1 Introduction

The variations of the seasonal rainfall associated with the south Asian monsoon are enormously important for millions of lives on the Indian subcontinent and beyond. The spatial and temporal variations of rainfall have a significant impact on the agrarian economies of India, Bangladesh and Pakistan. While interannual variations in Indian summer monsoon rainfall (ISMR) are only \(\approx\) 10% of the long term mean, the high and low extremes of the seasonal mean ISMR result in floods and droughts (Shukla and Moolay 1987). Food production in the Indian region is strongly correlated with ISMR (Gadgil et al. 1999), and these floods and droughts can cause devastating human and economic losses. The south Asian monsoon is recognized as a prominent feature of the global circulation (Lau and K.-M. Kim 2006). Continental-scale land-sea contrast has been suggested as primary cause for the monsoon (Webster et al. 1998), while other studies suggest it is driven by the meridional movement of the Intra-Tropical Convergence Zone (ITCZ) (Gadgil et al. 2003). Besides these two basic components the ISMR is also influenced by the topography of Great Himalaya, which introduces an elevated heating source and helps to set the meridional tropospheric temperature gradient. The local reversal of the meridional tropospheric temperature gradient during the summer is thought to be important for the onset of the ISMR. This gradient is maintained in part by the heat fluxes and diabatic heating due to precipitation (Yanai et al. 1992; Wu and Zhang 1998). The topography of Himalaya isolates the Indian monsoon thermal maximum from the dry and cold air in the interior of Asian continent (Chakraborty et al. 2002; Boos and Kuang 2010), and numerical modeling studies have found that by removing the topography the northern extent of the precipitation is greatly reduced (e.g., Hahn and Manabe 1975; Prell and Kutzbach 1992). Another key feature of the monsoon circulation is the climatological low over northwestern India and Pakistan, which is the deepest low in the global tropics during boreal summer (Joshi and Desai 1985; Sikka 1997). It develops in April–May concurrently with the south-westerly wind regime (Ramage 1996). The high winds associated with the monsoon trough not only bring moisture over the land but also natural dust and aerosols. Aerosols can influence the monsoon through direct (interaction with solar radiation) and indirect (interaction with cloud microphysics) effects (Bollasina et al. 2011; Lau and K.-M. Kim 2006). Slowly varying boundary conditions such as SST, snow cover and soil moisture are also key components of the Indian monsoon, particularly in terms of its potential predictability (Charney and Shukla 1981). The teleconnection between southern oscillation and ISMR is among the oldest observed teleconnections (Walker 1925). Observational analysis shows that indian summer monsoon rainfall found below average during El Niño events, while La Niña events lead to above normal rainfall (e.g. Sikka 1980; Pant and Parthasarathy 1981; Rasmusson and Carpenter 1983; Gadgil et al. 2003, 2004). Niño 3.4 index (standardized area average SST average over the region 170\(^{\circ }\)E–120\(^{\circ }\)W, 5\(^{\circ }\)S–5\(^{\circ }\)N) is negatively correlated with ISMR. The observed negative correlation between the ISMR and Niño 3.4 index can be explained to some extent by the modulation of the Walker circulation (Shukla and Paolino 1983; Palmer et al. 1992). Thus, the Indian monsoon includes a complex orographically influenced structure, interaction between convection and large-scale atmospheric circulation, wave propagation in both the zonal and meridional directions, air-sea interaction, and cloud-aerosol interaction. Due to the presence of all the above components and their nonlinear interactions, Indian monsoon rainfall is an extremely challenging phenomenon to simulate (Gadgil et al. 2005).

Uncertainties and model errors in climate prediction can be classified into two groups: (1) uncertainties and errors in model initialization and (2) uncertainties and errors in model parameterizations and model physics (Buizza et al. 2005; Schwierz et al. 2006). The multi-model ensemble (MME) is recognized as one approach to address the above-mentioned uncertainties and errors (Palmer et al. 2004, 2005; Hagedorn et al. 2004). MMEs typically have higher skill for predicting weather and climate as compared to single models, and also provide estimates of model uncertainty. The simulation and prediction of ISMR at both inter-annual and intra-seasonal time scales has been evaluated in several such MMEs (Gadgil and Sajani 1998; Kang et al. 2002; Rajeevan and Nanjundiah 2009; Sperber et al. 2013; Wang et al. 2004, 2004). All MMEs examined previously have been shown to simulate large-scale feature of Indian rainfall with modest skill. Some studies (Wang et al. 2003; Sharmila et al. 2013) have highlighted the importance of air-sea interactions and suggest that coupled ocean-atmospheric models are crucial for monsoon seasonal predictions. Preethi et al. (2010) and Rajeevan et al. (2012) evaluated the seasonal forecast skills of Development of European multi-model ensemble system for seasonal to interannual predictions (DEMETER) (Palmer et al. 2004) and ENSEMBLE (Hewitt and Griggs 2004) projects respectively and found that these multi-model ensembles predict ISMR with positive (modest) skill. The realized skill is still below the limit of potential predictability (Saha et al. 2016).

In this study we investigate the ability of the North-American Multi Model Ensemble (NMME) models to reproduce and predict the seasonal mean and interannual variability of the Indian summer monsoon rainfall. The NMME is a collaborative effort between several modeling centers for seasonal forecasts. The NMME simulations provides us with the opportunitiy to compare the simulations from multiple seasonal models for the same phenomenon. The analysis of the multi-model simulations for identical scenarios will aid us in identifying and understanding the similarities and differences of the various model simulations. The study of Kirtman et al. (2014) have shown that modeling system improvements and data assimilation system improvements led to improved NMME-2 forecast quality. The second objective of this study is to compare the seasonal forecast skill of NMME phase 1 with the currently operational NMME phase 2 to understand whether the improvements in modeling systems and data assimilation systems have contributed to improved seasonal prediction of the Indian summer monsoon.

2 Data and methodology

The NMME is an MME producing both retrospective and real-time intraseasonal to interannual predictions and is comprised of global coupled atmosphere-ocean models from modeling centers in the United States and Canada (Kirtman et al. 2014). The NMME provides retrospective seasonal forecasts for 1982–2010. In this study nine models are selected from the first implementation of the NMME (phase 1; denoted here as NMME:1) and nine models from the current implementation (phase 2; denoted here as NMME:2 as summarized in Table 1. CFSv2, CanCM3 and CanCM4 are the common models in both of the NMME phases (denoted by \(\oplus\) in Table 1). The 15 models have a common re-forecast period of 28 years from 1982–2009. The number of ensemble members for each model ranges from 6 to 24, with 109 total ensemble members from nine models for NMME:1, and 110 ensemble members from nine participating models for NMME:2. Model runs are initialized every month with forecast lengths ranging from 6 to 11 months. In the present study we analyze the June–September (JJAS) seasonal means of precipitation and SST for forecasts starting from May 1 initial conditions. Equal weights are given to each model in calculating the average over all models and ensemble members, denoted the multimodel ensemble mean (MMEM). The choice of reforecasts initialized in May was made in order to avoid inclusion of potential skill from the atmospheric initial conditions. It is assumed that after one month of model integration, the atmospheric initial conditions, which provide much of the skill for numerical weather forecasts at 1–15 days lead-time, have a minimal impact on the forecast skill of the ensuing seasonal mean. It is possible that forecasts initialized in May are subject to the spring predictability barrier (Torrence and Webster 1998), which may mask some of the difference in skill among models. All NMME models are re-gridded to a common 1\(^{\circ } \times 1^{\circ }\) resolution. The Climate Prediction Center Merged Analysis of Precipitation (CMAP) (** intraseasonal prediction. Bull Am Meteorol Soc 95:585–601" href="/article/10.1007/s00382-018-4203-6#ref-CR26" id="ref-link-section-d139343495e3001">2014) indicated that improvement in data assimilation and modeling systems contributed to improved forecast quality in NMME phase 2. However, we find the skill of seasonal prediction of Indian summer monsoon rainfall is nearly the same in NMME:2 (0.46) as compared to NMME:1 (0.40); the NMME is still not able to accurately predict extremes (drought/floods) of rainfall. Therefore seasonal monsoon rainfall forecast is not improved by the improvement in data assimilation system and modeling system in the NMME phase 2. The inability to predict extremes can also be seen in both the DEMETER and ENSEMBLE experiments (Preethi et al. 2010; Rajeevan et al. 2012). Both DEMETER and ENSEMBLE, as well as NMME predicted droughts during the normal monsoon years of 1997 and normal monsoon year during flood year of 1983. This suggests that similar biases found in the DEMETER and ENSEMBLE models exist in the models used in the NMME.

The interannual and intraseasonal time scale variability of ISMR is strongly influenced by SST variability in the Pacific and Indian Oceans. Pointwise correlation of seasonal mean SST from NMME and observations revealed that the skill of interannual predictions is high (0.6–0.9) for most ocean basins, and improved in NMME:2 relative to NMME:1. The most common seasonal mean SST biases in NMME models are cold equatorial Pacific and subtropical Atlantic Ocean and warm biases in northern Pacific Ocean. These biases also remain in the MMEMs, and while the cold bias over the equatorial Pacific is improved in NMME:2, the re-forecasts of the Indian Ocean warm bias worsen. We find that the NMME simulates the observed interannual variability of the NINO3.4 index with correlations greater than 0.8. We also find that predictions of the ENSO anomalies are remain same in both NMME:1 NMME:2.

In this work we also examine teleconnection patterns that affect the monsoon, and find that teleconnections in the MMEMs are stronger than in the observations. The MMEMs capture the ENSO-monsoon, Atlantic-monsoon and west Pacific-monsoon teleconnections correctly, but fail to correctly represent the association with the Indian Ocean. The EQUINO-ISMR relationship in particular is opposite to what is observed. The teleconnection between the ISMR and Indian Ocean SST also was not represented well in the DEMETER and ENSEMBLES models. This again suggests a common systematic error in coupled model forecasts. This error in association may be the reason why the NMME predicted droughts during the normal monsoon years of 1997 and a normal monsoon year during the flood year of 1983, as SST anomalies in the Indian Ocean during 1997 and 1983 played an important role in overcoming the negative impact of El Niño events (Gadgil et al. 2007). The NMME captures the negative correlation between ENSO and the monsoon, but the influence of ENSO on ISMR is stronger in the NMME than is observed. The overly strong ENSO-ISMR relationship suggests that oceanic influence on atmosphere may be too strong in NMME, particularly when comparing the MMEM to observations.

Overall the NNME shows modest skill in predicting Indian summer monsoon rainfall and its interannual variability. However, the NMME models show common biases in rainfall over Indian Ocean, are unable to predict the extremes in seasonal rainfall, and show only modest increases in skill from NMME:1 to NMME:2. The failure to represent the monsoon-EQUINO teleconnection in particular may be a critical limitation of the models comprising the NMME, and the association between this link and the prediction of extremes of seasonal rainfall clearly warrants further investigation.