A machine learning approach for prediction of seasonal lightning density in different lightning regions of India.

Lightning is one of the most severe weather events causing significant loss of human lives and resources. Increasing number of lightning fatalities due to recent climatic changes is emerging out to be a serious concern for India during last few years. Proper characterization and parameterization of the same, therefore, is extremely crucial. However, lightning is an extremely dynamic phenomenon having enormous spatio‐temporal inhomogeneity especially over such a vast country like India with varied topographic and climatological features. Therefore, proper parameterization of lightning activity over India needs consideration of different lightning climatologies. This study has attempted to resolve the issue by regionalizing Indian subcontinent in different lightning climatologies based on lightning density and associated atmospheric variables that is, CAPE, specific humidity at different pressure levels, temperature, k index and cloud particle size and identified seven distinct lightning climatologies over India. A regression model is proposed for estimating the annual and seasonal (monsoon and pre‐monsoon) lightning activities over the seven resulting lightning zones based on the said atmospheric variables using machine learning techniques. Four machine learning models have been tested among which Random forest has shown the best accuracy. The regression model has shown an R‐squared score of 0.81 during monsoon season and 0.71 during the pre‐monsoon. The atmospheric features based on their influences on the lightning activity in these seven climatologies has been ranked which presented the evidences of largely varied interplay between different atmospheric variables and lightning over different parts of the country and during different seasons. [ABSTRACT FROM AUTHOR]