Mitigation of Spatial Bias in Social Media Data for Disaster Relief

Abstract

Social media platforms generate large volumes of geotagged data that are becoming more valuable for disaster response and situational awareness. However, the opportunistic and skewed nature of Volunteered Geographic Information (VGI) makes it spatially biased, making such data less representative and distorting disaster relief efforts. This study proposes a new method to mitigate the effects of this spatial bias in geotagged social media data, specifically tweets made during Hurricane Sandy in New York City. The study uses environmental covariates like building density, elevation, distance to shoreline, and Modified Normalized Difference Water Index (MNDWI) to measure the representativeness of tweet samples in each covariate space. Two optimization techniques, Genetic Algorithm (GA) and Stochastic Gradient Descent (SGD)-based optimization, were used to determine an optimal weight for individual tweets to adjust their spatial representation by improving their distribution in the covariate space. This study found that the mitigation process is effective and that the SGD-based approach significantly outperformed GA, with overall similarity scores increasing from 0.679 (initial) to 0.895. The improvement was particularly notable in flood-sensitive areas, as evidenced by the MNDWI similarity score rising from 0.605 to 0.948. The weighted tweet maps generated using the optimized weights identified the disaster-affected areas more effectively and were in close agreement with the official inundation data published by FEMA. This research contributes a scalable and data-driven framework for spatial bias mitigation in social media data and helps enhance the reliability of such data in post-disaster mapping and response planning.