Enhancing Tropical Cyclone Forecast Applications using GLM with Machine Learning

Abstract:

Tropical cyclone (TC) intensity forecast improvement, including better rapid intensification (RI) prediction, remains a top priority for the National Hurricane Center. Large-scale atmospheric and oceanic environmental factors, such as vertical wind shear and sea surface temperature, have relatively well-understood influences on TC intensity change, but convective-scale processes also play important roles in intensity evolution. Geostationary satellite longwave infrared imagery is currently used to characterize the symmetry and areal extent of cold cloud tops; however these images are limited by their inability to see through the cirrus canopy that obscures convection below. The Geostationary Lightning Mapper (GLM) provides a new opportunity to observe convection below the cloud tops. Improvements to RI forecasts have been obtained by incorporating lightning predictors from ground-based networks, but recent research has found even stronger relationships between lightning and intensity change when TC structure is also considered. We have summarized the current state-of-the-art knowledge on using GLM for TC forecast applications in a “GLM TC Quick Guide” that will be described in our presentation.

This presentation will also introduce a new project that seeks to better understand the relationships between GLM lightning, TC structure, and intensity change to improve statistical intensity models. Lightning can occur at all stages of a TC’s lifecycle, and the lightning location and pattern can impact intensity change differently based on their relation to TC structure. This project will fuse in-situ and multi-satellite datasets to assess the links between spatial lightning features and TC structure, including their relation to the storm size, 3-D wind field, and eyewall(s). Variations in spatial lightning features, including connections to the diurnal cycle, will then be used to assess the temporal evolution of TC intensity. Linking GLM features to TC structure and intensity change will be completed using a combination of expert physics-based knowledge that is already understood and machine learning-based techniques, including dense and convolutional neural networks.