The use of crowdsourced social media data to improve flood forecasting

Abstract

Reliable flood forecasting systems are essential for predicting and mitigating the impact of flooding worldwide. However, minimising flood forecast uncertainties remains a challenging task due to many sources of uncertainty in underlying flood simulation modelling. Such uncertainties can be reduced by employing data assimilation techniques to dynamically incorporate the most recent available observations into the system while accounting for existing uncertainties in both models and observations. However, traditional observations often lack the necessary temporal or spatial resolution, limiting the adoption of data assimilation methods for real-time applications. In contrast, data collection through crowdsourcing has grown in popularity with the potential to provide high spatiotemporal resolution data, especially in urban areas. Nevertheless, the use of crowdsourcing is still impacted by validation uncertainties and data quality, which makes it a complementing data to traditional observations rather than an alternative data source. This paper presents a novel methodology for assimilating crowdsourced social media data to improve a 2D flood forecasting model through various update strategies. The methodology was tested against a real case flood event of the 2017 Phetchaburi flood (Thailand), and the performance of different update strategies was evaluated with reference to the calibrated model output obtained from a particle swarm optimisation algorithm. Empirical results demonstrate that global state updates suffer from inconsistencies in predicted water levels, whereas topographically based local state updates provide encouraging results. Specifically, the improvement due to the local state update alone is short-lived, and findings indicate that a longer lasting improvement in flood forecasting performance can be achieved through a combination of both state and boundary updates. Overall, the results indicate the feasibility of utilising crowdsourced social media data to improve the performance of flood forecasting systems for urban environments.