Sensitivity of Indian summer monsoon rainfall forecast skill of CFSv2 model to initial conditions and the role of model biases
K Rajendran, Sajani Surendran, Stella Jes Varghese, Arindam Chakraborty
SSensitivity of Indian summer monsoon rainfall forecastskill of CFSv2 model to initial conditions and the role ofmodel biases
K Rajendran , , ∗ , Sajani Surendran , ,Stella Jes Varghese and Arindam Chakraborty Multi-Scale Modelling Programme (MSMP), CSIR Fourth Paradigm Institute (CSIR-4PI), Bangalore, India Academy of Scientific and Innovative Research (AcSIR), Ghaziabad, India Indian Institute of Science (IISc), Bangalore, India
January 20, 2021 ∗ Corresponding author address: Dr. K. Rajendran, Multi-Scale Modelling Programme (MSMP), CSIRFourth Paradigm Institute (CSIR-4PI), CSIR-NAL Belur, Wind Tunnel Road, Bangalore, 560 037. E-mail: [email protected] 1 a r X i v : . [ phy s i c s . a o - ph ] J a n bstract This study analyses Indian summer monsoon (ISM) seasonal reforecasts by CFSv2 model, initiated fromJanuary (4-month lead time, L4) through May (0-month lead time, L0) initial conditions (ICs), to exam-ine the cause for highest all-India ISM rainfall (ISMR) forecast skill with February (3-month lead time,L3) ICs. The reported highest forecast skill for L3 ICs is based on correlation between observed and pre-dicted interannual variation (IAV) of ISMR. Other skill scores such as mean error, bias, RMSE, mean,standard deviation and coefficient of variation, indicate higher or comparable skill for April (L1)/May(L0) ICs. Although theory suggests that skill degrades with increase in lead-time, CFSv2 shows highestskill with L3 ICs, due to predicting ISMR excess of 1983 for which other ICs fail. But, this correctforecasting is caused by wrong forecast of La Ni˜na, cooling of the equatorial central Pacific (NINO3.4)during the monsoon season, by L3 ICs. In observation, near-normal sea surface temperatures (SSTs) pre-vailed over NINO3.4 and ISMR excess was due to variation of convection over equatorial Indian Oceanor EQUINOO which CFSv2 failed to capture with all ICs. Major results are reaffirmed by analysingan optimum number of experimental reforecasts by current version of CFSv2 initiated from five late-April/early-May ICs having shorter yet useful forecast lead time. These experimental reforecasts arefound to have least seasonal biases and highest correlation skill score if 1983 is excluded. Model defi-ciencies such as over-sensitivity of ISMR to SST variation over NINO3.4 (ENSO) and unrealistic influ-ence of ENSO on the EQUINOO, contribute to errors in ISMR forecasting. Whereas, in observation,ISMR is influenced by both ENSO and EQUINOO. The forecast skill for Boreal summer ENSO is foundto be deficient in CFSv2 with the skill being the lowest for L3/L4 ICs, hinting the possible influence ofdynamical drift induced by long forecast lead-time. Rainfall occurrence despite strong cold bias overNINO3.4 in CFSv2, is associated with a stronger ocean-atmosphere coupling, with a shift of the SST-rainfall relationship pattern to slightly colder SSTs than the observed. These results warrant the need forminimisation of biases in SST boundary forcing to achieve improved ISMR forecasts.
Keywords:
Indian Summer Monsoon Rainfall, Seasonal Reforecasts, Forecast Skill, Model Biases,ENSO, EQUINOO 2
Introduction
Rainfall received over India during the summer season (June to September, JJAS) is termed as the In-dian summer monsoon rainfall (ISMR). There has been considerable year to year variation (known asinterannual variation or IAV) in the quantum of ISMR that has a profound effect on the agricultural sec-tor and the socioeconomic well-being of India. Hence, it is essential to predict ISMR or its departurein a season correctly to facilitate effective planning of agricultural and economic strategies, and waterand hydel power management. Despite the challenges in modelling Indian summer monsoon due to itscomplex features and multiple processes involved, coupled ocean-atmosphere general circulation mod-els (CGCMs) have become an essential tool for dynamical seasonal prediction. The climate forecastsystem version 2 (CFSv2) model of the National Centers for Environmental Prediction (NCEP), USA,is an outcome of such efforts in recent years to improve dynamical prediction and its forecast skill iswidely studied (Saha et al. 2010, Krishnamurthy and Rai 2011, Pattanaik et al. 2012, Saha et al. 2014etc.). Recently, this model is adopted by Ministry of Earth Sciences, Government of India, for dynamicalseasonal prediction of Indian summer monsoon.Over the tropics, the existence of slowly varying boundary conditions constitutes the basic premiseof seasonal prediction (Charney and Shukla 1981). Anomalous IAV of sea surface temperature (SST)over the equatorial central Pacific associated with El Ni˜no-Southern Oscillation (ENSO, Rasmusson andCarpenter 1983), is considered to be the primary source of predictability (Shukla and Wallace 1983).Krishnamurthy and Shukla (2011) examined the predictability of ISMR in eight CGCMs including CFSfor forecast and predictability errors and estimated the doubling time of errors for rainfall over India, tobe 4-14 days in the CFS against 4-7 days in other models. Forecast skill is to get better as the initialconditions (ICs) get closer to the prediction period and thus the highest forecast skill is expected for ICswith 0-month lead-time (L0). In other words, the skill is expected to increase (decrease) with decreasing(increasing) lead-time when considering the development of dynamical shift in model with time (Slingoand Palmer 2011) and systematic biases caused due to deficient representation of physical processes inthe model. Kumar et al. (2011) analysed CFS forecast skill of monthly mean SST and precipitation andshowed that the skill rapidly decays with lead time. After a lead-time of about 30-40 days, the forecast3kill for monthly mean was found to deteriorate, with the SST anomalies in the tropical central/easternPacific playing a dominant role. Thus, for seasonal predictability, the conditions of the ocean state alsobecome very important. They suggested the reduction in skill is due to the large contribution from theatmospheric internal variability to monthly means.Contrary to expectation and the understanding of significant ENSO spring predictability barrier andlow predictability of ENSO forecasts during February-March (Webster and Yang 1992), CFSv2 predic-tions of ISMR with February ICs (3-month lead time, L3), are reported to have the maximum forecastskill (Pokhrel et al. 2016, Ramu et al. 2016, Pillai et al. 2018, Rao et al. 2019), based on the correlationbetween observed and predicted IAV of ISMR during the analysis period. Further, the skill scores re-ported in previous studies vary considerably among themselves depending upon i) the region selected foraveraging the summer season rainfall to estimate ISMR for each year, ii) the reference dataset used as ob-servation and iii) the duration of the analysis period. These seasonal forecast verifications are performedwith datasets rarely exceeding 29 samples, which can also lead to highly uncertain scores (Schaeybroeckand Vannitsem 2019). However, an understanding of the impact of different ICs on ISMR forecast skillis fundamental and central to improving its predictability. Thus, it is imperative to understand whatcontributes to the forecast skill of February (L3) ICs. We focus on the factors which influence ISMRvariability in CFSv2 by comparing its seasonal reforecasts (hindcasts) with observations/reanalyses, withemphasis on its dependence on SST boundary forcing. We analysed large datasets of 124 CFSv2 refore-casts initialised with ICs from 1 st January to 31 st May, which are made available by NCEP. To reconfirmmajor results and to understand the advantage of choosing ICs which are nearer to the forecast period(JJAS) yet having useful lead-time, we assessed the performance skill of the current version of CFSv2by analysing its reforecasts initialised with an optimum subset of 5 late-April/early-May ICs.
CFSv2 is a coupled dynamical forecast system with Global Forecast System model in triangular trunca-tion of T126 ( ∼ ◦ horizontal resolution) with 64 hybrid sigma-pressure levels as the atmosphericcomponent and Geophysical Fluid Dynamics Laboratory Modular Ocean Model v.4 (GFDL MOM4)4ith 0.25 ◦ horizontal resolution in equatorial region ( ± ◦ latitude and 0.5 ◦ elsewhere) as the oceancomponent (Saha et al. 2014).To examine the dependence of ISMR forecast skill of CFSv2 on ICs, we have analysed 124 nine-month retrospective seasonal reforecasts or hindcasts (hereafter referred to as ’CFSv2-NCEP’) initiatedfrom CFS Reanalysis based ICs on every 5 th day starting from 1 st January (4-month forecast lead time,’L4’) to 31 st ◦ × ◦ gridded India Meteorological Department (IMD) rainfall (Pai et al.5014), Global Precipitation Climatology Project (GPCP) version 2.3 data (Adler et al. 2003) and HadleyCentre Ice and SST (HadISST) data (Rayner et al. 2003) are used. Daily optimum interpolation SSTversion 2.1 (OISSTv2.1 at 0.25 ◦ × ◦ horizontal resolution) data (Reynolds et al. 2007) is also anal-ysed. For validation of 850 hPa winds, we use 5 th generation European Centre for Medium RangeWeather Forecast Reanalysis (ERA5) data (Hersbach et al. 2020). NINO3.4 (170 ◦ -120 ◦ W; 5 ◦ S-5 ◦ N)SST anomaly (normalised by standard deviation) is used as ENSO index. ENSO index > <
1) indi-cates El Ni˜no (La Ni˜na). It is important to note that our analysis focuses on ENSO during the Indiansummer monsoon season (i.e. JJAS). Negative of the anomaly of surface zonal wind at the equatorial In-dian Ocean (IO, 60 ◦ -90 ◦ E; 2.5 ◦ S-2.5 ◦ N) estimated from ERA5 is used as the index for equatorial IndianOcean oscillation (EQUINOO, Gadgil et al. 2004).For estimating seasonal mean ISMR for each year, we use the rainfall averaged over the monsoonregion (Gadgil et al. 2019). Anomalies of ISMR, and indices of ENSO and EQUINOO are standardisedwith their standard deviation. For assessing the performance of forecasting the IAV of ISMR, basicskill scores such as ISMR temporal mean, standard deviation and coefficient of variation (CV) are used.Deterministic skill scores such as mean error, bias and root-mean-square error (RMSE) are also used.Details of these methods are provided in Appendix 1. To assess climatological mean monsoon rainfallpattern over India, statistics with respect to IMD rainfall viz. spatial pattern correlation coefficient (PCC),ratio of standard deviation against observed (SD) and climatological bias as percentage of observed(bias), are computed.
Figure 1a shows the interannual variation (IAV) of standardised ISMR anomalies from IMD observationand deterministic ensemble mean of CFSv2-NCEP L3 reforecasts. The correlations for deterministicreforecasts against the observed for 1982-2010 period ( γ ) for CFSv2-NCEP reforecasts with L0 to L4ICs are written in Fig. 1a. It can be seen that the ISMR forecast skill based on correlation is the highestfor CFSv2-NCEP L3 ( γ =0.44) followed by CFSv2-NCEP April (L1) ICs ( γ =0.35). We can see that theperformance of ensemble mean of CFSv2-CSIR reforecasts with late-April/early-May ICs (Fig. 1a) is6omparable with that of L3 ( γ =0.38).Corresponding IAV of Boreal summer ENSO index (NINO3.4 SST) anomalies is shown in Fig. 1b.ISMR excess of 1983 in L3 is associated with erroneous Boreal summer (JJAS) La Ni˜na forecast whenthe observed SST condition was neutral over NINO3.4 (Fig. 1b). The extreme ISMR departure in 1983is captured only by February ICs, in magnitude and sign. To some extent, the departure of 1994 is alsocaptured by CFSv2-NCEP L3. However, all ICs fail to forecast the departures of 1985, 1990, 1997,1998 and 2006. For these special years, departures of ISMR and ENSO index predicted by the ensem-ble means of CFSv2-NCEP reforecasts with L4 to L0 ICs are compared in Table 1. We can see that theISMR departures in CFSv2-NCEP are largely influenced by the sign and magnitude of their ENSO indexforecasts; ISMR deficits are associated with El Ni˜no or the anomalous warming of NINO3.4 SST andexcesses are associated with La Ni ˜na or the anomalous cooling of SSTs over NINO3.4 region. The in-verse relationship and interaction between ENSO and ISMR during boreal summer, are well-documented(Walker and Bliss 1932, Walker 1933, Sikka 1980, Torrence and Webster 1999 etc.). This relationshipis modulated on decadal timescales (Kumar et al. 1999, Chen et al. 2010, Kumar et al. 2006, Azadand Rajeevan 2016, Fan et al. 2017). Most importantly, in CFSv2-NCEP reforecasts, ISMR is shown tobe having over-sensitivity to ENSO, especially to the SST fluctuation over the equatorial central Pacificregion (Vishnu et al. 2019).Further examination of yearly ISMR departures, reveals that skill is not better in CFSv2-NCEP L3compared to CFSv2-CSIR (Fig. 1a), though their correlations are 0.44 and 0.38 respectively. The differ-ence between these two correlations is significant only at 0.73% confidence level. However, the previousstudies (Pokhrel et al. 2016, Ramu et al. 2016, Pillai et al. 2018, Rao et al. 2019) have all reported theskill improvement in CFSv2 with February ICs with such differences in correlations, though the exactcorrelation values vary from one study to another depending upon the region selected to compute averageseasonal ISMR, the data used as reference/observation and the duration of the analysis period.In observation, the ISMR departures of 1983, 1994, 1985, 1990, 1997, 1998 and 2006 are not strongly(and inversely) related to ENSO anomalies (Table 1). The excesses of 1983 and 1990 are associated withneutral ENSO phases and excesses of 1994 and 2006 are associated with mild warming and deficit of7985 is associated with strong cooling over NINO3.4. In spite of strong El Ni ˜no in 1997 and La Ni˜nain 1998, the ISMR remained close to normal in observation. In comparison, ISMR departures of 1985,1990, 1997, 1998 and 2006, are associated with very intense NINO3.4 SST anomalies in the modeland all ICs forecast the inverse ISMR departures. During these years the inverse relationship is strongand evident in the model and the dominant driving force determining the ISMR departure remains tobe ENSO with all ICs. Compared to observation, the model tends to show amplified ENSO anomalies(skewed for cold events), more for earlier ICs of L4-L2 and most conspicuously for L3. For 1983 and1994, larger errors are seen in ENSO predictions, with the largest error amplitudes for L3.It can be seen that the correlation for CFSv2-NCEP L3 falls to 0.4 which is lower than the corre-sponding score ( γ =0.42) for CFSv2-CSIR, if we exclude 1983 (Fig. 2). So the improved ISMR forecastskill of L3 is contributed by its prediction of 1983 ISMR excess. Next, we applied other deterministicverification scores such as the mean error, bias, and RMSE during the analysis period, for assessing theforecast skill of CFSv2-NCEP L3 and CFSv2-CSIR (Table 2). These skill scores are clearly improved inCFSv2-CSIR reforecasts compared to CFSv2-NCEP L3 (Table 2). The deficiencies of underestimationof the mean (dry bias) and standard deviation (reduced variability) of ISMR, also get improved and CVbecomes the closest to the observed in CFSv2-CSIR.Model intercomparison studies in the past had suggested that models which are skilful in representingclimatological mean summer monsoon rainfall are more adept in simulating IAV of ISMR (Sperber andPalmer 1996, Gadgil and Sajani 1998). CFSv2 is found to have reasonable skill in capturing the spatialdistribution of climatological JJAS mean rainfall, SST and 850 hPa winds over Indian region. The meanbias is lower in CFSv2-CSIR than in L3 (Fig. 3). Still, there exists underestimation of rainfall (drybias) over central India coinciding the seasonal monsoon trough zone, and uniform wet bias and wide-spread underestimation of SST (cold bias) over the Indian Ocean and West Pacific. Oceanic regionswith enhanced rainfall are associated with convergence and colder SSTs. Wet bias over equatorial IOhas zonal asymmetry with the maximum over the eastern equatorial Indian Ocean (EEIO) with strongwesterly wind biases and low-level convergence. Cold bias over the equatorial central Pacific is foundto be associated with the strengthening of ITCZ (wet bias) and with cyclonic wind bias slightly north of8quator (not shown).Dry bias over India is larger in CFSv2-NCEP L3 compared to CFSv2-CSIR. The pattern correlationcoefficient (PCC), standard deviation (SD) and mean bias are largely comparable among ensemble meansof CFSv2-NCEP and CFSv2-CSIR reforecasts (Table 3). But, the PCC is slightly larger for reforecastswith May ICs. Similarly, standard deviation and bias are clearly improved in CFSv2-CSIR reforecastswith late-April/early-May ICs and it is better than L3 in representing mean monsoon rainfall over theIndian region. This is expected as atmospheric and oceanic states are close to JJAS. The increase in biasas lead-time increases, indicates the role of dynamical drift in the model. The leading factor determining IAV of ISMR is the strong relationship between ISMR and ENSOin which there is an increased propensity of droughts during El Ni˜no and of excess rainfall during LaNi˜na (Sikka 1980). It can be gleaned from Fig. 1 that 8 out 12 excess events are associated with LaNi˜na and 8 out of 12 deficit events are associated with El Ni˜no in CFSv2-NCEP L3. There are no largeexcess (large deficit) associated with El Ni˜no (La Ni˜na). All large excesses (large deficits) are associatedwith La Ni ˜na (El Ni ˜no). Thus, ISMR-ENSO relationship is much stronger in CFSv2-NCEP L3 with acorrelation of -0.85 than in observation ( γ =-0.44) where other factors do influence ISMR (Fig. 4). Thestrongest correlation is seen for L4 followed by L3 and the correlation is the lowest for L0 ICs. ForCFSv2-CSIR reforecasts with late-April/early-May ICs, the correlation is -0.79 which is closer to theobservation compared to L3. It is to be recalled that its ISMR forecast skill is also comparable with L3for 1982-2010 period which becomes better ( γ =0.42) than that of L3 ( γ =0.40) when 1983 is excludedfrom the analysis period (Fig. 2). Thus, the correct forecast of 1983 ISMR excess as a result of anerroneous La Ni˜na forecast by L3 contributed to the seemingly higher IAV correlation for L3. But, otherskill scores do not show higher ISMR forecast skill for L3 (Table 2). Moreover, the Boreal summerENSO forecast skill is the lowest for L3 (Fig. 1b). This makes it necessary to analyse its ENSO forecastskill during Boreal summer, in detail. 9 oreal summer ENSO forecast skill CFSv2-NCEP L3 appears to have serious deficiency in forecasting summer-time ENSO (Fig. 1b).The forecast skill for JJAS ENSO index is found to be the lowest in CFSv2-NCEP L3 and L4 ( γ =0.59)compared to those in L2 to L0, and in CFSv2-CSIR ( γ =0.76). We have seen that the skill is much higherwhen all 124 reforecasts are considered together with the correlation of their median with the observedbeing 0.74 (not shown). The verification of performance of CFSv2 in predicting the warm and coldSST anomalies over the NINO3.4 region can be done from the classification of hits, misses and falsealarms in CFSv2-NCSP L3 and CFSv2-CSIR reforecasts (Table 4). The forecasts miss several eventsand there are a few false alarms as well. The number of misses and false alarms for cold and warmevents is more for L3 forecasts. Thus, the performance is slightly better for CFSv2-CSIR reforecastswith late-April/early-May ICs.Monthly forecast skill scores for ENSO indices against the observed during the analysis period (writ-ten in Figs. 5a-d) clearly manifest the bias in L3 ENSO forecasts for June, July, August and September.During 1983, L3 predicted neutral condition in June and thereafter strong La Ni˜na which kept intensi-fying from July to September. In contrast, in observation, NINO3.4 was having El Ni ˜no in June, neutralconditions in July and August, and a mild cold anomaly in September. The forecast skill for 1982-2010period, systematically drops from June to September with the least skill exhibited in September (Fig. 5).Correspondingly, the relationship of NINO3.4 SST with local NINO3.4 rainfall and remote impact onISMR in 1983, show model biases in L3. In observation, there is enhanced NINO3.4 rainfall associatedwith El Ni˜no in June which tends to become normal as SSTs approach climatology and then develops toa cold anomaly by September. Accordingly, ISMR varies from below normal in June to normal in July tolarge excesses in August and September. This is consistent with the inverse relationship between ISMRand ENSO. In L3, ENSO condition is near-neutral with excess rainfall over NINO3.4 in June which dropsto strong La Ni˜na in July which intensifies thereafter with deficit rainfall over NINO3.4. This results inabove-normal ISMR in June and large excesses in July to September of 1983.The strong association of local rainfall with NINO3.4 SST, even with cold bias over the equatorialPacific Ocean in CFSv2 (not shown), can be understood from the SST-rainfall relationship over NINO3.410rom June to September of 1982-2010 (Fig. 6). Figure 6 shows the number of points for each 0.25 ◦ CSST and 0.5 mm/day rainfall bin, along with the variation of mean rainfall with SST. The observedrelationship ( γ =0.59) shows that the rainfall along with the mean steadily increases with SST from about27 ◦ C with high propensity of rainfall for SSTs above this threshold. In CFSv2-NCEP L3, there is a slightshift in the SST-rainfall relationship towards colder SSTs (Fig. 6) with the number of observations above28 ◦ C becoming much lower than the observed. This is consistent with the finding that the SST-rainfallpattern in coupled models is similar to the corresponding observation or atmosphere-only version, exceptfor a shift of the pattern to colder/warmer SSTs as per their seasonal mean cold/warm bias (Rajendran etal. 2012).
In CFSv2, the SST-rainfall association/relationship over NINO3.4 is stronger with a correlation of 0.65than observed (Fig. 6), which in turn seems to have a remote impact on ISMR. Thus the ISMR predictiondepends highly on ENSO. This indirectly implies that the reduction of SST bias over central Pacific cancontribute to improvement in ISMR forecast skill. Further, daily SST averaged over the NINO3.4 regionshows that SST starts falling sharply after the beginning of monsoon season in 1983 (Fig. 7). Thedropping of SST is steep and large. The characteristics of the evolution of 1983 SST over NINO3.4for L3 ICs remain the same in the current version of CFSv2 as well (i.e., CFSv2-CSIR initiated withFebruary ICs). The build-up of bias hints at the role of dynamic drift and model bias resulting in colderSSTs by the summer months for L3. Given the high sensitivity of ISMR to NINO3.4 SST boundaryforcing, systematic approach to minimise SST bias is essential to achieve the potential predictability.
Another mode of SST variability in the equatorial Indian Ocean (IO), is the occurrence of oppositeSST anomalies over eastern equatorial IO (EEIO, 90 ◦ -110 ◦ E; 10 ◦ S-0 ◦ ) and western equatorial IO (WEIO,50 ◦ -70 ◦ E; 10 ◦ S-10 ◦ N), known as the Indian Ocean dipole (IOD, Saji et al. 1999). Climatologically,IO is warmer in the east supporting more convection than in the west during monsoon. Positive IODphase is characterized by weakening or reversal of climatological zonal SST gradient with suppression11enhancement) of convection over east (west) and anomalous winds blow from east to west along theequator, lifting up of thermocline and mixed layer of the east. However, the relationship between ISMRand IOD during JJAS is found to be rather weak, with the correlation coefficient not significantly differentfrom zero, and only about 1% of ISMR variance explained by IOD (Sajani et al. 2015).The atmospheric counterpart of IOD, the equatorial Indian Ocean Oscillation (EQUINOO), withits positive (negative) phase associated with enhanced convection over WEIO (EEIO) and suppressedconvection over EEIO (WEIO) is found to play an important role in determining IAV of ISMR (Gadgilet al. 2004) with the positive (negative) phase favourable (unfavourable) for ISMR. As the positive(negative) EQUINOO phase is associated with an easterly (westerly) anomaly of the zonal wind over thecentral equatorial IO, the EQUINOO index is based on the surface zonal wind anomaly over this region.Although EQUINOO is considered to be the atmospheric component of the coupled IOD mode, unlikeENSO, they are not as tightly coupled (Sajani et al. 2015) with correlation between their indices beingonly ∼ γ =0.14, Fig. 9a). Resultantly, ISMR-EQUINOO relationship is also too strong in CFSv2-NCEP L3( γ =-0.77, the largest among ICs, γ =-0.56 for CFSv2-CSIR) which is opposite to the observed ( γ =0.54)relationship (Fig. 9b).Forecast of 1994 ISMR departure by L3 was also due to an erroneous La Ni˜na forecast when inreality ISMR was excess only due to positive EQUINOO (coloured as green for positive EQUINOO andred for negative EQUINOO events in Table 1). In CFSv2, ISMR departure is almost entirely decidedby ENSO whereas in observation EQUINOO is found to play a decisive role in several years (Table1). It can be seen that excesses of 1983 (with neutral ENSO condition) and 1994 and 2006 (with mildwarm ENSO anomalies) are due to positive EQUINOO events. ISMR of 1985 was below normal dueto negative EQUINOO despite having very strong cold ENSO anomaly. Normal monsoons of 1997 and1998 are due to positive and negative EQUINOO events despite having very strong El Ni˜no and La Ni ˜narespectively. Inability of CFSv2 to forecast EQUINOO events independent of ENSO, made it impossiblefor the model to forecast ISMR anomalies of 1985, 1990, 1997, 1998 and 2006 with almost all ICs (Figs.1a and Table 1). This study attempts to understand what contributes to the highest ISMR forecast skill for CFSv2 Febru-ary (3-month forecast lead time, L3) ICs as reported in previous studies. We analysed 124 retrospective13ine-month reforecasts by CFSv2 with January (4-month forecast lead time, L4) through May (0-monthforecast lead time, L0) ICs, provided by NCEP for 1982-2010 period (referred to as CFSv2-NCEP re-forecasts). Our analysis reveals that the reported higher forecast skill for February (L3) ICs was based ona single skill score of correlation between observed and predicted ISMR departures during the analysisperiod. In contrast, other skill scores such as the mean error, interannual bias and RMSE, and the mean,standard deviation and coefficient of variation, indicate higher or comparable forecast skill for April/May(L1/L0) ICs. Climatological bias in mean summer monsoon rainfall over India is also found to be theleast with L1/L0 ICs. These results are reconfirmed through the analysis of a set of experimental refore-casts by the current version of CFSv2 with an optimum subset of 5 late-April/early-May ICs which arehaving shorter yet useful forecast lead times (referred to as CFSv2-CSIR reforecasts). Correspondingly,reforecasts with late-April/early-May ICs yield a correlation skill score comparable to that of L3 and thedeterministic ISMR forecast skill is found to be the best with late-April/early-May ICs for 1982-2010period, if 1983 is excluded.The success of CFSv2-NCEP L3 in forecasting a single event, i.e., excess ISMR departure in 1983,contribute to its higher IAV correlation of 0.44. The correlation is 0.38 for CFSv2-CSIR late-April/early-May ICs which is significantly different from that of CFSv2-NCEP L3 with only 73% confidence. Thesecorrelations become 0.40 and 0.42 for CFSv2-NCEP L3 and CFSv2-CSIR respectively, if 1983 is ex-cluded from the analysis period of 1982-2010. Further, we find that the success of CFSv2-NCEP L3in forecasting 1983 ISMR excess is due to its wrong forecast of La Ni ˜na (unlike L1 and L0 ICs) duringBoreal summer of 1983. Our analysis thus suggests the importance of initialising seasonal forecasts fromApril/May ICs.CFSv2’s common deficiencies such as the over-intensified influence of ENSO on ISMR and on vari-ation of SST, rainfall and circulation over the equatorial Indian Ocean, are also important factors whichcontribute to errors in ISMR forecasting. In CFSv2, ISMR is almost entirely decided by ENSO relatedSST boundary forcing, with no link between variabilities of ISMR and convection over equatorial IndianOcean associated with EQUINOO. In contrast, in observation, ISMR is influenced by both ENSO andEQUINOO independently. 14entral Pacific was under the sway of El Ni˜no till June 1983. All forecasts were initiated when ElNi˜no was prevailing with active convection over NINO3.4. CFSv2 is known to develop pronouncedwet and cold bias over the central Pacific. The fact that CFSv2-NCEP L3 with long lead-time endedin forecasting La Ni ˜na in summer hints at the possible role of wet bias and associated winds resultingin stronger cooling of NINO3.4 ocean surface for L3. This also implies that the persistence of errorsin atmospheric circulation due to imperfections in physical processes could eventually lead to large-scale bias in ocean circulation and surface temperatures. This can be manifested in larger magnitudes inforecasts with longer lead-times. Improvements in atmospheric model physics schemes and experimentswith observed SST forced atmosphere-only component of CFSv2 can throw further light on these aspects.It is also important to see if the ocean model of CFSv2 can simulate oceanic modes correctly when forcedwith realistic atmospheric circulation and fluxes. Our analysis suggests the need for a systematic approachto minimise the biases in SST boundary forcing in CFSv2, to achieve improved ISMR forecasts.
Acknowledgements ◦ × ◦ gridded rainfall (http://imdpune.gov.in/Clim Pred LRF New/Grided Data Download.html, Pai et al. 2014), Global Precipitation Climatology Project (GPCP) version2.3 data (https://psl.noaa.gov/ data/gridded/data.gpcp.html, Adler et al. 2003), Hadley Centre Ice andST (HadISST) data (http://badc. nerc.ac.uk/view/badc.nerc.ac.uk ATOM dataent hadisst, Rayner etal. 2003), 5 th eferences
1. Adler, R. F., Huffman, G. J. Chang, A. and co-authors, The version 2 Global Precipitation Clima-tology Project (GPCP) monthly precipitation analysis (1979-present). 2003,
J. Hydrmeteorol. , ,1147-1167.2. Azad, S., and Rajeevan, M., Possible shift in the ENSO-Indian monsoon rainfall relationship underfuture global warming. 2016, Sci. Rep. , , 20145.3. Charney, J. G. and Shukla, J., Predictability of monsoons. 1981, Monsoon dynamics , , 99-109.4. Chen, W., Dong, B., and Lu, R., Impact of the Atlantic Ocean on the multidecadal fluctuation of ElNi˜no-Southern Oscillation-South Asian monsoon relationship in a Coupled General CirculationModel. 2010, J. Geophys. Res. , , D17109.5. Fan, F., and co-authors, Revisiting the relationship between the south Asian summer monsoondrought and El Ni˜no warming pattern. 2017, Atmos. Sci. Lett. , , 175-182.6. Gadgil, Sulochana, and Sajani, Surendran, Monsoon precipitation in the AMIP runs. 1998, Clim.Dyn. , , 659-689.7. Gadgil, Sulochana., Vinayachandran, P. N., Francis, P. A., and Gadgil, S., Extremes of Indiansummer monsoon rainfall, ENSO, equatorial Indian Ocean oscillation. 2004, Geophys. Res. Lett. , , L12213.8. Gadgil, Sulochana., Rajendran, K., and Pai, D. S., A new rain-based index for the Indian summermonsoon rainfall. 2019, Mausam , , 485-500.9. Hersbach, H., Bell, B., Barrisford, P., Hirahara, S., and co-authors, The ERA5 global reanalysis.2020, Quart. J. Roy. Meteorol. Soc. ,
146 (730) , 1999-2049.10. Krishnamurthy, V. and Rai, S., Predictability of South Asian monsoon circulation in the NCEPClimate Forecast System. 2011,
Adv. Geosci. , , 65-76.171. Krishnamurthy, V. and Shukla, J., Predictability of the Indian monsoon in coupled general circu-lation models. 2011, COLA Tech. Rep. , .12. Kumar, A., Chen, M. and Wang, W., An analysis of prediction skill of monthly mean climatevariability. 2011, Clim. Dyn. , , 1119-1131.13. Kumar, K. K., Rajagopalan, B., and Cane, M. A., On the weakening relationship between theIndian monsoon and ENSO. 1999, Science , , 2156-2159.14. Kumar, K. K., Rajagopalan, B., Hoerling, M., Bates, G., and Cane, M. A., Unraveling the mysteryof Indian monsoon failure during El Ni˜no. 2006, Science , , 115-119.15. Pai, D. S., Latha, S., Rajeevan, M., Sreejith, O. P., Satbhai, N. S. and Mukhopadhyay, B., Develop-ment of a new high spatial resolution (0.25 ◦ × ◦ ) long period (1901-2010) daily gridded rainfalldata set over India and its comparison with existing data sets over the region. 2014, Mausam , ,1-18.16. Pattanaik, D. R., Mukhopadhyay, B. and Kumar, A., Monthly Forecast of Indian Southwest Mon-soon Rainfall Based on NCEP’s Coupled Forecast System. 2012, Atm. Clim. Sci. , , 479-491.17. Pillai, P. A., Rao, S. A. and co-authors, Seasonal prediction skill of Indian summer monsoonrainfall in NMME models and monsoon mission CFSv2. 2018, Int. J. Climate , , e847-e861.18. Pokhrel, S., Saha, S. K. and co-authors, Seasonal prediction of Indian summer monsoon rainfallin NCEP CFSv2: forecast and predictability error. 2016, Clim. Dyn. , , 2305-2326.19. Rajendran, K., Nanjundiah, R. S., Gadgil, Sulochana, and Srinivasan, J., How good are the simula-tions of tropical SST-rainfall relationship by IPCC AR4 atmospheric and coupled models?. 2012, J. Earth Syst. Sci. , , 595-610.20. Ramu, D. A. and co-authors, Indian summer monsoon rainfall simulation and prediction skill inthe CFSv2 coupled model: Impact of atmospheric horizontal resolution. 2016, J. Geophys. Res.Atmos. , , 2205-2221. 181. Rao, S. A. and co-authors, MONSOON MISSION: A targeted activity to improve monsoon pre-diction across scales. 2019, Bull. Amer. Meteorol. Soc. , , 2509-2532.22. Rasmusson, E. M. and Carpenter, T. H., The relationship between eastern equatorial Pacific seasurface temperatures and rainfall over India and Sri Lanka. 1983, Mon. Wea. Rev. , , 517-528.23. Rayner, N. A., Parker, D. A. and co-authors, Global analyses of SST, sea ice and night marineair temperature since the late nineteenth century. 2003, J. Geophys. Res. Atm. , , 4407,10.1029/2002JD002670.24. Reynolds, R. W., Smith, T. M., Liu, C., Chelton, D. B., Casey, K. S. and Schlax, M. G., Daily High-Resolution-Blended Analyses for Sea Surface Temperature. 2007, J. Climate , , 5473-5496.25. Saha, S., Moorthi, S., and co-authors, The NCEP Climate Forecast System reanalysis. Bull. Amer.Meteorol. Soc. , 2010, , 1015-1057.26. Saha, S., Moorthi, S., Wu, X. and co-authors, The NCEP Climate Forecast System version 2. 2014, J. Climate , , 2185-2208.27. Sajani, Surendran, Gadgil, Sulochana, Francis, P. A. and Rajeevan, M., Prediction of Indian rainfallduring the summer monsoon season on the basis of links with equatorial Pacific and Indian Oceanclimate indices. 2015, Environ. Res. Lett. , , 094004.28. Saji, N. H., Goswami, B. N., Vinayachandran, P. N., and Yamagata, T., A dipole in the tropicalIndian Ocean. 1999, Nature , , 360-363.29. Schaeybroeck, B. V. and Vannitsem, S., Postprocessing of Long-Range Forecasts. 2019, In: Sta-tistical postprocessing of ensemble forecasts , Chapter 10 , 267-290.30. Shukla, J. and Wallace, J. M., Numerical simulation of the atmospheric response to equatorialPacific sea surface temperature anomalies. 1983,
J. Atmos. Sci. , , 1613-1630.191. Sikka, D. R., Some aspects of the large-scale fluctuations of summer monsoon rainfall over Indiain relation to fluctuations in the planetary and regional scale circulation parameters. 1980, Proc.Indian Acad. Sci. (Earth Planet. Sci.) , , 179-195.32. Slingo, J. and Palmer, T. N., Uncertainty in weather and climate prediction. 2011, Philos. Trans.R. Soc. , , 4751-4767.33. Sperber, K. R., and Palmer, T. N., Interannual Tropical Rainfall Variability in General CirculationModel Simulations Associated with the Atmospheric Model Intercomparison Project. 1996, J.Climate , , 2727-2750.34. Torrence, C., and Webster, P. J., Interdecadal changes in the ENSO-monsoon system. 1999, J.Climate , , 2679-2690.35. Vishnu, S., Francis, P. A., Ramakrishna, S. S. V. S., and Schenoi, S. S. C., On the relationshipbetween the Indian summer monsoon rainfall and the EQUINOO in the CFSv2. 2019, Clim. Dyn. , , 1263-1281.36. Walker, G. T., and Bliss, E. W., World weather. V. 1932, Mem. Roy. Meteorol. Soc. , , 53-84.37. Walker, G. T., Seasonal weather and its prediction. 1933, Nature , , 805-808.38. Webster, P. J. and Yang, S., Monsoon and ENSO: selectively interactive systems. 1992, Quart. J.Roy. Meteorol. Soc. , , 877-926. 20 ppendix 1: Forecast Skill Scores The methods used for verifying forecasts are the following.1. Mean error is average error.
M ean Error = 1 N N (cid:88) i =1 ( F i − O i )
2. Bias is comparison of average forecast magnitude to the observed.
BIAS = N (cid:80) Ni =1 F i N (cid:80) Ni =1 O i
3. RMSE is average magnitude of forecast errors and Anomaly correlation is comparison of forecastanomalies to observed.
RM SE = (cid:118)(cid:117)(cid:117)(cid:116) N N (cid:88) i =1 ( F i − O i )
4. In addition, the amount of climatological JJAS rainfall over the Indian land region ( µ ), the cor-responding standard deviation of JJAS mean rainfall ( σ ) and its temporal coefficient of variation(CV) in percentage for 1982-2010, are estimated as: µ = (cid:80) Ni =1 India Rain ( JJAS ) i Nσ = (cid:115) (cid:80) Ni =1 (cid:0) India Rain ( JJAS ) i − µ (cid:1) NCV = σµ × td. Anomalies 1983 1994 1985 1990 1997 1998 2006IMD/HadISST January ICs
February ICs
March ICs
April ICs -0.06 0.13 0.93 -0.36 -1.89 2.03 -0.630.04 -0.23 -1.20 -0.01 1.81 -1.19 0.65
May ICs -1.11 -0.14 1.38 -0.75 -2.07 1.81 -0.850.59 0.10 -1.13 0.37 2.32 -2.05 0.92
Table 1: Standardised anomalies of ISMR and ENSO index for special years of 1983 and 1994, and1985, 1990, 1997, 1998 and 2006, for ensemble means of CFSv2-NCEP reforecasts with January (L0)to May (L4) initial conditions (ICs). For each entity, the first and the second sub-rows are the anomaliesof ISMR and ENSO index respectively. Cells are coloured in green (red) if anomaly of EQUINOO indexis positive (negative) for each year.
IAV CFSv2-NCEP CFSv2-CSIR ObservationSkill Scores Feb. ICs. Ens. Apr/May Ens. (IMD)
Mean Error (IAV) -3.11 -1.99Bias (IAV) 0.64 0.69RMSE (IAV) 3.14 2.12ISMR µ (mm/day) 3.50 4.51 6.50ISMR σ (mm/day) 0.51 0.66 0.76ISMR CV (%) 14.5 10.3 11.7 Table 2: Skill scores for interannual variation (IAV) of standardized (with standard deviation) anoma-lies of Indian summer monsoon rainfall averaged over monsoon region (ISMR), for ensemble means ofCFSv2-NCEP reforecasts with February (L3) initial conditions (ICs) and CFSv2-CSIR reforecasts withlate-April/early-May ICs, with respect to corresponding IMD observation for 1982-2010. Time seriesstatistics against the observed are mean error, bias, and RMSE are shown as skill scores. The mean ( µ ),standard deviation ( σ ) and coefficient of variation (CV in %) of ISMR for models and IMD observationare also given. 22 FSv2 reforecasts ICs PCC SD bias
Ensemble means ofCFSv2-NCEP reforecasts with L4-L0 ICs Jan (L4) 0.69 0.86 -33.2%Feb (L3) 0.69 0.85 -34.0%Mar (L2) 0.69 0.86 -33.2%Apr (L1) 0.69 0.90 -27.6%May (L0) 0.71 0.96 -19.6%CFSv2-CSIR reforecasts andensemble mean 21 Apr 0.69 0.92 -27.4%26 Apr 0.69 0.95 -24.3%01 May 0.70 0.93 -24.3%06 May 0.71 0.95 -21.6%11 May 0.70 0.95 -19.8%5 ICs Mean 0.70 0.94 -23.5%
Table 3: Statistics for climatological summer (JJAS) mean rainfall over India for ensemble means ofCFSv2-NCEP reforecasts with January (L0) to May (L4) initial conditions (ICs) and CFSv2-CSIR re-forecasts with late-April/early-May ICs and their ensemble mean with respect to corresponding IMDobservation for 1982-2010 period. Statistics are spatial pattern correlation coefficient (PCC), ratio ofmodel spatial standard deviation with the observed (SD) and bias in simulating the mean with respect toIMD observation. 23 old EventsHadISST CFSv2-NCEP CFSv2-CSIR
Warm EventsHadISST CFSv2-NCEP CFSv2-CSIR γ ) between observedISMR anomalies and those from the ensemble means of CFSv2-NCEP reforecasts with L4 to L0, andCFSv2-CSIR reforecasts with late-Apr/early-May ICs, are also given. 1983 ISMR anomalies are high-lighted in light red background color. b) Same as a) but for anomalies of ENSO index defined as stan-dardised anomalies of NINO3.4 SST. Correlation coefficients between observed HadISST based ENSOindex anomalies and those from the ensemble means of CFSv2-NCEP reforecasts with L4 to L0, andCFSv2-CSIR reforecasts with late-Apr/early-May ICs, are also given.25 FSv2 NCEP L3 CFSv2 CSIR Apr/MayInitial Conditions00.10.20.30.40.50.6 I S M R F o r e c a s t S k ill ( γ ) Figure 2: Correlations of standardised ISMR anomalies of deterministic (ensemble mean) CFSv2-NCEPreforecasts with February (L3) initial conditions (ICs) and CFSv2-CSIR reforecasts with late-April/early-May ICs, against the IMD based observation for i) 1982-2010 period and ii) 1982-2010 period excluding1983. 26igure 3: Climatological summer mean (JJAS) rainfall (shaded), SST (contours) and 850 hPa winds (vec-tors) from i) observation (top), and ensemble means of ii) CFSv2-CSIR reforecasts with late-April/early-May ICs (middle) and iii) CFSv2-NCEP reforecasts with L3 ICs (bottom).27igure 4: Anomalies of ISMR plotted against ENSO index for CFSv2-NCEP L3 (blue) and observation(red). Respective regression lines are drawn. Correlations ( γ ) for ensemble means of CFSv2-NCEP L4,L2, L1 and L0, and CFSv2-CSIR with late-Apr/early-May ICs, are given in bottom-left corner and forCFSv2-NCEP L3 and observation are given in top-right corner.28igure 5: Monthly standardised anomalies of NINO3.4 rainfall (NINO3.4-PR) plotted along with thecorresponding anomalies of ENSO index (NINO3.4-SST) and ISMR from observation and ensemblemean of CFSv2-NCEP L3 for a) June, b) July, c) August and d) September. Corresponding correlations( γ ) are also given. 29 R a i n f a ll ( mm / d a y )
11 2021 4041 6061 80>90 γ =0.59(a) Observation24 26 28 30SST ( ° C)03691215 R a i n f a ll ( mm / d a y )
11 2021 4041 6061 80>90 γ =0.65(b) CFSv2 NCEP Figure 6: Distribution of the number of points for each 0.25 ◦ C SST and 0.5 mm/day rainfall bins alongwith the mean rainfall versus mean SST for each bin (red curve) showing the relationship between rainfalland SST, for June, July, August and September of 1982-2010 period over NINO3.4 in e) observation andf) ensemble mean of CFSv2-NCEP L3. 30igure 7: Evolution of daily SST averaged over NINO3.4 from OISST observation, OISST daily clima-tology and ensemble mean CFSv2-NCEP reforecasts with February (L3) initial conditions (ICs), andensemble mean CFSv2-CSIR reforecasts (current version of CFSv2 being used in India) with February(L3) ICs. 31igure 8: Seasonal summer mean (JJAS) anomalies of rainfall (shaded), SST (contour) and 850 hPawinds (vectors) from i) observation (top), and ensemble means of ii) CFSv2-NCEP reforecasts with L3ICs (middle) and iii) CFSv2-CSIR reforecasts with 5 ICs (bottom).32igure 9: a) Anomalies of EQUINOO index plotted against ENSO index for CFSv2-NCEP L3 (blue) andobservation (red). b) Anomalies of EQUINOO index plotted against ISMR for CFSv2-NCEP L3 (blue)and observation (red). Respective regression lines are drawn and correlations ( γγ