Sean Ernst - February 12

Convective Meteorology (Mesoscale Dynamics) Seminar

Cracking the TORFF Code: Testing a Neural Network Coding Scheme on Broadcaster Coverage of TORFF Events

Sean Ernst

Friday, February 12^th

3:00pm

Join Google Meet

https://meet.google.com/ksh-txvg-kni

Simultaneous Tornado and Flash Flood (TORFF) events are a difficult to message and potentially deadly hazardous weather threat that is common in the contiguous US.  As the primary source of weather information for the general public, the burden of communicating the threat from these events falls on the shoulders of broadcast meteorologists, though studying the messages broadcasters share can be extremely time- and resource-intensive.  In this study, we seek to develop a neural network-based deterministic model using Snorkel that can automatically code transcribed broadcaster coverage using labeling functions derived from a scheme developed by researchers.  As an early test of the efficacy of this model, we use the model to code broadcaster coverage of TORFF events, specifically to define how often broadcasters discuss flood and tornado threats and compare those coverage rates across time, location, and event type.  Findings suggest that the deterministic model can code broadcaster coverage without human intervention reasonably well for the hazards code, though continued training of the model will improve its output.  The model’s coding reveals that broadcasters heavily favor coverage of tornadoes during TORFF events and may be focusing coverage on tornadoes more with time.  This result suggests that broadcaster coverage can be efficiently interrogated with this automatic coding tool, and that broadcaster coverage during TORFFs is biased towards the more sensational threat posed by tornadoes.  Future work will seek to identify what actions and impacts broadcasters focus on during these TORFF events, as well as to analyze more TORFF events to confirm these findings in a larger sample.