An explainable machine learning technique to forecast lightning density over North-Eastern India

Increasing lightning fatalities over India is a concerning subject. Especially, it is pretty crucial over North-Eastern part of the country where lightning is extremely frequent. Given the complex nature of the problem, machine learning can be an excellent option in such forecasting scenarios. However, such dynamic processes seek proper transparency of the model. The current work attempts to devise a model for short range prediction (one month ahead) of lightning density based on primary atmospheric parameters from satellite data with a lead time of one month over North –Eastern and Eastern part of the country. Random Forest regression seems to outperform other models explored, with a R2 of 0.86 and an MAE of 0.0071. The interpretation of the model output using SHAP index reveals that 2 meter temperature at previous two months and CAPE and K-index at previous month has a positive impact on the output of the model whereas, instantaneous surface heat flux of previous month and two month prior K-index has an inhibiting effect on model`s output. The use of machine learning techniques for atmospheric predictions without the shed of the black box can be of importance to the scientific community. Such studies especially over lightning prone tropical regions can be crucial in meteorological forecasting applications.