Social sensing of floods in the UK

Abstract

"Social sensing" is a form of crowd-sourcing that involves systematic analysis of digital communications to detect real-world events. Here we consider the use of social sensing for observing natural hazards. In particular, we present a case study that uses data from a popular social media platform (Twitter) to detect and locate flood events in the UK. In order to improve data quality we apply a number of filters (timezone, simple text filters and a naive Bayes 'relevance' filter) to the data. We then use place names in the user profile and message text to infer the location of the tweets. These two steps remove most of the irrelevant tweets and yield orders of magnitude more located tweets than we have by relying on geo-tagged data. We demonstrate that high resolution social sensing of floods is feasible and we can produce high-quality historical and real-time maps of floods using Twitter.

Fig 1. Number of tweets collected per day during the whole collection period 22/10/2015 and 25/11/2016.

Fig 2. Number of tweets collected per day during the whole collection period 22/12/2015 and 04/01/2016 at each filter level.

Fig 3. Number of relevant tweets collected with location info in each field: GPS-tagged tweets, location field GPS coordinates, location field toponyms, message text toponyms.

Fig 4. Floodiness grid, 64 × 64, over England and Wales on 5/12/2015 using (r, α, T) = (1.0, 0.15, 0.1).

Using tweets collected in 1 hour windows. White indicates no tweets. Colour bar units are floodiness relative to daily max. Top left: 10am-11am. Top Right: 1pm-2pm. Bottom Left: 4pm-5pm. Bottom Right: 9pm-10pm.

Fig 5. Floodiness grid, 64 × 64, over the North East on 5/12/2015 between 4pm and 5pm using (r, α, T) = (1.0, 0.15, 0.1).

White indicates no tweets or zero population. Colour bar indicates floodiness relative to daily max.

Fig 6. Tuning relative floodiness threshold T by varying text versus location weighting r and population scaling exponent α.

Each point corresponds to the average precision and recall over 15 days for a different triple of r, α, T.

Fig 7. Tuning absolute floodiness threshold T by varying text versus location weighting r and population scaling exponent α.

Each point corresponds to the average precision and recall over 15 days for a different triple of r, α, T.

Fig 8. Flood map generated by twitter converted into FFC format for validation.

White indicated no tweets. Colour bar units are relative floodiness. Top Left: Floodiness grid (64 × 64) over England and Wales on 28/10/2015 using (r, α) = (1.0, 0.15). Top Right: Showing only grid squares above threshold 0.1. Bottom Left: Counties with floods on 28/10/2015 according to Twitter. Bottom Right: Counties with floods on 28/10/2015 according to the FFC, with g_h set to 1 for flooded counties.