An Examination of Dryline Convection Initiation Using Machine Learning

Forecasters often have difficulty in determining whether deep, moist convection (DMC) will occur along the dryline. Sometimes convection does not initiate despite the apparent presence of the key ingredients of DMC: moisture, instability and lift. What are the differences between drylines that produce convection and those that don't? Are any variables better discriminators than others in determining whether DMC will occur?

Following the creation of a climatology of dryline convection in the Southern Great Plains, a machine learning technique was applied to a sample of dryline days to attempt to identify variables important for predicting convection initiation. Dryline days were identified using Weather Prediction Center surface analyses for April, May and June 2006—2015. NEXRAD radar mosaics, visible and infrared satellite imagery were then used to establish whether DMC occurred along the dryline. A dryline was considered convective if an identifiable cell that formed along the dryline had a reflectivity of at least 40 dBZ for a continuous period of one hour or more. Rapid Refresh Model analyses were obtained for convective and non-convective dryline days, before a gradient boosting algorithm was applied to assess the predictability of dryline convection.

The skill of the algorithm in predicting dryline convection initiation varies with location relative to the dryline. Gradient boosting is more skillful using data obtained from the region immediately south-east of the dryline than elsewhere. In describing how the gradient boosting technique was implemented we will attempt to explain why this regional variation in skill occurs. We will reveal which variables are the best predictors of DMC before discussing the strengths and weaknesses of using gradient boosting as a tool to examine drylines.