.

Open image in new window