Conjunction of hard k-mean and fuzzy c-mean techniques in clustering and identifying some critical meteorological parameters for thunderstorm formation over a metro city of India during pre-monsoon season.

Among the metro cities in India, Kolkata is chosen incidentally for the present study. The study aims at clustering of pre-monsoon days of the urban area, Kolkata (22º32′N, 88º20′E) (India) in two groups (thunderstorm days denoted by TS and non-thunderstorm days denoted by NTS in the literature) using hard k-mean technique, backward selection procedure and fuzzy c-mean algorithm (FCM). Various thermodynamic and dynamic parameters that are already identified by several scientists as responsible for thunderstorm formation have been considered here for different atmospheric layers up to 500 hPa. The study is performed in two stages: In the first stage, the hard k-mean technique is applied to cluster the days of a semi-supervised dataset in the two categories mentioned above. Then the backward selection procedure is used to find the best possible combinations of the theoretically influential atmospheric parameters considered in the study, that plays the dominant role in the categorization on the basis of performance score (PC). In the second stage of the work, fuzzy c-mean algorithm is applied to the same semi-supervised dataset of parameters to clarify the results obtained in the first stage. This study is performed separately for the morning (0000 UTC) and afternoon (1200 UTC) atmosphere as it is already revealed that there is a structural difference between the morning and afternoon atmosphere of Kolkata, India. In the first stage with the (thunderstorm and non-thunderstorm) dates of 0000 UTC (morning) data reveals that the combination of maximum vertical velocity and P–PLCL at 1000 hPa level performs better in detecting the pre-monsoon thunderstorm days, whereas with those at 1200 UTC (afternoon) data showed vertical wind speed shear for (1000–850) hPa layer, maximum vertical velocity, P–PLCL at 1000 hPa level and (θes–θe) at 850 hPa level dominate better in detecting thunderstorm days. It is interesting to note that these findings are supported by FCM in the second stage for both morning and afternoon atmosphere. Research highlights: The study in the first stage reveals that the hard k-means clustering technique with two different metrics, Euclidean distance produces more stable results than Manhattan distances. In the second stage the fuzzy c-means algorithm helps to clarify the reasons behind the results obtained in the first stage. In this study proportion correct, HK skill score and backward selection procedure indicates that combination of maximum vertical velocity and P–PLCL furnish better results in morning whereas in afternoon along with the above two parameters, vertical wind speed shear (1000–850 hPa) and (θes–θe) at 850 hPa also gives better results for the categorization of pre-monsoon days of Kolkata (India), in two groups, thunderstorm and non-thunderstorm. [ABSTRACT FROM AUTHOR]