Doctor of Philosophy (PhD)


Environmental Sciences

Document Type



With the rapid development of social media, many researchers chose to use social media data to conduct studies of disaster resilience. This dissertation addresses three fundamental challenges related to social media use and disaster resilience analysis including: how the spatial analysis scale affects the results of social media use disparity and its impacts on disaster resilience; how different vulnerability and resilience measurement methods can generate very different results; and how social media use disparity can be identified to help improve the fairness of AI-based emergency rescue model development. Therefore, the first part of the dissertation measured the relationships between disparities in social media use and disparities in community resilience at both the county level and zip code area level, using Twitter use data during Hurricane Sandy as an example. We found zip code areas that have major transportation hubs and commercial activities or low night-time population are major factors increasing Twitter use indices and hence the correlations. The second part compared three indices of vulnerability and resilience at the block group level, including the CDC Social Vulnerability Index (SVI), the Community Resilience Estimates (CRE) from the Census Bureau, and the Resilience Inference Measurement (RIM) method developed by our team, using the Mississippi River Delta as the study region. We found that SVI and CRE are highly similar to each other, whereas RIM only agrees with the two for about half of block groups in the study area. However, the RIM model has empirical validation and inferential ability. The third part of the dissertation used Twitter data during Hurricane Harvey as a case study, and aimed to determine which type of communities have positive difference (sending more rescue requests via social media than expected) and which are negative (sending less rescue requests via social media than expected). This information can be utilized in a fairness-aware AI rescue prediction model to minimize social media representation bias. We found positive block groups had higher social vulnerability, higher flood risk, lower elevation, and lower digital access. These findings provide new information and insights on the effects of analysis scale, measurement methods, and social media use disparity on fairness modeling and the resilience of communities.



Committee Chair

Lam, Nina

Available for download on Tuesday, March 18, 2025