A Deep Learning Approach to Lightning Nowcasting and Forecasting

Lightning plays a fundamental role in Earth’s climate system and is a frequently occurring natural hazard. However, lightning remains a relatively unpredictable area of meteorology, especially in terms of lightning frequency per convective event, with limited ability for nowcasting and forecasting of lightning occurrence. The goal of this study is to develop a deep learning algorithm able to replicate lightning stroke density on a climatological average, as well as on a convective feature basis. We use a convolutional neural network (CNN) containing combinations of the following variables at 0.5-degree by 0.5-degree spatial resolution and a 3-hourly temporal-resolution over a domain that encompasses most of the Western Hemisphere: lightning from the World-Wide Lightning Location Network (WWLLN), precipitation rate from NASA’s Integrated Multi-satellite Retrievals for GPM (IMERG) and convective available potential energy (CAPE), cloud base height (CBH), two-meter temperature (T2M) and zero degree level (ZDL) from the European Centre for Medium-Range Weather Forecasting (ECMWF). We train the CNN on the years from 2010 to 2018, and tested on the years 2019 to 2022. Model performance was evaluated on a four-year average through changes to the initial seed used to train the model, the loss function used, transformations to the lightning dataset, and changing the spatial and temporal resolution of the input datasets. We further examined the value of 11 input variable combinations, from single variables to all five variables used in training. Preliminary results show that changing the initial seed, as well as changing the loss function from mean squared error to mean-squared logarithmic error, does not greatly impact model performance when running the model with more than one input variable. Results vary amongst the variable combinations, but amongst the different initial seeds and loss functions, the r-squared values remain above 0.75 for every model configuration over both land and ocean. Model performance is improved when using higher time resolution training set but not necessarily a higher spatial resolution. For example, a 1-degree by 1-degree spatial resolution and a 3-hourly time resolution resulted in an r-squared between predicted and observed lightning frequency 0.1 higher than that using 0.5-degree by 0.5-degree spatial resolution and a daily time resolution. The model is able to reproduce the approximate evolution of lightning stroke density of individual convective events, but tends to overestimate the stroke density on a 3-hourly basis. Future work will include a steeper penalty for overestimating lightning occurrence during training. These results show that larger-scale weather forecasting and earth system models could significantly improve lightning stroke density parameterizations by incorporating deep learning results.