Skill of precipitation prediction with GCMs over north India ... - IIT Delhi

INTERNATIONAL JOURNAL OF CLIMATOLOGY Int. J. Climatol. (2014) Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/joc.3921

Skill of precipitation prediction with GCMs over north India during winter season P. R. Tiwari,a S. C. Kar,b U. C. Mohanty,c* S. Kumari,a P. Sinha,a A. Naira and S. Deya a

c

Centre for Atmospheric Sciences, Indian Institute of Technology Delhi, New Delhi, India b National Centre for Medium Range Weather Forecasting, Noida, India School of Earth Ocean and Climate Sciences, Indian Institute of Technology Bhubaneswar, Odisha, India

ABSTRACT: This study aims to analyse the skill of state-of-the-art of five general circulation models (GCMs) in predicting winter precipitation over northern India. The precipitation in winter season (December, January and February) is very important for Rabi crops in north India, particularly for wheat, as it supplements moisture and maintains low temperature for the development of the crops. The GCM outputs (seasonal mean forecasts issued in November) from various organizations are compared with the observed high-resolution gridded rainfall data obtained from India Meteorological Department (IMD). Prediction skill of such GCMs is examined for the period 1982–2009. The climatology, interannual standard deviation (ISD) and correlation coefficients have been computed for the five GCMs and compared with observation. It is found that the models are able to reproduce the climatology and ISD to varying degrees; however, skill of predictions is too low. Multi-model ensemble (MME) approaches have been employed. It is found that the weighted MME using multiple linear regression technique improves the prediction skill of winter precipitation over northern India. The teleconnection between the sea surface temperature (SST) and winter precipitation revealed that the SST over the Pacific Ocean affects the precipitation over north India in winter season. While this observed feature is represented well by some models with high fidelity, most models are unable to respond to SST variations in the Pacific Ocean in a realistic manner. Lagged correlations between the north India rainfall and SST over the Niño-3.4 region reveal that only two of the five GCMs get the observed simultaneous teleconnection correctly. Furthermore, only one of these two models has the observed phase lag with the strongest correlation as observed. KEY WORDS

north India; winter precipitation; predictability; general circulation models; MME

Received 21 June 2013; Revised 16 November 2013; Accepted 15 December 2013

1. Introduction The northern part of India, known as the ‘wheat bowl’ of the country, is a vast land mass consisting of six meteorological subdivisions, namely, Jammu and Kashmir (J&K), Himachal Pradesh (HP), Haryana, Punjab, Uttarakhand (UK) and west Uttar Pradesh. This region receives most of its precipitation during winter season [December, January and February (DJF)], which contributes around 15–20% to the annual precipitation over this region. Precipitation during winter season in north India is very important for Rabi crops, particularly for wheat, as it supplements moisture and maintains low temperature for the development of the crops. Precipitation in the form of snow over the hilly regions of north India also helps in glacier maintenance for the supply of water to rivers throughout the year, hydropower production, transport and logistics, etc. Therefore, precipitation during the winter season is critically important for the agriculture and economy of the country. * Correspondence to: U. C. Mohanty, School of Earth Ocean and Climate Sciences, Indian Institute of Technology Bhubaneswar, A2708, Toshali Bhawan, Satya Nagar, Bhubaneswar 751007, Odisha, India. E-mail: [email protected]

 2014 Royal Meteorological Society

Precipitation over this region is mainly associated with western disturbances (WDs) (Pisharoty and Desai, 1956; Mooley, 1957; Agnihotri and Singh, 1982; Kar and Rana, 2013), which bring heavy bursts of rain and snow. These extra-tropical low-pressure systems originate in the eastern Mediterranean Sea and are carried towards India by the subtropical westerlies, which are the prevailing winds blowing at north India’s range of latitude. Once their passage is hindered by the Himalayas, they release significant precipitation over the western Himalayas and adjoining plains. The frequency and amplitude of these WDs in a given month or season decides whether the winter season will experience above-normal or belownormal precipitation. The winter precipitation over this region has been relatively less explored due to its limited spatial extent and total amount. Yadav et al. (2009, 2010) have examined the precipitation variability over the north-west Indian region and provided a mechanism on how the El Nino/Southern Oscillation (ENSO) can influence this variability. They proposed a physical mechanism by which WDs are intensified over north-west India because of a baroclinic response to large-scale sinking motion over the western Pacific Ocean during the warm phase of ENSO. This response causes an upper level cyclonic circulation anomaly over north of

P. R. TIWARI et al.

India and a lower level anti-cyclonic anomaly over southern and central India. The cyclonic circulation anomaly intensifies the WDs passing over north-west India. Several operational centres around the globe routinely use general circulation models (GCMs) to prepare predictions at various space and time scales for Northern Hemisphere winter seasons. However, very few or no studies exist in documenting the skill of these models in predicting winter precipitation over northern India. Hatwar et al. (2005) did two case studies for intense WD, which affected the north-west India using the India Meteorological Department (IMD) operational limited area analysis and forecast system. This analysis showed that 24-hour model forecasts are in good agreement with the observations in respect of WD movement and intensification. Statistical/global dynamical models have also been used by Kripalani and Kumar (2004), Zubair and Ropelewski (2006) and Nayagam et al. (2009) for studying winter precipitation. In all the above-mentioned studies, the focus of the research has been WDs, which is a cause of heavy precipitation over the northern part of India. With the availability of climate predictions produced by several global dynamical models, techniques have been developed to combine the multi-model ensemble (MME) forecasts to a single reliable forecast that carries higher skills when compared to the individual member models. These include the simple ensemble mean (Doblas-Reyes et al., 2000; Stephenson and DoblasReyes, 2000; Palmer et al., 2004), regression-improved ensemble mean (Kharin and Zwiers, 2003) and MME (Krishnamurti et al., 2000; Yun et al., 2003). Kar et al. (2006) have used several multi-model approaches to estimate the economic values of the forecasts and have found that the MME schemes improve the value of the forecasts over the single model. Kug et al. (2008) in a comprehensive paper described the skill of many MME methods for seasonal prediction. Compared to a single control forecast, an ensemble forecast not only provides a more accurate estimate of the first moment (the mean) of the probability density function (PDF) of future atmospheric states but also provides higher order moment estimations such as the forecast error variance. Although, the GCMs are critically analysed in the context of south-west monsoon rainfall in many studies (Prasad et al., 2009; Pattanaik and Kumar, 2010; Kar et al., 2011), yet no systematic effort has been made to analyse the usefulness of GCMs for forecasting precipitation variability in winter season over northern India. No attempt has also been made so far to examine the usability of MME techniques for the prediction of wintertime rainfall over north India. Therefore, there is a huge gap in the understanding and prediction of wintertime precipitation over north India using GCM products. Keeping the above in view, a comparison between the observed winter precipitation and the hindcasts from state-of-the-art GCMs has been made in this study. The continuous development in GCMs in recent decades which represent the physical processes in the atmosphere, ocean and the land surface that simulate the global  2014 Royal Meteorological Society

climate system (Randall, 2000) has motivated this study to analyse the predictions of winter season rainfall in the GCMs. To carry out this study, climatology, interannual variability (IAV) and correlation coefficients have been computed for the five GCMs and compared with observation. An attempt has also been done to find the teleconnection patterns of the model predicted precipitation with the sea surface temperature (SST) variability and compared with the observed pattern. MME approaches have been employed. A simple arithmetic mean of the model predicted precipitation as well as weighted MME technique has been used and their skill has been evaluated. Brief description of the observed data and GCMs products as well as the analysis methodologies are provided in Section 2. The results are discussed in Section 3. Finally, this study is concluded in Section 4.

2. Data and method of analysis 2.1. Observed data High-resolution (1◦ × 1◦ ) (Rajeevan et al., 2006) rainfall data from IMD over the Indian land mass (6.5◦ N–38.5◦ N and 66.5◦ E–100.5◦ E) for the winter season (DJF) has been used as the observational data. These data are based on 2140 rain gauge stations. The station data have been interpolated to the specific grid points using objective analysis. It is to be noted that DJF seasonal rainfall data for a year is constructed by taking average of that year’s December rainfall and next year’s January and February rainfall. For example, values of 1982 DJF seasonal rain is obtained by averaging rainfall values of December 1982, January 1983 and February 1983. In addition to the rainfall data, SST data are also used in this study. These data are obtained from NOAA ERSSTv3b extended reconstructed monthly sea surface temperature data set for the whole globe (Smith et al., 2008). 2.2. Model products In this study, precipitation outputs from state-of-the-art of five global models are used. These outputs are 1month lead prediction for precipitation. The 1-month lead seasonal prediction means that the seasonal average precipitation predictions for DJF is obtained by initializing the GCMs around November 2001. The GCMs used in this study are initialized either towards the end of October or in the beginning of November. The date of initialization of each model is given in Table 1. In this study, one-tier models used are NCEP_CFSv2, MOM3_AC1 and MOM3_DC2. The two-tier model is ECHAM_CFS, which is an atmosphere-only model, forced with predicted SSTs from the CFS. Among the one-tier models, NCEP_CFS2 is a coupled ocean atmosphere model in which the oceanic component is Geophysical Fluid Dynamics Laboratory (GFDL) Modular Ocean Model (MOM) version 4. The other set of models, MOM3_AC1 and MOM3_DC2, uses MOM3 ocean model coupled with ECHAM atmospheric model. These Int. J. Climatol. (2014)

PREDICTABILITY OF GCMS OVER NORTH INDIA DURING WINTER SEASON

Table 1. Description of GCMs/Atmosphere-Ocean General Circulation Models (AOGCMs). Model

Resolution

Atmosphere General Circulation Model (AGCM)

Ocean General Circulation Model (OGCM)

ECHAM_CFS ECHAM_GML

(T42) 2.7◦ × 2.8◦ (T42) 2.7◦ × 2.8◦

ECHAM4p5 ECHAM4p5

24 12

Roeckner et al., 1996 Roeckner et al., 1996; Lee and De Witt, 2009

MOM3_AC1

(T42) 2.7◦ × 2. 8◦

ECHAM4p5

CFS-predicted SST 30th October CFS-predicted SSTs 1st November prescribed over the tropical Pacific basin (semi-coupled) MOM3 1st November (anomaly-coupled)

24

MOM3_DC2

(T42) 2.7◦ × 2. 8◦

ECHAM4p5

MOM3 (direct-coupled)

1st November

12

NCEP_CFSv2

(T126) 0.9◦ × 0.9◦

GFS (2009 version) MOM4

9th, 14th, 19th, 24th, 29th November and 3rd December

24

Roeckner et al., 1996; Pacanowski and Griffes, 1998 Roeckner et al., 1996; Pacanowski and Griffes, 1998 Saha et al., 2006

two models differ in their coupling scheme. The former is anomaly coupled (only anomalies from the mean seasonal cycle are coupled) while the later is directly coupled. ECHAM_GML is a semi-coupled model with a mixed layer model for oceans except for the Pacific basin where predicted SSTs from CFS are used. Predictions of all these models are collected from the International Research Institute for Climate and Society (IRI), Columbia University, New York, data library except for the CFSv2. The CFSv2 model products are obtained from National Center for Environmental Prediction (NCEP). Brief descriptions of these models are presented in Table 1. We have considered the period of study from 1982 to 2009, as this is the common period for data, which are available for all the models. As these GCMs are in very low resolution (except for CFSv2), a bi-linear interpolation technique is employed to bring them to observed data’s resolution (1◦ × 1◦ grid). Together with the precipitation data, SST hindcast runs from these models are also used in the analysis. 2.3. Analysis methods The first step of diagnosing these GCMs is to examine the climatology and IAV that provides the very basic step in evaluating how well these GCMs tend to capture the observed characteristic features. Several skill metrics have been used to evaluate the performance of the GCMs in predicting the seasonal mean precipitation patterns over the domain of interest. In this study, firstly, a comparison between the model simulated climatology and the observed climatology has been made over northern India from 1982 to 2009. For this purpose, the precipitation model data were interpolated onto the same grid resolution 1◦ × 1◦ latitudelongitudes for performing spatial inter comparison and to avoid dependence on resolution. The climatology is obtained by finding average of 27 years of rainfall data at each of the Indian grid point. The basic means of analysing how model behaves compared with observation is to determine the correlation coefficient between  2014 Royal Meteorological Society

Date of initializations

Ensemble member

Reference

the two time series, which gives a degree of association between them and is defined as Cov (X , O) CC = σx σo n 1 Xi − X Oi − O (n−1) =

1 (n−1)

i =1 n i =1

Xi − X

2

n 2 1 × (n−1) Oi − O

,

i =1

(1) where σ x and σ o are the standard deviations of precipitation for model and observation. The total number of points over space is n, X i and O i are the model and observed precipitation for i th point and X and O are the mean value averaged over space for model and observed precipitation. The root mean square error (RMSE) is used as one of the skill measures of GCMs. RMSE is computed as n (Xi − Oi )2 i =1 . (2) RMSE = n Bias in the model predictions is defined as X −O Bias (in %) = × 100. (3) O A comparison of IAV between observation and model simulations is analysed over the entire country from 1982 to 2009. The IAV is obtained by finding standard deviation of 27 years of rainfall data at each of the Indian grid point. In order to examine precipitation and circulation difference between excess and deficit years, the precipitation data have been divided in to three categories. According to a normal (Gaussian) distribution, if the precipitation values for a year lies between ± 0.43 standard deviation, Int. J. Climatol. (2014)

P. R. TIWARI et al.

Figure 1. Precipitation climatology (mm) from observation and models from 1982 to 2009 for December to February.

that year is categorized as a normal year and other years are categorized as either above-normal or below-normal years. Excess or deficit years are located on the tails of the Gaussian distribution. In order to identify excess (top 85 percentile) and deficit year (bottom 15 percentile), ±1 standard deviation criteria has been used. The years having standardized precipitation anomaly greater than 1 are considered as excess years, while years having less than −1 standardized precipitation anomaly are defined as deficit years. Therefore, among 27 years, there are 4 years in the category of excess precipitation (1990–1991, 1994–1995, 1995–1996 and 1997–1998) and 4 years in the category of deficit precipitation (1996–1997, 2000–2001, 2004–2005 and 2007–2008). In addition to above, 4 years in the category of normal precipitation (1988–1989, 1993–1994, 2001–2002, 2003–2004) years have also been chosen. After choosing the years in a particular category, mean of the meteorological variables  2014 Royal Meteorological Society

(composite analysis) in that category has been computed. Results are further analysed by computing the differences between the composite data for excess and normal and deficit minus normal precipitation years. In this study, two MME methods are also used. In the first method, all the individual member models have been assigned same weight while carrying out ensemble average, i.e. simple averaging of all the individual models. The second MME method is a multiple linear regression (MLR) approach in which individual member models are assigned weights based on their skill during the training period following leave-one-out cross-validation technique. 3. Results and discussion As discussed above, the study is aimed at analysing the prediction skill of winter rainfall in five GCMs. Int. J. Climatol. (2014)

Precipitation (mm day-1)


Year

Figure 2. Time series of observed and model predicted area-averaged precipitation (mm day−1 ) over north India from 1982 to 2009 for December to February.

The GCMs are individually analysed for the prediction of winter rainfall on the basis of certain statistical measures as described above. Two MME approaches have been employed: (1) simple combination (MME; arithmetic mean) of all the GCMs and (2) weighted MME using MLR. 3.1. Observed analysis The observed climatology of precipitation (seasonal, total) is depicted in Figure 1(a). It is seen that the maximum value of rainfall occurs during winter is over northern Kashmir, which is in the range of 250–350 mm. Over north-east part of India, the climatology of rainfall lies in the range of 50–150 mm. Maximum rainfall of about 150 mm can be observed over Arunachal Pradesh and then the rainfall gradually decreases westward. On the other hand, in the southern part of peninsular India especially over eastern part of Tamil Nadu, the value of climatology is in the range of 100–150 mm. So it can be seen that the maximum rainfall occurs during the winter season over northern parts of India compared to rest of the country. Observed area-averaged time series of precipitation over north India have been compared with precipitation from the models in Figure 2. Over the region of interest, maximum observed rainfall was in 1994 and minimum rainfall occurred in 2000 (∼1 mm day−1 ). Figure 3(a) represents the observed interannual standard deviation (ISD). In observation, the standard deviation is found to be high over north India and over some pockets of north-east India and southern peninsular India. The values of ISD over these regions range from 20 to 35 mm. It can be seen that the highest ISD is found over the areas experiencing large amount of rainfall during the season that is the northern part of India in observation. Over rest of India, the values of ISD are in between 10 and 20 mm. 3.2. Skill of predictions by GCMs The climatology of precipitation (seasonal, total) simulated by each state-of-the-art of five GCMs is compared  2014 Royal Meteorological Society

against observation and is depicted in Figure 1(b)–(f). It is seen that only the NCEP_CFS2 model is able to reproduce the high-rainfall zone to some extent, i.e. the region is analogous to what is observed over the Kashmir region (Figure 1(d)). This model shows a zone of high precipitation over north India as well as over south-eastern parts of peninsular India. The patch of high maxima zone seems to be extending towards southward in this model than observation. All the other models show a pocket of high-rainfall zone over north India, but the magnitude over the region is found to be very less (less than 250 mm) compared to the observation. In Figure 2, it is seen that the models generally produce very less rainfall over north India compared to observed rainfall. In 1994, observed rainfall during the season was more than 4 mm day−1 . However, rainfall values from the models are within 1–2 mm day−1 , indicating large bias in the model rainfall. Moreover, it is seen from the figure that IAV of the model-simulated rainfall does not agree with that of observed rainfall. Figure 3(b)–(f) represents the ISD of precipitation from individual models. Almost all the models are able to show the total variance of precipitation as in observation; however, the MOM3_DC2 is somewhat able to show the region of maximum ISD similar to observation. In the case of NCEP_CFS2, the ISD is less than observation, especially over eastern parts of India. Over northern India, the ISD is less than 30 mm in observations as well as in all the models except it reaches up to 40 mm in CFS2. Therefore, the figure suggests that most of the GCMs are able to predict the observed magnitude of ISD. Figure 4(a)–(e) represents the bias (in %) between the model and observed precipitation. It is seen that all the models (except NCEP_CFS2) show almost the same bias pattern, ranging from 50 to 200% over J&K, HP and UK region. On the other hand, the NCEP_CFS2 model has stronger bias over eastern part of J&K and north-eastern parts of India. As shown in Figure 4(c), this model has higher bias up to 500% of observed precipitation. Figure 5(a)–(e) outlines the spatial correlation coefficient between the precipitation from the models and observation. It can be clearly seen from the diagram that the values of correlation coefficients for most models over northern part of India are low except MOM3_DC2 and MOM3_AC1 models. The best among the models is MOM3_DC2 which has correlation coefficients of 0.4 and above over the region of interest. The other models have varying degrees of skill and correlation coefficients range from 0.3 to as low as −0.2. It may be noted that for 27 years of data that are analysed in this study, correlation values greater than 0.33 are statistically significant at 95% confidence level. Therefore, most models do not have statistically significant correlation skill for precipitation over the north Indian region during winter season. Figure 6 shows the RMSE of precipitation between the models and observations. It can be noticed that the RMSE values ranges from 0.5 to 3 mm day−1 over J&K, HP and UK region. The maximum RMSE in case of all the models (except NCEP_CFS2) are seen over north-west Kashmir Int. J. Climatol. (2014)

P. R. TIWARI et al.

Figure 3. IAV for precipitation between models and observation from 1982 to 2009 for December to February.

while for NCEP_CFS2 model higher RMSE values are located over the eastern part of J&K. A Taylor diagram (Taylor, 2001) is presented in Figure 7. In this diagram, the skill of individual GCMs for prediction of rainfall during winter at the all-India level in terms of correlation, RMSE and standard deviation is shown. The figure clearly indicates the significant correlation skill of three coupled GCMs (MOM3_DC2, ECHAM_CFS and MOM3_AC1) out of the five GCMs used in the study with less RMSE. Also from Figure 7, a smaller, or even negative, correlation is observed in  2014 Royal Meteorological Society

the case of ECHAM_GML and NCEP_CFSv2 with very large RMSE. A composite analysis has been carried out by computing the precipitation anomalies between excess minus normal and deficit minus normal precipitation years. Figures 8(a) and 9(a) show observed precipitation difference between excess minus normal and deficit minus normal precipitation years. It is seen in Figure 8(a) that over entire northern and central India, a coherent positive precipitation pattern has emerged and the positive difference ranges from 0.2 to 0.6 mm day−1 over the Int. J. Climatol. (2014)


Figure 4. Bias (in %) for precipitation between models and observation from 1982 to 2009 for December to February.

northern part of India. On the other hand, over north-east and south peninsular India, it shows negative rainfall difference. In Figure 9(a), a coherent negative precipitation pattern has emerged and the negative difference ranges from 0.2 to 0.6 mm day−1 over entire northern and central India. While over north-east and south peninsular India, it shows positive rainfall difference. In order to study the upper air circulation pattern, composite analysis for wind at 500 hPa has also been carried out. For this purpose, the difference of winds between excess and normal and deficit and normal years has been computed. For excess minus normal, the observation (Figure 10(a)) shows westerly anomalies of 1–2 m s−1 over central part of India succeeded by cyclonic flow due to hindrance of Himalayas. As already mentioned, precipitation occurs over northern parts of India when WDs pass over the region forming cyclonic anomaly over J&K and adjoining regions. This feature is well brought out in observed data as years with excess precipitation have higher frequency  2014 Royal Meteorological Society

and intensity of WDs. On the other hand in the case of deficit minus normal, the observation (Figure 11(a)) shows anomaly of strong easterlies (1 m s−1 ) over central part of India proceeded by anti-cyclonic flow due to hindrance of Himalayas. A critical evaluation of performances of the GCMs is made on the basis of composite analysis of excess minus normal and deficit minus normal precipitation years (Figures 8 and 9). In case of individual GCMs, it has been found that only MOM3_DC2 and MOM3_AC1 (Figure 8(c) and (f)) are able to bring out the positive rainfall (∼ 0.2 and ∼ 0.1 mm day−1 ) over north Indian region. On the other hand, rest of the models show just the opposite pattern of observation. None of the models are able to bring out the large-scale positive precipitation difference over northern and central parts of India in the difference plots. In the case of deficit minus normal precipitation years (Figure 9), none of the models are able to bring out the large-scale negative Int. J. Climatol. (2014)

P. R. TIWARI et al.

Figure 5. Spatial correlation pattern of the GCMs. The correlations were calculated between observed precipitation and model simulations from the period 1982–2009 for December to February.

precipitation difference over northern and central parts of India. In order to study the performance of GCMs in simulating the upper air circulation, composite analysis of wind at 500 hPa has been carried out for the individual GCMs. It has been found that only MOM3_DC2 and ECHAM-CFS models (Figures 10(b) and (e) and 11(b) and (e)) are able to bring out the westerly/easterly flow over central India. MOM3_DC2 model (Figures 10(e) and 11(e)) is showing the westerly/easterly with less strength (1–1.5/0.6–1 m s−1 ) followed by cyclonic/ anti-cyclonic flow over northern India. On the other hand, ECHAM_CFS model (Figures 10(b) and 11(b)) is showing the westerlies/easterly over stretched area and is of less strength. It can also be noticed from the figure that the ECHAM_CFS model is not able to capture the cyclonic/anti-cyclonic flow over northern India while it is captured by MOM3_DC2 model. Apart from these two models, all the other models are not able to delineate these features. Therefore, it is seen that models predict their own excess, deficit and normal years and those do not match with observed excess, deficit and normal precipitation years. As a result, when composite difference  2014 Royal Meteorological Society

plots based on observed years are made, the differences are small (statistically not significant) as well as their patterns do not match with observed patterns. So, overall, the above analysis suggests that models are able to replicate some aspects of the observed precipitation climatology to varying degrees of accuracy. One of the one-tier models shows the extreme precipitation zone over northern India but it underestimates the precipitation over other parts of the country. The composite analysis shows that only two models (MOM3_DC2 and ECHAM_CFS) of five models are able to show the rainfall and upper air wind anomaly up to certain extent. Further between these two models, the performance of MOM3_DC2 is marginally better. 3.3. Skill of MME predictions 3.3.1. Simple MME As a first step, a MME method is used. In this method, all the individual member models have been assigned same weight while carrying out ensemble average, i.e. simple arithmetic mean of all the individual models. Int. J. Climatol. (2014)


Figure 6. RMSE (mm day−1 ) of precipitation from the models against observations from 1982 to 2009 for December to February.

Hagedorn et al. (2005) have described the rationale behind the success of such MME techniques for seasonal prediction. Figure 1(g) shows the climatology obtained from ensemble mean of the GCMs. It can be seen that the climatology of MME is able to delineate the climatological precipitation reasonably well compared to individual models over J&K and HP regions. However, it underpredicts the rainfall (50–200 mm) against the IMD observation (50–350 mm). The ISD for MME is shown in Figure 3(g). It is noticed from the diagram that the ISD of MME (10–20 mm) is smaller over J&K, HP and UK regions compared to IMD observation (10–30 mm). Figure 4(f) represents the bias of MME prediction. Over the northern part of Kashmir, the bias of MME model is reduced compared to the individual models. The correlation of the MME predictions with observed precipitation is shown in Figure 5(f). Whereas the MME predictions have a good correlation (0.2–0.4) over central and north-east India, they do not have any skill over J&K, HP and UK regions. In fact, the correlations are mostly negative over the region of interest. It may be noted that at least one model (MOM3_DC2) has positive  2014 Royal Meteorological Society

correlations over northern India. It is seen that the skill of MME predictions is worse than the skill of the best model among the models used in this study. This is due to the fact that the forecasts from other poorer models dominate when equal weights are given to all member models of the MME process. Therefore, despite the scientific rationale behind the success of MME predictions by computing simple arithmetic means of all the available models, for the northern Indian region, the MME predictions of precipitation in winter seasons are not at all useful. Figure 6(f) shows the RMSE for MME model. It can be noticed from the diagram that the RMSE over northwest part of Kashmir is reduced for the MME compared to individual GCMs. The Taylor diagram (Figure 7) indicates that the MME has the best correlation (0.42) with least RMSE compared to individual GCMs at allIndia level. However, for the region of interest, the simple MME scheme does not improve the seasonal mean predictions of precipitation. The MME scheme also fails to bring out the observed excess minus normal/deficit minus normal precipitation pattern in the composite plot shown in Figures 8(g) and 9(g) respectively. Int. J. Climatol. (2014)

P. R. TIWARI et al.

ECHAM_CFS MOM3_AC1 NCEP_CFSv2 ECHAM_GML MOM3_DC2 MME

Figure 7. Taylor diagram for the all-India average precipitation prediction skill of the GCMs.

3.3.2. Weighted MME For carrying out weighted MME mean, point-by-point MLR method has been employed (Yun et al., 2003; Kar et al., 2011). Seasonal mean (DJF) precipitation from observations and the global models have been used for the period 1982–2009 (27 years). A major component of weighted MME forecast techniques is training of forecast data set. The prediction skill from these techniques during the forecast phase could be degraded if the training was executed with poorer forecasts. That means the prediction skill is improved when higher quality training data set is deployed for the evaluation of the multi-model bias statistics (Yun et al., 2003). Cross-validation technique has been used in which each year has been successively withheld from the training data set, and the remaining 26 years have been used for calculation of the model and observed statistics (i.e. seasonal means and regression coefficients). These means and regression coefficients are used for calculating the forecast for the verification year (the year that was withheld). It may be noted here that cross-validation with 1-year withheld increases the bias and leakages. If one follows the procedure of withholding 3–5 years of contiguous years and validating for the central years of the withheld year, such bias and leakages get reduced by reducing the negative cross-validation. In the case of weighted MME method that is employed in this study, training phase plays an important role. Krishnamurti et al. (2000), Yun et al. (2003) and Kug et al. (2008) have highlighted the need of longer period of training data for obtaining better skill of MME predictions. This study uses only 27 years of hindcast data, which is not very long, and for such shorter period of hindcast data, the MME predictions are developed using leave-one-out cross-validation. Therefore, in this study, weights are computed for individual models by withholding 1-year  2014 Royal Meteorological Society

data successively. If large number of years is withheld, number of years in training phase would get reduced affecting the skill of predictions. Several studies such as Acharya et al. (2011), Singh et al. (2012) and Sinha et al. (2013) have utilized MME and PCA/CCA approaches for rainfall prediction. Computation of regression coefficients with singular value decomposition (SVD) technique has several advantages, which include removal of the singular matrix problem while calculating covariance among models, which cannot be entirely solved with the Gauss–Jordan elimination method. In this study, SVD technique has been employed similar to Yun et al. (2003) for the computation of the regression coefficients. This method is referred in the following paragraph as MLR method. Figure 12 shows the correlation of the precipitation obtained from the MLR scheme with that of observed precipitation. It is seen that the weighted MME scheme is able to improve the prediction skill over northern India. As compared to individual models and the MME, the correlation values are positive over J&K, HP and UK regions. However, the correlation values are very low and mostly statistically insignificant at 95% confidence level (0.33 for 27 years of data), except over some regions in the domain of interest. Correlation values are lower over central parts of India in MLR compared to the MME scheme. Therefore, even though the skill of MLR is marginally higher over the domain of interest, significant improvement is not achieved when this point-by-point MLR approach is utilized to develop the MME predictions. Multi-model linear regression forecast is superior to all other linear combinations of the individual model forecasts when very large training data sets are available for estimating the regression coefficients. In practice, the coefficients depend on the estimates of covariance of model with observation and Int. J. Climatol. (2014)


Figure 8. Difference of composite precipitation for excess and normal years (mm day−1 ) from observation and the models.

between models. In these circumstances, estimating too many regression coefficients leads to over fitting that causes a degradation in skill (Kharin and zwiers, 2003). It may be noted that the total period of data used in this study is only from 1982 to 2009, which may not be very large for estimation of regression coefficients. 3.4. Teleconnection patterns ENSO is known to be a major forcing for the year-toyear climate variability. Association between ENSO and summer monsoon rainfall over India has been rigorously studied (Sikka, 1980; Rasmusson and Carpenter, 1983),  2014 Royal Meteorological Society

and it has been observed that this relationship has been weakening in recent two decades (Krishna Kumar et al., 1999). Influence of ENSO on north-east monsoon rainfall by Jayanthi and Govindachari (1999) has brought out that Tamil Nadu received record rainfall in 1997 that happened to be one of the strongest El Nino episodes. Yadav et al. (2009, 2010) have examined the precipitation variability over the north-west Indian region and provided a mechanism on how the ENSO can influence this variability. However, no studies exist on the performance of the GCMs in bringing out the observed teleconnection patterns for the winter precipitations in winter season Int. J. Climatol. (2014)

P. R. TIWARI et al.

Figure 9. Difference of composite precipitation for deficit and normal years (mm day−1 ) from observation and the models.

over northern parts of India. The study of teleconnection patterns in the models, i.e. the remote influence of SST on DJF precipitation over north India, is necessary as it is one of the major factors influencing the performance of GCMs. Therefore, in this section, an effort has been made to explore the suitable linkages in terms of correlation coefficients between global SST and observed area averaged winter rainfall over north India from the period 1982–2009. The correlation patterns for each individual model with corresponding SSTs have also been computed and are shown in Figure 13. It can be noticed in Figure 13(a) that there is a significant positive relationship between observed precipitation  2014 Royal Meteorological Society

and eastern equatorial Pacific SSTs. This increase in strength of positive correlation since 1980 has been noted by Zubair and Ropelewski (2006). Positive correlation can also be noticed in the Indian Ocean. It is noticed that even if the SST predictions by individual models are reasonably good (figure not shown), the SST precipitation (over north India) remote response is not well simulated by these models. Most of the models are not able to capture the positive relationship between rainfall over north India and SST over eastern equatorial Pacific except MOM3_DC2 and MOM3_AC1 models. The possible reason for such behaviour of these models (ECHAM_GML, ECHAM_CFS and NCEP_CFS) could Int. J. Climatol. (2014)


Figure 10. Difference of composite winds for excess and normal years (m sec−1 ) from observation and the models.

be that the internally generated variability of precipitation in the model is not governed by with the SST variability over the Pacific Ocean. Teleconnection patterns for MOM3_DC2 and MOM3_AC1 show that the models are able to bring out the remote response rather well. However, the model’s rainfall has large correlations with southern Pacific SST that is not delineated in the observations. Summing up the above discussion, it can be concluded that the SST over Pacific Ocean and Indian Ocean affects the rainfall over north India in winter season. Some models such as MOM3_DC2 and MOM3_AC1 represent this observed feature with high fidelity. In order to examine the phase relationship of teleconnection between the precipitation and SST, monthly lead/lag (−1, 0, +1 month) correlations between north India rainfall (t) and Niño-3.4 (5◦ S–5◦ N, 120◦ –170◦ W) SST (t + lag) are computed. The selection of the Niño-3.4 region is done as the strongest correlations between north India rainfall and SST occur over this region in observed data (Figure 13(a)). The correlations are also calculated separately for each of the five models. Though lead/lag  2014 Royal Meteorological Society

correlations have been computed for each month of the winter season, we only examined the correlations for the month of January, as for other winter months, the results are essentially similar. In observations, simultaneous correlation (at 0 lead) between Nino 3.4 SST and precipitation over north India is maximum (Figure 14). Positive correlations also occur at 1-month lead or lag. The observed maximum correlation at 0 lead suggests that maximum intensity of the monthly (DJF) precipitation is potentially influenced by the Niño-3.4 SST of the same month. Among the GCMs, except two models (MOM3_AC1 and MOM3_DC2), all the other models are not able to capture this relationship. While two of the models reasonably represent the predictability, seen as the near-zero correlations during these months, the ECHAM_GML, ECHAM_CFS and NCEP_CFS2 models are incorrect in this respect, with the presence of pronounced negative correlations from the preceding winter months. Of the five models, MOM3_DC2 best captures the timing in the relationship correctly. Overall, the results presented so far indicate that the relationship between north India precipitation and Int. J. Climatol. (2014)

P. R. TIWARI et al.

Figure 11. Difference of composite winds for deficit and normal years (m sec−1 ) from observation and the models.

Niño-3.4 SST is highly correlated and it has been brought out well by the two (MOM3_AC1 and MOM3_DC2) models.

4.

Conclusion

The skill of five different state-of-the-art GCMs is examined for the period 1982–2009 in predicting winter precipitation over northern India, which is critically important for agriculture. Seasonal hindcast precipitation data from these GCMs have been obtained at 1-month lead. The major findings of the study are enumerated as follows: • The GCMs in general underestimate the observed climatology and IAV of precipitation. Though the NCEP_CFS-coupled model shows the extreme rainfall zones reasonably well, it has lower correlation with observation as far as IAV is concerned. On the other hand, the other two-coupled models (MOM3_DC2 and MOM3_AC1) underestimate the high rainfall zones but they have reasonable correlations over north-west and north-east part of India. Most of the GCMs are able to predict the observed IAV magnitude to some extent  2014 Royal Meteorological Society

Figure 12. Skill of weighted MME predictions of precipitation.

but none of the models is able to depict the observed IAV correctly. • A simple multi-model (MME) ensemble approach with all the models getting same weight does not improve Int. J. Climatol. (2014)

PREDICTABILITY OF GCMS OVER NORTH INDIA DURING WINTER SEASON January_Rain (t) vs SST (t+Lag) 0.7 0.6 0.5 0.4

Correlation Coeffiicient

0.3 0.2 0.1 0 -0.1 -0.2 -0.3 -0.4 -0.5 -1

1

0 Lag IMD

ECHAM_CFS

MOM3_AC1

NCEP_CFS2

ECHAM_GML

MOM3_DC2

Figure 14. Correlation between precipitation over north India and lag −1, 0, +1 monthly SST over Nino 3.4 region.

much the forecast skill. Weighted MME approach using MLR technique has also been employed. It is found that the prediction skill is improved; however, correlation values are not statistically significant over most parts of India. • Further, to understand the teleconnection pattern, the remote influence of SST is analysed and it is seen that the SSTs over Pacific Ocean and Indian Ocean affect the rainfall during winter season over north India. Most of the GCMs (except MOM3_DC2 and MOM3_AC1 coupled models) do not respond to SST variability over the Pacific in a realistic manner. Therefore, in view of the above discussion, it can be concluded that to further improve the precipitation predictions over northern India from GCM outputs, it is necessary to use appropriate statistical post-processing techniques such as the model output statistics approach. Acknowledgements This research has been conducted as part of the project titled “Precipitation and temperature variability and extended range seasonal prediction during winter over Western Himalayas” at IIT, Delhi, sponsored by the Snow Avalanche Study Establishment (SASE), Chandigarh. We gratefully acknowledge the International Research Institute for Climate and Society (IRI), Data Library group for making five of their GCM-based seasonal forecasting systems available to this study. Also, the authors are sincerely thankful to the India Meteorological Department (IMD) for providing the gridded rainfall data for this study. References

Figure 13. Correlation of observed SST with the DJF precipitation over north India (a) from observation; (b) MOM3_AC1 model; (c) MOM3_DC2 model and (d) ECHAM-GML model.


Acharya N, Kar SC, Kulkarni MA, Mohanty UC, Sahoo LN. 2011. Multi-model ensemble schemes for predicting northeast monsoon rainfall over peninsular India. J. Earth Syst. Sci. 120: 795–805. Agnihotri CL, Singh MS. 1982. Satellite study of western disturbances. MAUSAM 33: 249–254. Doblas-Reyes FJ, Deque M, Piedelievre JP. 2000. Multi-model spread and probabilistic seasonal forecasts in PROVOST. Q. J. R. Meteorol. Soc. 126: 2069–2088. Int. J. Climatol. (2014)

P. R. TIWARI et al. Hagedorn R, Doblas-Reyes FJ, Palmer TN. 2005. The rationale behind the success of multi-model ensembles in seasonal forecasting – I. Basic concept. Tellus 57: 219–233. Hatwar HR, Yadav BP, Rama Rao YV. 2005. Prediction of western disturbances and associated weather over Western Himalayas. Curr. Sci. 88: 6–25. Jayanthi N, Govindachari S. 1999. El-Nino and northeast monsoon rainfall over Tamilnadu. MAUSAM 50: 2–217. Kar SC, Rana S. 2013. Interannual variability of winter precipitation over northwest India and adjoining region: impact of global forcings. Theor. Appl. Climatol., DOI: 10.1007/s00704-013-0968-z. Kar SC, Hovsepyan A, Park CK. 2006. Economic values of the APCN multi-model ensemble categorical seasonal predictions. Meteorol. Appl. 13(3): 267–277. Kar SC, Acharya N, Mohanty UC, Kulkarni MA. 2011. Skill of monthly rainfall forecasts over India using multi-model ensemble schemes. Int. J. Climatol. 32: 1271–1286. Kharin VV, Zwiers FW. 2003. Improved seasonal probability forecasts. J. Clim. 16: 1684–1701. Kripalani RH, Kumar P. 2004. Northeast monsoon rainfall variability over south peninsular India vis-à-vis the Indian Ocean dipole mode. Int. J. Climatol. 24: 1267–1282. Krishna Kumar K, Rajagopalan B, Cane A. 1999. On the weakening relationship between the Indian monsoon and ENSO. Science 284: 2156–2159. Krishnamurti TN, Kishtawal CM, Shin DW, Williford CE. 2000. Multimodel superensemble forecasts for weather and seasonal climate. J. Clim. 13: 4196–4216. Kug JS, Lee J, Kang I-S, Wang B, Park CK. 2008. Optimal multi model ensemble method in seasonal prediction. Asia Pac. J. Atmos. Sci. 44(3): 259–267. Lee DE, De Witt DGA. 2009. New Hybrid Coupled Forecast System Utilizing the CFSSST, 2009 Forecasts, Retrieved February 28, 2010. http://portal.iri.columbia.edu/portal/server.pt/gateway/PTARGS_0_ 49 72_5734_0_0_18/ NOAAabstract2009. Mooley DA. 1957. The role of western disturbances in the production of weather over India during different seasons. Indian J. Meteor. Geophys. 8: 253–260. Nayagam LR, Janardanan R, Ram Mohan HS. 2009. Variability and teleconnectivity of northeast monsoon rainfall over India. Glob. Planet. Change 69: 225–231. Pacanowski RC, Griffes SM. 1998. MOM 3.0 Manual . NOAA/ Geophysical Fluid Dynamics Laboratory: Princeton, NJ; 608. Palmer TN, Alessandri A, Andersen U, Cantelaube P, Davey M, Delecluse P, Deque M, Diez E, Doblas-Reyes FJ, Federsen H, Graham R, Guladi S, Gueremy J-F, Hagedorn R, Hoshen M, Keenlyside N, Latif M, Lazar A, Maisonnave E, Marletto V, Morse AP, Orfila B, Rogel P, Terres J-M, Thomson MC. 2004. Development of a European Multimodel Ensemble System for Seasonal to Interannual Prediction (DEMETER). Bull. Am. Meteorol. Soc. 85: 853–872. Pattanaik DR, Kumar A. 2010. Prediction of summer monsoon rainfall over India using the NCEP climate forecast system. Clim. Dyn. 34: 557–572.


Pisharoty PR, Desai BN. 1956. Western disturbances and Indian weather. Indian J. Meteor. Geophys. 8: 333–338. Prasad K, Dash SK, Mohanty UC. 2009. A logistic regression approach for monthly rainfall forecasts in meteorological subdivisions of India based on DEMETER retrospective forecasts. Int. J. Climatol. 30: 1577–1588. Rajeevan M, Bhate J, Kale J, Lal B. 2006. High-resolution daily gridded rainfall data for the Indian region: analysis of break and active monsoon spells. Curr. Sci. 91: 296–306. Randall D (ed). 2000. General Circulation Model Development: Past, Present and Future. Academic Press: New York, NY; 781. Rasmusson EM, Carpenter TH. 1983. The relationship between eastern equatorial Pacific sea surface temperatures and rainfall over India and Sri Lanka. Mon. Weather Rev. 110: 517–528. Roeckner E, et al.. 1996. The atmospheric general circulation model ECHAM4: model description and simulation of present-day climate. Max-Planck-Institutitute fur Meteorologie Rep. 218. Hamburg, Germany, 90. Saha S, Nadiga S, Thiaw S, Wang J, Wang W, Zhang Q, Van Den Dool HM, Pan HL, Moorthi S, Behringer D, Stokes D, Pena M, Lord S, White G, Ebisuzaki W, Peng P, Xie P. 2006. The NCEP climate forecast system. J. Clim. 19: 3483–3517. Sikka DR. 1980. Some aspects of large-scale fluctuations of summer monsoon rainfall over India in relation to fluctuations in the planetary and regional scale circulation parameters. Proc. Indian Acad. Sci. (Earth Planet. Sci.) 89: 179–195. Singh A, Kulkarni MA, Mohanty UC, Kar SC, Robertson AW, Mishra G. 2012. Prediction of Indian southwest monsoon rainfall using canonical correlation analysis of global circulation model products. Meteorol. Appl. 19(2): 179–188, DOI: 10.1002/met.1333. Sinha P, Mohanty UC, Kar SC, Dash SK, Robertson A, Tippett M. 2013. seasonal prediction of the Indian summer monsoon rainfall using canonical correlation analysis of the NCMRWF global model products. Int. J. Climatol. 33: 1601–1614, DOI: 10.1002/joc.3536. Smith TM, Reynolds RW, Thomas CP, Jay L. 2008. Improvements to NOAA’s Historical Merged Land-Ocean Surface Temperature Analysis (1880–2006). J. Clim. 21: 2283–2296. Stephenson DB, Doblas-Reyes FJ. 2000. Statistical methods for interpreting Monte Carlo ensemble forecasts. Tellus 52: 300–322. Taylor KE. 2001. Summarizing multiple aspects of model performance in a single diagram. J. Geophys. Res. 106: 7183–7192. Yadav RK, Rupa Kumar K, Rajeevan M. 2009. Increasing influence of ENSO and decreasing influence of AO/NAO in the recent decades over northwest India winter precipitation. J. Geophys. Res. 114: D12, 10.1029/2008JD011318, 1–12. Yadav RK, Yoo JH, Kucharski F, Abid MA. 2010. Why is ENSO influencing northwest India winter precipitation in recent decades? J. Clim. 23: 1979–1993. Yun W-T, Stefanova L, Krishnamurti TN. 2003. Improvement of the superensemble technique for seasonal forecasts. J. Clim. 16: 3834–3840. Zubair L, Ropelewski CF. 2006. The strengthening relationship between ENSO and northeast monsoon rainfall over Sri Lanka and southern India. J. Clim. 19: 1567–1575.

Int. J. Climatol. (2014)