A minimalistic seasonal prediction model for Indian monsoon based on spatial patterns of rainfall anomalies

Abstract

Seasonal prediction of Indian Summer Monsoon Rainfall (ISMR, rainfall during June to September over India) has remained an important scientific challenge for decades, due to its complex multi-scale nature. Statistical and dynamical seasonal ISMR predictions have traditionally relied on the relatively small variability of the spatially averaged monsoon rainfall over India, known as All India Monsoon Rainfall (AIMR). While this has served to mitigate socioeconomic impacts to some extent, overall prediction skill has remained relatively low (Wang et al. in Nat Commun 6:7154. https://doi.org/10.1038/ncomms8154 , 2015) while the spatial variability is anything but small. Here we find that the classification of deficit/ dry or surplus/ wet monsoon years based on AIMR does not add value at a regional scale due to the very high heterogeneity of monsoon rainfall, even in the extreme years. To demonstrate the need and the potential to predict this important spatial heterogeneity, we improve the classification of monsoon years by focusing on the spatial patterns of rainfall anomalies within different meteorological subdivisions of India. We apply the k-means clustering methodology and also offer cluster validation. Cluster validation reveals the existence of nine clusters of monsoon years with distinct spatial patterns of monsoon rainfall anomalies. The composite anomalies of sea surface temperature (SST) and winds during March to May (MAM) and June to September (JJAS) show distinct hydroclimatic teleconnections indicating potential predictability of regional monsoon rainfall at seasonal scale. To demonstrate the potential prediction pathways for spatial patterns of ISMR, we develop a statistical seasonal prediction model based on Classification and Regression Tree (CART) between SST over different oceanic regions as predictors and monsoon classes as predictands for the period 1901–2010. Search for the potential regressors reveals the importance of new predictors such as Atlantic Nino and SST over North Pacific region along with conventional predictors such as El Nino Southern Oscillation (ENSO), Indian Ocean Dipole/Zonal Mode (IODZM), etc. Validation of the method is performed for 2011–2015 and the model is able to predict the regional pattern of monsoon rainfall for 4 out of the 5 years. The purpose of this prediction exercise is to demonstrate the need to focus on the process and predictive understanding of these clusters and their predictability.