(Please see Weather, 50, 12, pp.449-467 for full paper)
In 1877, India experienced a serious famine caused by highly deficient monsoon rainfall. Following this, the Government of India called upon H.F. Blanford, who had by then established the India Meteorological Department (IMD) in 1875 and was the Chief Reporter of IMD, to prepare monsoon forecasts. Thus, Blanford was the first to attempt a forecast of monsoon based on the hypothesis that 'varying extent and thickness of the Himalayan snows exercise a great and prolonged influence on the climate conditions and weather of the plains of northwest India' (Blanford, 1884). The success of Blanford's tentative forecasts during 1882-85 encouraged him to start operational LRF of monsoon rainfall covering the whole of India and Burma in 1886. Since then the LRF of monsoon has become an important operational task of the IMD. Sir John Elliot, who succeeded Blanford in 1895, utilized the weather conditions over the whole of India and surrounding regions to prepare an elaborate forecast of the monsoon rainfall. The forecasts after 1895 were based on (i) Himalayan snow cover, October to May, (ii) 'local peculiarities' of pre-monsoon weather in India and (iii) 'local peculiarities' over Indian Ocean and Australia (Thapliyal, 1987).
Later, Sir Gilbert Walker (1908, 1918, 1923), the then Director General of IMD, has initiated extensive studies of worldwide variation of weather elements such as pressure, temperature, rainfall etc., with the main aim to develop an objective method for LRF of monsoon rainfall over India, expanding on the works of Hildebrandsson (1897), Lockyer and Lockyer (1904) and others who had drawn attention to the global-scale oscillations in surface pressure. These studies led Walker to identify three large-scale pressure seesaw patterns; two in the Northern Hemisphere (North Atlantic Oscillation, NAO and North Pacific Oscillation, NPO) and one in the Southern Hemisphere (Southern Oscillation, SO). While the NAO and NPO are essentially regional in nature, the SO has since been recognized as a phenomenon with global-scale influences. The SO was later linked to the oceanic phenomenon called El Niño in the east-equatorial Pacific characterized by warming of the sea surface along the Peru coast; this led to the theory of Walker Circulation (Bjerknes, 1969). Walker also succeeded in removing the subjectivity in the earlier forecast methods by involving, for the first time, the concept of correlation in the field of LRF of monsoon rainfall. With these pioneering contributions to the field of LRF, Sir Gilbert Walker has been most closely identified with the early attempts at monsoon forecasting in India and his findings are relevant even today.
Walker (1924) also attempted LRF for sub-regions of India by dividing the country into 3 homogenous regions namely (i) Northeast India, (ii) Peninsular India and (iii) Northwest India. Regression formulae were developed separately for these three regions, which had been subsequently revised several times (Thapliyal ,1987). Subsequent to Walker's work, very little progress was made in LRF of monsoon rainfall until the early eighties when several studies have re-established the strong link between the monsoon rainfall variability and ENSO using better data sets.
Most of the studies in LRF are primarily based on statistical and empirical techniques. Diagnostic studies of historical data sets, over the years, have brought out several predictors for the monsoon rainfall forecasting. These parameters represent different components of the coupled atmosphere-ocean-land system. In this section, the characteristic features of the known predictors grouped into appropriate categories are described.
Identification of reliable predictors for LRF is achieved by analysing the relationships between AISMR and regional/global fields of several surface/upper-air parameters. Though correlation coefficients (CCs) have helped to broadly identify the various forcings on the monsoon as detailed below, they also exhibit marked sensitivity to the data window considered, leading to secular variations in their magnitude as well as sign, imposing some limitations on fixing a set of predictors in an absolute time frame. The LRF predictors, pertaining to the preceding winter and spring seasons can be broadly classified into 4 groups representing the following features:
Click on the above link for a comprehensive list of the predictor parameters showing statistically significant CCs with AISMR in the recent period.
Major Predictors of the Indian Summer Monsoon
The progressive development of surface heat low over central parts of Pakistan and adjoining northwest India and the decaying wintertime upper air westerly wind regime in the pre-monsoon season play a very important role in the evolution and performance of Indian summer monsoon. In view of this, several workers have developed predictors for the monsoon rainfall using data on various surface as well as upper-air meteorological parameters from India and adjoining regions. Details on some of the important predictors are presented below:
Monsoon being the result of land-sea heating contrast involving large-scale seasonal reversal of pressure, temperature and winds, many studies have been carried out to identify useful predictors based on the pressure and thermal fields during antecedent winter and pre-monsoon seasons. Parthasarathy et al. (1992c) developed a predictor parameter, which they called West Central India (WCI) pre-monsoon (MAM) pressure, represented by the mean of sea level pressure (SLP) at six stations (Jodhpur, Ahmedabad, Bombay, Indore, Sagar and Akola) located in the core region of high correlation, which showed a CC of -0.63 (significant at 1% level) with the AISMR during 1951-80. Earlier, Parthasarathy et al., 1990 found that the mean surface temperatures at these stations during MAM season also showed high correlation (0.6) with monsoon rainfall for the period 1951-80.
Mooley and Paolino (1988), using maximum and minimum temperature data for the period 1901-75, have shown that a predictor based on May minimum temperatures over western Indian region has good potential for LRF. Krishna Kumar et al. (1995) identified two predictors based on the minimum temperatures during March over East Peninsular India and during May over west central India.
The operational LRF model (Gowariker et al,, 1991) of the India Meteorological Department also uses three minimum temperature parameters representing northern, central and east coastal areas of India.
The mean latitudinal location of the 500 hPa ridge along 75°E in April over India, first identified by Banerjee et al. (1978), is considered to be one of the most important predictors. The mid-tropospheric anticyclone over southern India migrates from 11.5°N in January to its northernmost position of 28.5°N during July. From October, the ridge starts shifting back southward. Mooley et al. (1986) found a CC of 0.71 (significant at 0.1% level) between the April ridge location and AISMR during 1939-84; a more northward location indicates better performance of the monsoon and vice versa. It is conjectured that the northward and southward displacements of this mid-tropospheric anticyclonic circulation are related to the seasonal march of the solar radiation and the associated diabatic heat source. The anomalies in the seasonal evolution of the mid-tropospheric circulation, as measured by the April ridge location, can be taken to be a good precursor of the slowly varying planetary-scale circulation. A delayed northward displacement of the ridge is considered to indicate large-scale anomalous descending motion over the Indian region (Shukla and Mooley, 1987). In a detailed diagnostic study using daily locations of the 500 hPa ridge during the pre-monsoon months of March-May for the period 1967-87, Krishna Kumar et al. (1992) found that the ridge location in March showed a CC of -0.47 with AISMR, while in April it showed a CC of +0.63. They also found that the negative correlation of the March ridge was more dominant with the monsoon rainfall of the peninsular India, while the positive correlation of the April ridge was more dominant with the monsoon rainfall of northern India. The difference between the two ridge locations (April minus March) shows a CC of 0.73 with AISMR. Though the 500 hPa ridge has shown consistently significant relation with monsoon rainfall in recent years, the subjectivity involved in the determination of its location imposes some limitation on its reliability.
Monsoon circulation over India involves marked changes in the upper tropospheric wind field. Keeping this in view, many recent studies have shown that the upper air winds during their pre-monsoon transition phase can provide a useful predictor. Verma and Kamte (1980) and Joseph et al. (1981) have identified the association between Indian monsoon rainfall and 200 hPa meridional wind component for the month of May, and indicated its potential for prediction of the seasonal rainfall. Recently, Parthasarathy et al., (1991a) have further investigated the relationship between meridional wind index (arithmetic average of 200 hPa meridional component of wind for May at Bombay, Delhi, Madras, Nagpur and Srinagar) and AISMR by using an extended data set for the period 1964-88 and found a CC of -0.72 (significant at 0.1% level).
Owing to the importance of ENSO phenomenon on the climate variability in the tropics and over several other regions of the globe, many predictors have been developed representing the strengths of both its atmospheric component, the Southern Oscillation (SO), and its oceanic component, the El Niño. Walker (1924) developed an index of SO based on a combination of pressure, temperature and rainfall. Subsequent workers have developed several other indices of SO using different combinations of stations mainly based on pressure data (Wright, 1975; Trenberth, 1976). The most commonly used Southern Oscillation Index (SOI) as a measure of the strength of the Walker Circulation across the Pacific, is taken as the normalized difference between the SLP anomalies at Tahiti and Darwin.
However, the CCs between various SO Indices during preceding winter and spring seasons and AISMR are not sufficiently high for LRF as the relationship develops simultaneously and even follows the monsoon season (Bhalme and Jadhav, 1984; Parthasarathy and Pant, 1985). Shukla and Paolino (1983) found that the CC of AISMR with Darwin SLP during winter (DJF) was positive while that during spring (MAM) was negative. Because of this change of sign in the CCs, the CC of the winter to spring tendency (MAM-DJF) of SLP at Darwin with AISMR during 1901-81 was found to be much higher (-0.46, significant at 0.1% level) than the CCs of individual seasons. Thus, the winter to spring tendency is considered to be a reliable precursor for the nature
of SO during the monsoon and later months.
The planetary-scale tropical SLP anomalies associated with the SO occur in conjunction with the episodes of large-scale sea surface temperature (SST) anomalies (El-Niño/La Niña) in the tropical Pacific (Rasmusson and Carpenter, 1982; Lau, 1985; Shukla and Mooley, 1987). It has generally been observed that warmer SSTs in central and eastern parts of equatorial Pacific are associated with lower monsoon rainfall (Angell, 1981; Khandekar and Neralla, 1984). Parthasarathy et al. (1988), using COADS SST data during 1951-80 have identified three important regions in the Pacific Ocean whose SSTs have shown significant relationships with the AISMR. These three regions are, (1) 14°-26°N, 128°-140°E; (2) 14°-20°N, 176°-160°W and (3) 14°N-10°S, 148°-100°W, whose MAM-DJF tendencies in SST showed CCs of 0.4, -0.51 and -0.52 respectively with AISMR. The intensities of El Niño events are generally assessed on the basis of the average SSTs over three Niño regions in the Pacific Ocean, widely known as NINO1+2, NINO3 and NINO4. Among these, only NINO4 SST (MAM-DJF) shows statistically significant CC (-0.54 during 1951-80) with AISMR.
The observational studies of Saha (1974), Pisharoty (1976) and Cadet and Reverdin (1981) and modeling studies of Washington et al. (1977) and Shukla (1984) have established the importance of the role of cross equatorial flow over Indian Ocean and moisture flux from both Indian Ocean and Arabian Sea regions in Indian monsoon rainfall, particularly over the west coast. However, there is some difference of opinion regarding the relative dominance of fluxes from Indian Ocean and Arabian Sea. The East African low-level jet is one of the most important manifestations of the cross-equatorial flow involving large-scale moisture and momentum transport (Findlater, 1977). Cross-equatorial flow develops during the onset phase of the monsoon season and its strength, if predictable, can provide an important predictor for AISMR. Hastenrath (1987) suggested the use of SSTs over the Arabian Sea and Parthasarathy et al. (1988) attempted to represent the strength of cross-equatorial flow in the Indian Ocean region by Nouvelle-Agalega SLP difference as well as SSTs. However, none of them could find practical utility in the development of LRF schemes for AISMR. Thus, in spite of the known physical link of cross-equatorial flow with monsoon rainfall, there has been very little progress in identifying useful predictors based on it.
With the availability of long period data on hemispheric mean surface temperatures in the early 1980's, efforts have been made to link the Northern Hemispheric (NH) mean surface temperature anomaly with the strength of the Indian summer monsoon. Verma et al. (1985) identified the NH winter surface air temperature anomaly (January+February) as an important predictor for LRF of AISMR. This parameter showed CC of 0.56 during 1951-80 maintaining its significance even during the later years, and is recognized as one of the most important predictors.
The early attempts of Blanford and Walker brought out the association of greater Himalayan snow cover with deficient monsoon rainfall over India during 1880-1920. Subsequently, the reported snow accumulation showed very large variability and the relationship with the monsoon rainfall was found to be opposite to what it was in the earlier four decades (Shukla, 1987). Though this led to the dropping of snow cover as a predictor by the IMD in 1950, the recent availability of more reliable satellite estimates of snow cover extent revived interest on this parameter and several studies have shown a statistically significant inverse correlation between the Eurasian/Himalayan winter snow cover and AISMR (Hahn and Shukla, 1976; Dey and Bhanukumar, 1983; Dickson, 1984; Bhanukumar, 1989). The spatial extent of Eurasian/Himalayan snow cover during the winter is considered to be an important slowly varying boundary condition for the subsequent development of monsoon circulation and therefore is a potential predictor with strong physical link (Vernekar et al., 1995). Parthasarathy and Yang (1995) examined the relationship between Eurasian snow cover (1974-1992) and AISMR and found a CC of -0.47 for the month of February. Satellite data, unfortunately, are not available for long periods and also suffer from several inhomogeneities (Ropelewski et al., 1984) which need to be clarified to develop a useful predictor representing satellite derived snow cover.
The Quasi-Biennial Oscillation (QBO) in the mean zonal wind of the tropical stratosphere at a height range of 20-30 km, with easterly and westerly wind regimes alternating regularly with a period varying from 24 to 30 months, has been found to have strong association with the performance of monsoon over India. Mukherjee et al. (1985) found a significant simultaneous CC of +0.39 between the monsoon rainfall and zonal wind (JJA) at 30 hPa using wind data of Balboa (9°N, 80°W) during 1951-1982. Taking cue from the fact that the wind reversal first appears above 30 km and propagates downward at a speed of about 1 km/month, Bhalme et al. (1987) related the January 10 hPa zonal wind anomalies at Balboa with AISMR and found a CC of 0.52 (significant at 1% level) during 1958-85. AISMR tends to be less (more) than normal during easterly (westerly) anomaly. Though very promising, the 10 hPa data from Balboa are not regularly reported after 1988 and efforts are being made by some groups to use 10 hPa wind data from other tropical stations like Ascension Island (8°S, 14°24'W) and Singapore (1°40(N, 104°E) having reasonably long-period data.
The spatial distribution of the relationship of any predictor with Indian summer monsoon rainfall can be studied by computing its CCs with the sub-divisional monsoon rainfall. Most of the predictors are known to show the highest CCs over northwestern and central India, and the lowest CCs over northeast and extreme southern parts of the Peninsula. The spatial patterns of CCs of most of the predictors with the sub-divisional rainfall have been found to be strikingly similar to those of AISMR. These patterns are generally consistent with the predominant first EOF mode structure of the monsoon rainfall over India (Shukla, 1987) and therefore essentially reflect the spatial coherence of AISMR.
Most of the studies on LRF of Indian monsoon rainfall are based on empirical or statistical techniques. These statistical techniques range from simple correlation analysis to advanced procedures such as canonical correlation analysis and neural networking.
Almost all the predictors identified so far have been based on correlation analysis. Though correlation is a very useful diagnostic tool in bringing out the association between various meteorological fields/circulations, it is highly sensitive to the data window over which it is calculated, both in terms of the position and length of the window in the time domain. This imposes certain limitations on the reliability of the predictors.
The most commonly used statistical technique for LRF of monsoon rainfall is the linear regression analysis. A large number of regression models (simple as well as multiple) have been proposed so far (Hastenrath, 1991). The predictors for the model are either subjectively chosen representing various important forcings on the monsoon, as perceived by the individual scientists or entered into the scheme by following some objective criteria. Both approaches have their own limitations; the subjective selection may not optimize the variance explained while the objective selection is highly sensitive to the data window and may result in overfitting to the data sample (Thapliyal, 1987; Parthasarathy et al., 1988; Hastenrath and Greischar, 1993). As the regression models tend to acquire sample-specific characteristics, their reliability is better assessed by testing on as large an independent data set as possible.
Auto-regressive integrated moving average (ARIMA) models were also used to forecast the AISMR as well as the monsoon rainfall over Northwest-India and Peninsular India, which were reported to have shown marginally better forecast skill over the multiple regression models (Thapliyal, 1990). However, the autocorrelations (-0.12, 0.04, 0.08, -0.09 and -0.02 for lags 1 to 5 respectively) in AISMR during the period 1871-1990 are statistically insignificant (Parthasarathy et al., 1994). In view of this, the applicability of ARIMA models for monsoon rainfall forecasting is doubtful.
Gowariker et al. (1989) developed parametric and multiple power regression (MPR) models with 15 predictors for LRF of AISMR, which were later modified (Gowariker et al., 1991) to include 16 predictor parameters. The parametric model is qualitative and indicates the likelihood of the monsoon rainfall to be excess or deficient, depending upon the proportion of favourable/unfavorable parameters out of the total of 16 parameters. This method is highly sensitive to the nature of the predictor data set and its inclusion of an equitable and realistic representation of all the forcings involved. The power regression model claims to account for possible non-linear interactions of different climatic forcings with the Indian monsoon system, by iteratively determining the power to which each predictor has to be raised for optimum CC with AISMR. Notwithstanding the success claimed by this model in operational LRF during the period 1988-97, the model with 16 independent variables based on just 31 years of data does suffer from the problem of overfitting and high redundancy (intercorrelations) among the predictors. Further, the very idea of quantifying non-linear interactions of various climatic forcings on monsoon rainfall on which the so-called 'power' regression model is based, is questionable, as the predictors involved in the model were originally identified based on their 'linear' correlation with AISMR. The reliability of this model vis-à-vis other conventional regression models came into sharp focus in 1994 (forecast: 92% and observed: 110%) and 1997 (forecast: 92% and observed: 102%), when its predictions of below normal monsoon failed like many other models. The power regression model therefore needs rigorous statistical testing using longer and homogeneous data sets and independent verification.
Thapliyal (1987) has developed dynamic stochastic transfer (DST) models for the prediction of AISMR as well as the monsoon rainfall over peninsular and northwestern India. In this method, the position of 500 hPa sub-tropical ridge over India has been considered as an input for the dynamic transfer component, coupled with a stochastic transfer system represented by an ARIMA process and the monsoon rainfall as the output.
Thapliyal (1990) evaluated the relative performance of multiple regression, MPR, ARIMA and DST models and found that the DST model has the highest accuracy among those models. However, this model considers only one predictor as input and it is desirable to develop DST models involving multiple parameters representing various forcings of the monsoon system.
Due to considerable spatial variability of monsoon rainfall over India, the forecasting of countrywide mean rainfall is of limited practical utility. Keeping this in view, Krishna Kumar (1994) attempted LRF of sub-divisional monsoon rainfall by canonical correlation analysis (CCA) technique, using monsoon rainfall data from 29 meteorological sub-divisions in India as the predictand data set and the SST patterns (MAM-DJF) over the Pacific Ocean and minimum temperatures (March and May) over India as predictor data set during the period 1958-87. He found that the spatial extent and the magnitudes of useful skill scores for sub-divisional LRF are much larger than those obtained with multiple regression analysis (Prasad and Singh, 1992). Thus, CCA technique appears to be a promising method for LRF of the spatial patterns of monsoon seasonal rainfall.
Navone and Ceccatto (1994) have used 'feed-forward' neural network technique for the prediction of Indian monsoon rainfall with two predictors (500 hPa ridge location and Darwin SLP tendency from January to April). They demonstrated that the neural networks could make a better use of the predictive information available in the predictor data. However, further work needs to be done to conclusively establish the superiority of this advanced statistical tool over other methods.
Compared to the number of studies using empirical methods, the studies on the seasonal prediction of monsoon rainfall using general circulation models (GCMs)
are very few. This may be partly because of the lack of skill in the simulation of monsoon rainfall over the Indian subcontinent by most of the GCMs (Gadgil et al., 1992). Also, there are marked differences found in the monsoon precipitation simulated by different GCMs (WCRP, 1992). The fact that the simulation of the summer monsoon rainfall over the Indian region is very sensitive to the initial conditions (Palmer et al., 1992) also throws up serious problems in this context. However, recently Ju and Slingo (1995) and Soman and Slingo (1995) have demonstrated that the seasonal integrations of GCMs could simulate strong and weak monsoon circulations based on the SST distributions over tropical Pacific and Indian Oceans. Thus, while a realistic simulation of monsoon rainfall in absolute terms is yet to be achieved by most GCMs, the sensitivity of the model monsoon to interannual variability in tropical circulation seems to be closer to the observed characteristics of the monsoon. This is an encouraging sign for LRF research using GCMs.
Reliability of any empirical/statistical forecast model depends largely on the stability of the relationships between the predictand and the predictor parameters. The presence of secular variation in the strength of correlation between monsoon rainfall and some predictor parameters has been realized since Walker's time. Some of the recent studies have addressed this problem either in terms of the stability of the relationship of the predictors when identified for the first time, or to understand the possible reasons for such variations in already known predictors.
Most of the predictors showed insignificant CCs with AISMR till 1950 and the CCs became significant only around the year 1951. However, there are some exceptions like Darwin SLP tendency and Pacific SST which showed significant CCs for relatively shorter periods around 1891 and 1921. The CCs of some predictors have even changed their sign during the early part of this century.
Parthasarathy et al. (1991b) have examined the relationship between Bombay SLP and AISMR with an extended data set for the period 1847-1990 and found turning points in the CCs around the years, 1870, 1900 and 1940. They attributed these turning points to delineate between two alternating regimes in the monsoon circulation, identified earlier as 'meridional' and 'zonal' types by Fu and Fletcher (1988). Parthasarathy et al. (1991b) concluded that the Indian summer monsoon had passed through two meridional (1871-1900; 1941-90) and two zonal (1847-1870; 1901-1940) circulation regimes during the past 150 years. They also found that the relationship between Bombay SLP and AISMR becomes dominant only when the ENSO variance in Bombay pressure is high and falls apart when the ENSO variance is small.
Thus, it appears that during the last 3 or 4 decades the ENSO phenomenon has played a dominant role in the climate variability in general and monsoon variability in particular. During this period, all the three facets of atmosphere-land-ocean system seem to have been strongly coupled. It is not clear how and why this coupling was not dominant in the earlier decades. Annamalai (1995) suggests that the presence of decadal-scale oscillations in the predictors themselves may possibly be responsible for the instability in the relationship between AISMR and its predictors. The sliding CCs also suggest that the predictability of the monsoon itself may be having secular variations, which probably is one of the reasons for the temporary 'lull' in LRF research immediately after Walker's time.
While there are a large number of predictors identified so far, most of them fall into one of the four categories (see table) and are also highly interrelated. Most of the parameters, even across the categories described above, show significant CCs between them. The presence of such high multicollinearity among the predictors imposes the problem of redundancy and unnecessary loss of degrees of freedom when they are used in large numbers in regression-based forecast schemes.
Though the predictors can be classified into four different groups based on their known physical linkage with the monsoon, the forcings represented by them are not entirely exclusive to their respective groups. The forcings represented by the various predictors can be objectively delineated and the common variance among them pooled into a set of independent principal components. However, very little work has so far been done using this approach. The results of a recent work by Krishna Kumar et al. (1997) indicate that a single component accounts for about half of the total variance in the predictors. This component apparently represents ENSO type variability in the predictors and is highly and significantly correlated with AISMR. This clearly shows that the ENSO has a ubiquitous influence on the monsoon circulation, playing a dominant role in the LRF.