A detailed regional drought study is carried out in the southern peninsula of India to characterize the spatio-temporal nature of droughts and to predict the drought magnitudes for various probabilities in the homogeneous drought regions. The method of several random initializations of the cluster centres of the K-means algorithm is suggested for the identification of the initial regions in the context of drought regionalization, which is shown to perform better than the initialization from the Ward’s algorithm and the Ward’s algorithm itself. The peninsula is classified into seven spatially well-separated homogeneous drought regions. The robust L-moment framework is used for the regional frequency analysis of drought magnitudes computed using the standardized precipitation index. The Pearson type III is found to be appropriate for regional drought frequency analysis in six of the regions, while the robust Wakeby distribution is suggested for one region. Low magnitude droughts are frequent and dominant in the northern part of west coast, the north-eastern coast and its adjoining inland region, while high magnitude droughts are less in number and are experienced in semi-arid central part, southern part of western coast, south-eastern part and north-western inland region. The spatial maps of drought magnitudes indicate that at higher return periods (100 and 200 years) the south-eastern part of the peninsula is likely to encounter high magnitude droughts, while the central region is likely to experience the same at lower return periods (10 and 50 years). Hence these regions need to be given special importance in the drought mitigation planning activities. © 2016, Springer Science+Business Media Dordrecht.