A Benchmark Dataset for Lightning Nowcasting in Latin America

In 2023, the World Meteorological Organization (WMO) began the Artificial Intelligence (AI) for Nowcasting Pilot Project (AINPP) initiative, with the goal to evaluate and possibly implement AI solutions for nowcasting convective hazards in developing nations. This initiative is part of the broader commitment to the Early Warning for All initiative (EW4All), a United Nations effort which aims to provide protection against extreme weather events for every individual on the planet. Benchmark datasets are being developed for the short-term prediction of precipitation and lightning over regions of Latin America, Africa, and Asia. This research details the benchmark dataset for lightning nowcasting in Latin America.

In collaboration with scientists from Mexico, Brazil, Argentina, and Japan, two domains were selected: one covering Mexico and a second domain covering most of South America. Thirty dates were selected for each domain from 6 convectively active months (5 days per month), with the months being different for each domain. The truth metrics for lightning nowcasting are derived from the GOES-16 Geostationary Lightning Mapper (GLM) flash-extent density product. Flash-extent density is simply the number of flashes that traverse a given location (or pixel) over a given period of time. For this dataset, the flash-extent density was summed over 10-minute increments at every location, and then maximum-accumulated over 1-hour and 2-hour periods. Thus, the final 2D target fields are the 1) maximum 10-minute flash-extent density over 1 hour, and the 2) maximum 10-minute flash-extent density over 2 hours. For the purposes of the AINPP, any lightning is the target (i.e., flash-extent density > 0), and thus the flash-extent density grids will be binarized for that evaluation.

The dataset will be publicly available in cylindrical equidistant projection (i.e., equally spaced latitudes and longitudes) with no correction for parallax effects. Routines to evaluate model data on the benchmark dataset will also be provided