NLP-driven crowdsourcing for urban flood monitoring: insights from mumbai

With increasing extreme rainfall events, densely populated megacities face recurrent flooding, necessitating real-time flood monitoring systems for citizens. Social media platforms like Twitter (now X) can complement conventional sensor-based flood monitoring. However, their utility is limited by low geotagging and a data stream cluttered with extraneous content. We take the case study of Mumbai to demonstrate how Natural Language Processing (NLP) can be leveraged to identify flood-relevant tweets. Applying NLP to a historical Twitter dataset (2017 – 2022), we match the text patterns in tweets against lexical keyword datasets to geocode them and classify their sentiment as ‘positive’ or ‘negative’. We observe a decline in the daily positivity ratio of tweets with increasing daily average rainfall for all years. We find that negative tweets are concentrated in areas of orographic rainfall, higher population density, and lower elevations. Further, we analysed extreme rainfall impacts on urban transportation using a network theory-based global efficiency loss (GEL) metric. For this, we configured Q-NEAT, a GIS-based network model, to perform time-optimized routing by integrating geolocated negative tweets alongside hotspots identified by the municipal corporation. We found that Twitter explained an additional 1.13 %, 20.96 %, and 67 % of the total GEL detected on the extreme rainfall days, 16th July 2021 (199.87 mm), 2nd July 2019 (200.12 mm), and 25th June 2018 (153.14 mm). Such impacts are not fully detected by the municipal corporation’s thumb rule for identifying waterlogging based on precipitation amounts and field surveys. Our findings underscore the advantages of crowdsourced data for enhancing urban flood monitoring and impact assessment.