GeoAI-driven analysis of urban activity shifts using geotagged social media data across the COVID-19 pandemic

Abstract

The COVID-19 pandemic has profoundly disrupted urban systems, reshaping human mobility and activity patterns. While previous studies have documented macro-level declines using aggregated mobility indicators, the evolution of localized activity patterns throughout the pandemic phases remains unclear. This study integrates geotagged social media imagery with a geospatial AI-based analytical framework to examine long-term shifts in urban activity in Seoul, South Korea. We collected 14,875 Flickr photographs from 2018 to 2023 and applied HDBSCAN to delineate urban areas of interest. We classified image content using a ResNet-50 model fine-tuned on the Places365 dataset to capture semantic activity patterns. The results revealed sharp contractions of indoor-oriented categories-such as shopping malls and coffee shops-during 2020-2021, alongside a rise in outdoor categories, including parks and plazas, reflecting both formal restrictions and heightened risk perceptions. By 2022-2023, indoor activities gradually reappeared while outdoor activities remained significant, suggesting a rebalancing of urban behavior rather than a simple return to pre-pandemic conditions. The analyses highlight differentiated adaptation pathways: cultural resilience in the central district, rapid commercial recovery in Gangnam-Songpa, and balanced leisure-work restoration in Mapo-Yeouido. Interpreted through the stimulus-organism-response framework, these findings provide empirical evidence of how external shocks reshaped perceptions and behaviors, mediated by district-specific functions. The study demonstrates the value of linking behavioral framework with geospatial AI-based spatial analysis to monitor place-specific adaptation. By offering a 6-year longitudinal perspective, it contributes to both conceptual integration and empirical insights for adaptive and resilient urban planning.