Towards an operational cloud-to-ground lightning probability model for the Alaskan tundra using machine learning and the Weather Research and Forecast (WRF) model
Jordan Alexis Caraballo-Vega, NASA GSFC, jordan.a.caraballo-vega@nasa.gov (Presenter)
Tatiana Loboda, University of Maryland, loboda@umd.edu
Mark Carroll, NASA GSFC, mark.carroll@nasa.gov
Allison Baer, University of Maryland, aebaer@terpmail.umd.edu
Dong Chen, University of Maryland, itscd@umd.edu
Although remote and inaccessible, Arctic tundra is a critical biome for the global community. Its accelerated warming trend is causing profound ecological changes that have the potential to significantly impact ecosystem services critical to Arctic people. These rapid rates of warming and changes in precipitation regime result in an increase in the frequency and extent of wildfires in a region that has been fire-resistant for millennia. Due to the inaccessibility of the Arctic tundra, cloud-to-ground lightning is the primary ignition source of wildfires. Thus, it is crucial to understand the mechanisms and factors driving lightning strikes in this cold and treeless environment in order to establish the foundations that better support the modeling of fire occurrence in the Arctic tundra. In this study, building on earlier work from our team, we have developed a machine learning model capable of simulating the probability of cloud-to-ground lightning across the Alaskan tundra by integrating weather variables from the Weather Research and Forecast (WRF) model and historically recorded lightning strikes from the Alaska Lightning Detection Network. Our workflow includes an end-to-end GPU-accelerated machine learning pipeline for (1) the parallel extraction and calculation of weather features from the WRF model output, (2) training and inference of cloud-to-ground lightning probability using single and multi-GPU resources, (3) and a containerized version of the WRF model for further portability across on-premises and cloud computing environments. A variety of temporal windows were used to apply the trained model and evaluate how well it performed under various scenarios throughout the study area. In addition, explainable artificial intelligence techniques were used to further analyze the drivers of the model and their impact in its performance. Our findings demonstrate the potential of forecasting cloud-to-ground lightning probability across the Alaskan tundra using WRF-simulated weather variables at 24-hour, 48-hour, and 10-days into the future.
Associated Project(s):
Poster Location ID: 3-32
Presentation Type: Poster
Session: Poster Session 3
Session Date: Thu (May 11) 3:00-5:00 PM
CCE Program: TE