| |
 |
|
|
|
 |
 |
|
 |
|
All forecasts are based on a Wavelet banding technique described in Webster and Hoyos (2004). Although the results obtained with the statistical scheme in hindcast mode were succesful, at this point the operational forecasts that are being issued are experimental.
This Page is Currently Disabled
Most attempts at predicting monsoon variability have concentrated on the seasonally averaged rainfall over the Indian subcontinent some months in advance. Predictors such as El Niño-Southern Oscillation phenomenon and sea-surface temperature variability in the Indian Ocean or the Northwest Pacific Ocean have been used. These relationships account for about 30% of the variance of mean seasonal monsoon rainfall but the statistics are nonstationary and correlations become insignificant or even disappear for decades at a time. Furthermore, numerical models have had difficulty in simulating the mean seasonal
precipitation patterns in monsoon regions and have not shown substantial skill in seasonal monsoon predictions. However, even if the statistics were stable and models could forecast the overall seasonal anomaly with some skill this knowledge of the seasonal mean anomaly is not downscalable to provide information about the spatial patterns of the variability or the temporal variability of the rainfall such as the timing of “active” and “break” periods of the monsoon. As the timing and magnitude of these intraseasonal 10-40 day variations of monsoon rainfall are largely responsible for the success or failure of regional agriculture it is argued that skillful and timely forecasts of intraseasonal variability would allow optimization
of agricultural and water resources management in monsoon regions.
A physically-based Bayesian scheme is developed for forecasting monsoon intraseasonal variability (MISO). The scheme employs a wavelet banding technique and linear regression to forecast 5-day average rainfall variability over regions of South Asia on 15-30 day time scales. The predictand can be any factor that is intimately tied to the MISO and for which there exists a long-term time series. Examples of predictands used in this site are: rainfall over Central India, rainfall over the Ganges valley, rainfall in the state of Rajasthan, and the river discharge of the Brahmaputra and Ganges at the borders of India and Bangladesh. Having determined the predictand, the choice of predictors is the critical problem with statistical schemes.
There are two philosophies used for choosing predictors: “frequentist” (von Storch and Zwiers 1999) and Bayesian (Leonard and Hsu 1999). In the former technique, predictors are chosen principally from their statistical relationship with the predictand. For example, Walker’s initial efforts to predict monsoon variability (e.g., Walker 1923) came from the examination of a wide range of climate parameters that has seemingly no apparent physical relationship. Ultimately, this a posteriori technique broke down for prolonged periods (e.g., Kumar et al. 1999, Torrence and Webster 2000). The second technique chooses predictors based on their
physical relationship with the predictand. Simply, the predictors are chosen in an a priori fashion based on the physics of the phenomenon one is trying to predict. Here the predictors become the Bayesian priors of the system.
Having chosen the predictand and the sets of predictors, we have a choice of a number of statistical tools that will produce a forecast. We choose a linear regression scheme to relate the predictand and the predictors. The predictors are chosen using the Bayesian philosophy. However, before the linear regression process is executed, we employ a “wavelet banding” technique to the predictand and to the predictors to sort the time series into specific spectral bands. This technique is described below. Together, the wavelet banding and the linear regression are referred to as the Wavelet Banding method or WB method.
To obtain the final forecasted value, the predictions in all bands are added together. The isolation of bands in a time series by wavelet analysis is the key factor in the WB statistical scheme (see appendix) because it allows the regression tool to identify, independently in each band, the existing relationship between the predictand and predictors. That is, the “noise” from other spectral bands is not allowed to affect the construction of the regression equation.
