Monsoon mission

AS the monsoon period approaches, there is understandably considerable anxiety about how it will fare this year after the devastating drought of 2009. To recall, the all-India Summer Monsoon Rainfall (ISMR) for 2009 was only 78 per cent of the long period average (LPA) a deficit of 22 per cent. The country was also unprepared for the drought because the operational forecast by the India Meteorological Department (IMD) had failed to predict it; before the onset of the monsoon, it had forecast a near-normal ISMR of 96 +/ 5 per cent of the LPA. Though its onset on May 23, 2009, was unusually early, the monsoon failed to advance for the next two weeks, resulting in a massive rainfall deficit of 54 per cent up to the last week of June. The IMD applied the customary mid-course correction on June 24 and scaled down the prediction to 93 +/ 4 per cent of the LPA. But the actual seasonal rainfall for June to September turned out to be way below the lower forecast limit.

As is known, for the long-range forecast (LRF) of the monsoon, since 1988 the IMD has been using an empirically determined statistical model based on an analysis of historical ISMR data and its correlation to a set of atmospheric and oceanographic variables (predictors) before the beginning of summer over different parts of the world. These statistical correlations are cast in a (regression) equation that is used to predict the ISMR on the basis of observed data of the predictors. After a string of 13 normal monsoons, following the major drought of 2002, that the statistical model failed to predict, the IMD modified the parameters being used for the 2003 monsoon forecast. While the prediction was 98 +/ 4 per cent of the LPA, the actual rainfall was 102 per cent, and the model was claimed to have been validated. However, there was highly deficient rainfall (87 per cent of the LPA) in 2004 though the new IMD model had predicted the ISMR to be 100 +/ 4 per cent of the LPA. Of course, many other statistical models (and even dynamical models) also failed to predict the deficit monsoons of 2002 and 2004. It is pertinent to note that the IMD model did not attract any criticism only because there were 13 successive normal monsoons. In quantitative terms, however, the model had failed to predict accurately 65 per cent of the time.

The failure in 2004 prompted the IMD to change its forecast strategy yet again. For the 2005 forecast, new statistical models with improved performance on the basis of new methods of predictor selection were claimed to have been developed and used. More advanced statistical techniques were stated to have been employed. Their better performance, as compared to the older ones based on pure climatology, were apparently proven by correct hindcasting of ISMR data of the past years, including 2002 and 2004. For 2005, the actual ISMR was 99 per cent of the LPA as against the model prediction of 98 +/ 4 per cent. The same model was used for the subsequent three years when the country once again had normal monsoons. For 2006, against a forecast of 92 +/ 4 per cent, the observed ISMR was 99 per cent of the LPA, somewhat higher than the upper limit of the model prediction; for 2007, against a prediction of 93 +/ 5 per cent, the actual was 105 per cent of the LPA; and for 2008, against 100 +/ 4 per cent, the actual was 98 per cent of the LPA.

The model, one may say, fared all right though in 2007 and 2008 it had significantly underestimated the ISMR. It must also be added that the model did not generally get the predictions right for the four homogeneous regions of the country northwest, central, southern peninsula and the northeast. It is not only important to predict the total rainfall (ISMR) but also the variation within a given season. But the model was also generally not able to predict correctly the rainfall during June, July and August when most of the seasonal rainfall is received. The year 2009 again was a bad one for the model. It remains to be seen whether the IMD tinkers with the model once again for the 2010 forecast.

The approach is clearly far from sound; it points to fundamental and inherent shortcomings of the statistical approach to the LRF of the monsoon. For nearly 125 years now, the IMD has been issuing the LRF on the basis of correlations of the atmospheric/oceanographic variables to the rainfall (regional or total). The first LRF, based on the relationship between the Himalayan snow cover and monsoon rainfall, was given in 1886. This was, however, subjective and qualitative, but in 1909 Sir Gilbert Walker introduced the technique based on correlation and statistical regression analysis. But, despite its initial success, the formula for the country as a whole began to fare badly, and it came to be recognised that the Indian region was too vast to be considered as one homogenous unit. Therefore, beginning 1924, the IMD started issuing the LRF for two homogenous rainfall regions: northwest India and the peninsula. This practice was followed up to 1987. But in 1988, following the development of a 16-parameter power regression model, the IMD returned to the practice of issuing forecasts for the country as a whole. Beginning 1999, it began to issue forecasts for the homogeneous regions also.

Comparing the performances of the LRF models during 1932-1987 and the power regression models-based LRF during 1988-2004, Sulochana Gadgil and Ravi Nanjundiah of the Indian Institute of Science (IISc), Bangalore, and M. Rajeevan of the Indian Space Research Organisations National Atmospheric Research Laboratory (NARL) at Gadanki, Andhra Pradesh (he was formerly of the IMD, Pune), concluded in 2005 that the models based on the relationship of monsoon rainfall to atmospheric/oceanic conditions had not been satisfactory. They showed in Current Science that there had not been any improvement in the magnitude of forecasting error over the past eight decades in spite of continual evaluation and improvement of the operational statistical models. They argued that while new approaches that incorporate inherent non-linearity in the relationships as well as new predictors based on the current understanding of the role of the Indian Ocean could be explored, the statistical approach pursued by the IMD has been far from successful. Moreover, correlations have been found to vary, vanish or even become negative over time (typically decadal scales), or new correlations appear. These may well be artefacts of the models but do lead to periodic modifications to them.

As against the empirical approach that underlies statistical models, the approach based on dynamical models is a deterministic one. In these, equations governing the appropriate (planetary scale or regional) physical processes of the atmosphere or the ocean-atmosphere coupled system are set up and numerically integrated over time to determine the systems evolution from an initial state (given by observed data of a set of variables) before the season. The process is numerically intensive and requires high performance computing (HPC) resources, the required computing power depending upon the fidelity or resolution with which the initial data can be specified and the corresponding grid points over which integrations have to be performed. The approach is also called numerical weather prediction (NWP).

Obviously, the success of the approach depends on how well the physical processes that play a role are understood, what (near and distant) elements of the atmosphere/ocean system need to be included and the scale of grid-resolution for numerical integration. With improved data-gathering methods, better understanding of the physics involved and the ever-increasing available computational power, the NWP has been getting better and better over the years. But the monsoon system has turned out to be a particularly complex one for meteorologists to understand the totality of underlying processes, and success with dynamical models both Atmospheric General Circulation Models (AGCMs) and atmosphere-ocean Coupled General Circulation Models (CGCMs) has been limited.

A major advance in seasonal forecasting over tropical regions has been due to the enormous progress in understanding El Nino, the phenomenon of waters warming in the eastern and central Pacific, resulting in increased atmospheric convection. El Nino has been found to play a crucial role in determining the summer weather over the tropics. Understanding it has led to the development of dynamic models that simulate the phenomenon fairly well and thereby the potential for improved realistic prediction of tropical weather, in particular the Indian monsoon. It is now well known that in an El Nino year there is a strong likelihood of deficient monsoon over India, while in a La Nina (opposite of El Nino) year, there is excess monsoon rainfall. Hence, the hope was that the improved ability to predict El Nino would result in improved forecasts of the Indian monsoon.

However, there have been significant examples that indicate that this correlation may not be absolute. There could be other significant offsetting factors as well. The complex nature of the influence of the Pacific on the monsoon became clearer in the 1990s when warm episodes of 1992, 1993 and 1994 did not lead to deficient monsoons. And, to top it, the strongest El Nino of the last century, in 1997, resulted in above-normal rainfall. Then came the severe drought of 2002 when there was a much weaker El Nino. This indicated yet again that the relationship of El Nino and the monsoon is yet to be fully understood and that there is much more to the inter-annual variability than can be attributed only to teleconnections to Pacific anomalies.

Way back in 1998, Sulochana Gadgil and S. Sajani of the IISc analysed the simulation of monsoon rainfall for the years 1979-95 by 20 state-of-the-art AGCMs that formed part of the international Atmospheric Model Intercomparison Project (AMIP). For AGCMs it is necessary to prescribe the Sea Surface Temperature (SST) as a boundary condition for the period of prediction. The AMIP simulations were thus made with the SST specified from observations and thus were expected to have better skill than forecasts made with predicted SSTs. But they found that very few models were able to capture all the fluctuations between good and poor monsoons observed during the period of AMIP simulation runs even as they could capture the Pacific anomalies fairly well.

Comparison of observed and simulated anomalies in the ISMR for the drought years (1979, 1982, 1987) and excess rainfall years (1983, 1988, 1994) showed that all but one AGCM could simulate the correct sign of the anomaly for the excess monsoon of 1988 and only one other model could do it for 1994. Similarly, all but three models simulated the deficit properly (with several of them showing a large deficit) for the drought of 1987. Most of the models simulated excess rainfall against the observed drought of 1979, with only one simulating a large deficit. While the drought years of 1982 and 1987 were associated with El Nino, the good monsoon of 1988 was associated with La Nina. Interestingly, analysis of non-El Nino seasons with significant rainfall anomalies showed that the number of models that could simulate these was much smaller than for the seasons associated with El Nino. Even though AGCMs have evolved to capture the impact of El Nino on the Indian monsoon, they are yet to reach a stage where they can fully simulate the year-to-year variation of the ISMR, particularly the extremes (droughts and excess rainfall), argued Sulochana Gadgil and her team in Current Science when they analysed the failure of these dynamical models to predict the deficient rainfalls of 2002 and 2004. CGCMs also fare no better in simulating the inter-annual variability. Analysis of simulations of the ISMR for the period 1959-2001 by the CGCM of the United Kingdoms Met Office (UKMO) showed that the correlation between simulated and observed rainfall was poor. Of the eight drought years in the period, the model could simulate anomalies with the correct sign only in four cases. Further, the magnitude of the anomaly was reasonable only for 1965 and 1987. For the major drought of 1972, the model simulated a slightly above normal rainfall, and for the famous El Nino year of 1997, the model gave a large deficit against the observed above normal rainfall. The model also failed to capture the years of excess rainfall. It failed to simulate the monsoon of 1961, the strongest in 150 years, and likewise those of 1983 and 1988 as well for which it actually gave a negative anomaly. For 1994, it correctly gave a positive anomaly, but the error was large. For the period beyond, the model consistently gave anomalies opposite to those observed. Thus at the present juncture, wrote Sulochana Gadgil and associates, neither the AGCMs nor the CGCMs are able to simulate correctly the inter-annual variation of the ISMR.

Most of the models in use and investigated for their performance over the Indian region have been developed by major international centres such as the UKMOs Hadley Centre, the National Centres for Environmental Prediction (NCEP) of the National Oceanographic and Atmospheric Administration (NOAA), the International Research Institute for Climate Prediction (IRI), the Centre for Ocean-Land-Atmosphere Studies (COLA) and the European Centre for Medium Range Weather Forecasting (ECMWF). Some of the models have been made available to Indian institutions where researchers have tweaked them suitably to make them represent the Indian region better and have run experimental forecasts of the monsoon. These include the IMD, the Indian Institute of Tropical Meteorology (IITM), the National Centre for Medium Range Forecasting (NCMRWF), the IISc, IIT Delhi, the Space Application Centre (SAC), the Centre for Development of Advanced Computing (CDAC), the National Aerospace Laboratories (NAL) and the Centre for Mathematical Modelling and Computer Simulation (CMMACS). After the deficit years of 2002 and 2004, which none of the overseas centres predicted, it was also found that none of the Indian groups forecasted these with the models that they had adapted.

An editorial in Current Science in July 2005 said: Since none of the atmospheric models were able to predict the recent droughts, an objective assessment of the performance of all the models used in the country for generating predictions of the fluctuations of the monsoon and particularly the droughts is a must. The performance and reliability of the models can be compared by running all the models for at least the last 20 years with identical initial and boundary conditions from data available before each monsoon season, in what is called the hindcast mode.[I]t would lead to a more focussed research effort in developing better models for monsoon prediction. We believe that organising such an intercomparison should be given a very high priority by the DST [Department of Science and Technology].

A project called Seasonal Prediction of the Indian Monsoon (SPIM) was indeed carried out last year under the aegis of the DST by a team of scientists from key institutions and led by Sulochana Gadgil and J. Srinivasan of the IISc and S. Purohit of CDAC (see Table 1 & 2 and figure). This project evaluated five AGCMs being used in the country in the prediction of ISMR by simulation runs of these for the period 1985-2004 in CDACs supercomputer Param Padma. The exercise also attempted to understand why errors occurred. Two sets of runs were carried out: one, with observed SSTs as the boundary condition and, two, for specific anomalous years (1987, 1988, 1994, 1997 and 2002) forced by SST by assuming that the April anomalies persist throughout the monsoon season.

The results showed that none of the models was able to simulate the correct sign for the ISMR anomaly for every year. Amongst the five models, the Portable Unified Model (PUM) of Hadley Centre was found to simulate the correct sign for the maximum number of years and the Seasonal Forecast Model (SFM) of the NCEP had the maximum skill in simulating extreme events. The findings were found to be consistent with the AMIP runs. It should be possible, the SPIM report concluded, to undertake an even more ambitious project that could involve simulations of the entire 20th century. The time may also be opportune to embark on a similar exercise with CGCMs to unravel the facets of monsoon variability which depend upon coupling and also address critical climate change issues.

It may be argued that the problem in correctly simulating the monsoon behaviour is partly because of the coarse resolution used in most GCMs. Inclusion of high-resolution regional models in conjunction with GCMs is another strategy often suggested. But a recent study of simulations with a high-resolution ECMWF model is not very promising. Though there has not been any systematic assessment of the performance of regional models, the one used at IIT Delhi, based on a French model, has had only limited success. Moreover, some experts believe that regional models are unlikely to do better than the GCMs they are used in conjunction with, because regional information gets swept out of the domain very soon and only global information survives during integration over longer periods. There are also arguments favouring forecasts based on an ensemble of GCMs rather than a single model.

That such approaches have not really proved successful is also borne out by a recent assessment by Nanjundiah of predictions made by major forecasting centres in April-May 2009 for the June-August 2009 rainfall using state-of-the-art AGCMs and CGCMs, including ensembles. It was a year of severe drought, with large deficits in rainfall in June, August and September. He found that all the CGCMs were able to predict the warming of the central Pacific associated with the El Nino. Also, all the AGCMs and CGCMs predicted the rainfall over the Pacific reasonably well. However, none of the models, except that of the Experimental Climate Prediction Centre (ECPC) of the Scripps Institute of Oceanography, United States, predicted a deficit, let alone a drought, over the Indian region for June-August 2009. In its April 2009 forecast, the ECPC predicted a deficit, and in its updated May forecast it, predicted a more severe deficit over a large part of India. We need to understand, said Nanjundiah, why this model succeeded and all the others failed to anticipate the drought. It is clear that more research is necessary.

Rajeevan and Nanjundiah have also recently looked at the skill of 10 CGCMs that were used by the Intergovernmental Panel for Climate Change (IPCC) in its Fourth Assessment report (AR4) of 2007 in simulating the 20th-century mean climate of the Indian monsoon. Their analysis has shown that there are large biases in the mean ISMR simulated by these models as compared to the observations. The models simulate excess rainfall over the equatorial Indian Ocean, while over the monsoon trough zone they simulate less rainfall.

The models, says their paper published in the Platinum Jubilee volume of the Indian Academy of Sciences, showed serious problems in simulating the northward migration of the rainfall belt from the equatorial Indian Ocean into the Indian landmass as insolation across the tropics varies during summer. They also have a cold bias in simulating SSTs over the Indian Ocean and the Pacific. The strength of monsoon flow into the Indian region is also underestimated. So even from a climate change perspective for the Indian region, a coordinated national effort at perfecting dynamical models is important.

Taking off from the recommendations made in the SPIM report, the Union Ministry of Earth Sciences (MoES) has recently proposed a National Mission on Monsoon that would develop reliable dynamical models for forecasting the monsoon over the next three- to five-year period. The objective is to put in place a coordinated multi-institutional effort. Though the focus will chiefly be on seasonal forecast, the mission will include aspects of short-range (up to 3 days) and medium-range (up to a week) predictions as well. However, improving and perfecting the IMDs statistical models will not form part of the mission, clarified Ajit Tyagi, Director-General of the IMD.

There are enough models and enough people around. We are almost a critical mass [of atmospheric scientists] now. It is the right time to put in a concerted effort to improve monsoon prediction. Nobody else is going to do it for us, said Sulochana Gadgil. According to her, the initial focus of the mission will be atmospheric models. Coupled models are not yet good enough. They get wrong SSTs and predict more rain over cold oceans, she added. There has also been no progress in the predictability of inter-seasonal and intra-seasonal variability in most models. They are not able to anticipate extremes of droughts and excess, she pointed out.

If we look at the [prediction] skill of most numerical models, we find that it is one of the lowest over the Indian region, pointed out Nanjundiah. This certainly implies that a lot needs to be done in improving these models. This would involve a lot of experimentation with models, identify processes that are going wrong, and try and fix them. This is easier said than done. Since considerable amount of signal for the seasonal scale comes from the oceans, we also need to worry about getting the oceanic state correctly. We would also need to think about ocean state data assimilation so that the initial condition for CGCMs can be improved. Monsoon mission is a step in the right direction. But dont expect miracles overnight. This is going to be a long and hard grind. Today, the major problem we could be facing in tackling the problem is not one of financial or computational resources; it is one of human resource, he added, striking a realistic note.

The mission is yet to be formalised as it has to be approved by the Planning Commission and then the Cabinet. According to Shailesh Nayak, Secretary, MoES, the details are being worked out. The project proposal is being prepared. In January, we had a meeting of people involved in model development and who have a working knowledge of models. Mostly these are models developed elsewhere. All that activity is slowly getting into frame. But these atmospheric or ocean-atmosphere coupled models need to be improved greatly through a coordinated effort. We have to create a forum for that, he added.

I feel, it is a good initiative, said Rajeevan. But the mission should be monitored and reviewed properly. Otherwise, it can end up like an ordinary project work.