Machine Learning Quality Control of Lightning Data for Tropical Cyclone Intensity Forecasting

Statistical models for tropical cyclone (TC) intensity prediction, such as the Statistical Hurricane Intensity Prediction Scheme (SHIPS) and Rapid Intensification Index (RII) incorporate the areal extent and symmetry of cold cloud tops from geostationary longwave infrared imagery. The cirrus canopy associated with TCs, however, can obscure the convection below cloud tops, limiting the low- and mid-level information provided by infrared imagery. The Geostationary Lightning Mapper (GLM) provides new information on the location and intensity of lightning-producing convection beneath the cirrus canopy. Research with ground-based lightning networks has shown improvements to the RII when lightning predictors are added, and more recent research has found even stronger relationships to intensity change when TC structure is also considered.

The potential for GLM observations to be used for forecasting and analyzing tropical cyclone structure and intensity has been complicated by inconsistencies and artifacts in the GLM data. Quality control (QC) techniques have been developed to help remove clear artifacts in the GLM dataset such as blooming events, spurious false lightning, bar effects, and sun glint. Simple QC methods include scaled maximum energy thresholds and minima in the variance of lightning group area and group energy are evaluated. More sophisticated methods using convolutional neural networks are also explored. Each QC method is successfully able to remove artifacts in the GLM observations while maintaining the fidelity of the GLM observations within tropical cyclones.