Estimating Individuals’ Suffering Levels Induced by Disasters via Social Media: A Case Study of the 2023 Beijing–Tianjin–Hebei Extreme Rainfall

Abstract Extreme hazard events can severely threaten the well-being of society. To understand their impacts on the well-being of people, social scientists have proposed various indicators related to individuals’ suffering, and analyzed them mainly via post-disaster surveys. Social media has shown its value in capturing people’s perceptions of disasters, but few scholars have investigated individuals’ suffering levels based on social media data. Accordingly, this study used social media data and developed a hybrid model that combines machine learning classifiers and lexicon-based approaches to estimate the well-being impacts of disasters, which are measured by individuals’ physical, emotional, and social suffering levels. Six machine learning models were trained to categorize the suffering as reflected by disaster-related posts on Weibo. Convolutional neural networks (CNN) were found to be the most accurate model, and was selected to classify all posts into four groups (no suffering, physical suffering, emotional suffering, and social suffering). In each classified group of posts, word co-occurrence analysis was then applied to construct suffering lexicons. By combining the classification results from CNN and suffering lexicons, this study proposed optimization-based algorithms to estimate suffering levels for posts across space and time. The proposed model was applied to the 2023 Beijing-Tianjin-Hebei extreme rainfall event. The temporal analysis revealed that individuals’ physical suffering levels declined more rapidly than mental and social suffering levels. Spatial analysis revealed that individuals’ suffering presented high spatial heterogeneity, and that the hazard-affected regions experienced significantly greater levels of suffering. This hybrid model provides an analytical tool for timely and human-centered disaster emergency management.