Exploring linkages between regional precipitation and sea surface temperatures using Bayesian learning

This thesis develops probabilistic models for forecasting inter-seasonal and intra-seasonal monsoon rainfall under uncertain inputs. Monsoons are characterized by seasonal reversal of winds that bring copious amount of oceanic water onto land masses thereby affecting regional hydrology and socio-economic life of the region. Monsoons exhibit substantial inter- and intra-seasonal variability that is linked to sea surface temperature (SST), making SST a key input to the state-of-the-art models for forecasting monsoon rainfall. Historical SST records that extend back to 1850s have been obtained from a wide range of sources ranging from traditional in situ measurements to remote sensing observations, and consequently have uncertainties varying in both space and time. In this study, a set of algorithms were developed to engage heterogeneous uncertainties of SST records in rainfall forecasting models by using correlation, regression, and principal component analysis. These new algorithms were developed within a common framework of Bayesian learning, and together they provide a comprehensive tool to account for uncertainty in forecast models. The developed algorithms were first tested and compared with traditional methods that ignore input uncertainty on synthetic data, and then applied to forecast the most extensive and intriguing of the Earth\u27s monsoon, the Indian summer monsoon. SST data and associated uncertainties were used as inputs to forecast monsoon\u27s inter-seasonal variability and its intra-seasonal oscillations. The latter are marked by active and break spells which were identified using hidden Markov models. The results suggest that engaging data uncertainty in hydrologic forecast models improves their prediction performance and provides better assessment of their predictive capabilities. The prediction skill of Indian monsoon using SST data was found to be low suggesting that (i) linkages between the two are weaker than expected from the past studies, or (ii) better and longer datasets are needed to identify these linkages