Natural Language Processing to Predict National Weather Service Products from Winter-Related Transportation Incidents

Adverse weather conditions are responsible for millions of vehicular crashes, thousands of vehicular deaths and billions of dollars in economic and congestion costs. The National Weather Service (NWS) has implemented a series of initiatives to improve and enhance stakeholder and public communication of risks associated with hazardous weather events. These initiatives include Weather-Ready Nation and Pathfinder, the latter of which is a joint effort with the Federal Highway Administration and statewide transportation agencies across the United States. The goal of Pathfinder is to provide critical weather information to transportation network operators and incident managers to respond and adapt quickly to changing weather conditions and situations. However, the goals of Pathfinder can further be enhanced through the usage of artificial intelligence, specifically natural language processing, to crowdsource both meteorological and transportation information in real‑time from sources such as social media. This research focuses on a case study approach of leveraging data from social media and NWS products issued during severe winter-weather associated with numerous transportation-related incidents.

Natural language processing obtains usable content from human language processed by a computer. This natural language processing is implemented using a Long-Short Term Memory (LSTM) neural network to perform classification of NWS winter weather products (e.g., winter weather advisories, winter storm/blizzard warnings) from input text data such as local storm reports, vehicular crash narratives, and road closure information. In addition to using an LSTM for classification, preprocessing techniques such as tokenization and word encoding are implemented as part of this overall process. The output from this LSTM can be used to assist with automated labeling of events, issuance of winter weather products, and providing motorist alerts on dynamic/variable message signs. Moreover, assessing the language of NWS products in conjunction with transportation incidents can better inform both the transportation and meteorology communities concerning the successes and remaining challenges of Pathfinder implementation. The findings of this work may suggest future automation that can alert both communities to rapidly developing situations and coordinate efforts more efficiently. Further, this automated framework can inform other hazard risk assessment and communication studies.