Forecast failures

Statistical models forecasting total rainfall from the monsoon in a given year need to be recast with suitable predictors if the present margins of error are to be reduced.

THE year's total rainfall from the southwest monsoon - from June 1 to September 30 - in the country as a whole was 92 per cent of the long-term average (LTA) of 88 cm. This means it was a normal monsoon - defined as +10 per cent of the LTA - for the 13th year in succession. But it also marks the failure, in quantitative terms, of the statistical long range forecast (LRF) model of the India Meteorological Department (IMD) for the sixth year running. The 16-parameter model, called the Power Regression Model, which uses 16 empirically derived atmospheric variables as predictors in a statistical relationship with the total rainfall, had forecast a total monsoon rainfall of 98 per cent of the LTA. The model has an inherent error margin of 4 per cent.

Indeed, since the IMD started issuing this 'operational forecast' based on the model in 1988, the actual rainfall has been beyond its error window as many as nine times - that is, 70 per cent - and the prediction for 1994 was particularly off. Clearly, that would suggest that the model needs a review. But because the monsoon has been normal all these years and the predictions have been right in the gross, the model continues to be the favoured one and influences government policies, planners and the stock market. Its actual impact on the agriculture economy, and the economy of the country in general, is minimal because in any case it cannot predict the temporal and spatial distribution of rainfall and, more important, the weather extremes of droughts and floods.

"The most important factor to remember is that if any model predicts that the monsoon will be normal or above normal, it will be right 90 per cent of the time irrespective of its intrinsic merits. Hence I do not think it is worthwhile to examine them too closely until they predict all the extreme events such as widespread droughts and floods correctly," says J. Srinivasan, Professor of Climate Studies at the Indian Institute of Science (IISc), Bangalore. "The true challenge in monsoon forecasting is to predict the extreme events. These are so rare that normal statistical methods cannot be applied," he adds.

It must, however, be said that notwithstanding the inherent methodological flaws in the model (Frontline, July 8, 2000), the IMD has begun forecasting rainfall for three homogeneous - in terms of inter-annual variations in rainfall - regions of the country since 1999. These are northwestern India, peninsular India and northeastern India. The model uses three different subsets of the 16 predictors with 6, 10 and 7 parameters respectively, and the forecast is said to have an error margin of 8 per cent. The regional forecasts for this year were 100 per cent, 96 per cent and 100 per cent of the regional LTA respectively. The actual rainfall was 93 per cent, 90 per cent and 94 per cent. In the last two years, however, the model did not do as well at the regional level.

While IMD scientists reject the criticisms of the model, they do acknowledge that the model has not performed as well as they expected it to. There is an ongoing effort to refine the model, the result of which was a change in the parameter set used in the forecast model last year. Four of the originally used parameters were replaced by four new ones. One of the replaced ones was the latitudinal position of the high pressure ridge along the longitude 75 E at 500 millibar level (at a height of 6 km). Since its discovery in 1978 by A.K. Banerjee and associates, this parameter was always regarded as being strongly correlated to the performance of the monsoon. In many statistical models, this parameter has often emerged as the most important single variable for monsoon forecast. The ridge basically governs the mid-tropospheric (at a height of 6-7 km) wind circulation pattern. It was considered to be indicative of the transition of the atmospheric circulation from winter type to summer type. Empirically, it had been found that a southward displacement of the ridge compared to its normal position in April (at about 15 N) was unfavourable for a good monsoon and a northward displacement was favourable.

According to V. Thapliyal, one of the authors of the 16-parameter model, the influence of this parameter has been weakening over the years. In an analysis he carried out in 1997, Thapliyal showed that while some parameters were very stable over the years, others were unstable. Their 'correlation co-efficients (CCs)' fluctuated from year to year, rendering the current influence of some of them on rainfall forecast very weak. In that analysis, he had argued that operational models should be revised every year by replacing those parameters which have been less stable during recent years. Indeed, according to M. Rajeevan, Director of Long-Range Forecasts at the IMD, Pune, the CC of the 500 mb ridge parameter has, from a strongly positive value in the past, become slightly negative in the last couple of years. It is from such a perspective that the IMD altered four of its parameters last year.

"We are perhaps in the midst of an epochal variation of the parameters," says Rajeevan. "Maybe the 500 mb ridge shows a decadal variation. Indeed, our analysis of historical data suggests that such variations were there in the past as well," he says. Some 100 years ago, scientists at the Indian Institute of Tropical Meteorology (IITM), led by its present Director G.B. Pant, had shown that rather than the April position of the ridge, the difference in the latitudinal position between March and April was better correlated to the performance of the monsoon. However, Rajeevan said that while over 30 land and ocean meteorological parameters were being constantly monitored for their influence on the monsoon, they had so far not considered this variant of the parameter.

"Monsoon is an interplay of several atmospheric factors and it is quite possible that while the influence of some of them weakens, the combined effect of the key forcing parameters is somehow still the same. Right from (Gilbert) Walker's days, an epochal variation has been observed in the correlation of parameters. But to appreciate the fact that the influence of certain parameters has declined, one should first know how much variance of the rainfall was accounted for by each one of these. But this the IMD has never done in order to understand the relative importance of each parameter in the context of the model," points out D.R. Sikka, former Director of IITM.

Striking a different note, Srinivasan says: "I am not really surprised that parameters such as the 500 mb ridge are losing their importance. When there is a long-term climate change (owing to natural or man-made factors), you can expect the basic parameters influencing the monsoon to change. If you look at the factors governing droughts during the last 120 years, only two factors were found to be causing drought. One was El Nino (the warming of waters in the Pacific Ocean off the Peruvian coast) and the other was the Eurasian snow cover. These two are not necessarily independent factors. In these 120 years there has been a warming of more than 0.6 C. The Eurasian snow cover appears to have declined steadily in spring during the past 25 years. The fact that the 1997 El Nino did not influence the monsoon has been a major puzzle. Until that is understood, one cannot know for sure why monsoons behave the way they do." According to Rajeevan, it is not clear if global warming is the cause of epochal variations in the monsoon variables though suggestions to that effect has been made.

EVEN though the four El Nino-related parameters - El Nino temperature anomalies (in the previous and current years), the Southern Oscillation (the pressure see-saw between Tahiti and Darwin in Australia) and the Darwin Pressure Tendency, as well as the Eurasian snow cover - continue to be among the 16 parameters being used currently, IMD scientists do recognise that the influence of ENSO (the combined effect of El Nino and the Southern Oscillation) on the monsoon has considerably reduced in recent times. "The effect of ENSO has become minimal and the 1997 monsoon was a clear reflection of that," says Rajeevan. Indeed, it is to the credit of the IMD model that, while many meteorologists had expected a deficient monsoon on the basis of the severe El Nino that year, the IMD predicted a normal monsoon, though on the lower side of the LTA at 92 per cent. The actual rainfall was, however, 102 per cent, implying that the IMD model too was off the mark, quantitatively speaking.

According to Rajeevan, more than the influence on the Pacific, conditions over the Indian Ocean and the neighbouring seas, notably sea surface temperatures (SSTs), have begun to show a strong correlation with the monsoon rainfall. The fact that two of the four new parameters - the South Indian Ocean SST and the Arabian Sea SST - relate to the Indian Ocean region is a recognition of the growing influence of the seas around it on the monsoon rather than on "teleconnections" such as the El Nino. Having said this, it must also be pointed out that there is evidence to suggest that the influence of El Nino on the monsoon is dependent on SST anomalies over the Indian Ocean. A positive SST anomaly has been found to offset the effect of the El Nino. The two new parameters related to SST anomalies over the Indian Ocean region are in fact positively correlated to the monsoon rainfall. While this may be the reason for retaining the ENSO-related variables in the model as well as for adding SSTs in the Indian Ocean region to the parameter set, it is also a reflection of the complexity of the forces that drive the monsoon and their changing facets. Similarly, the influence of the Eurasian snow cover is also being watched by IMD scientists. Data from the Eurasian snow cover monitoring project at Rutgers University in the United States clearly show that its extent fell significantly in the 1990s.

The upshot of the above is that the LRF of the monsoon, rather than becoming simpler with years of modelling, is turning out to be more and more complex with even the atmospheric variables that are believed to influence the monsoon strongly becoming unstable over time. The surprising thing is that, in spite of this the monsoon visits the subcontinent with unfailing regularity. Some have even begun to take another look at the physical basis that is currently applied to study the monsoon.

The canonical explanation for the monsoon is that the differential heating between the land and the oceans during summer drives the monsoonal winds. In a research work published early this year, Srinivasan has tried to give a new insight to the mechanism. According to him, in the tropics the contribution of horizontal temperature gradients and moisture gradients to the overall energy balance is negligible compared to the temperature gradients and the moisture balance in a vertical atmospheric column. He has shown that precipitable water (that is, the total amount of water vapour in the column) plays an important role in determining the strength of the monsoon. Based on this thermodynamic approach, he has been able to model successfully the rainfall pattern in Africa, Asia and South America over the years. In particular, his model reproduces the 1997 Indian monsoon rainfall fairly well.

If Srinivasan is right, moisture content at different heights - just as wind and pressure gradients - could well become a key predictor that needs to be included in statistical models. With the availability of moisture data from the Indian Remote-Sensing Satellite IRS-P4 (Oceansat), this should emerge as a new research area for Indian meteorologists.