Abstract: Gentrification is a process that indicates the socioeconomic and physical transformation of a geographic area over time. Gentrification usually happens when affluent residents and businesses gradually move into an under-invested area, often leading to the displacement of the lower-income, original residents. Gentrification status is a crucial indicator for understanding urban dynamics and its implications on residents and communities within the region. Prediction of future gentrifying areas is essential for ensuring sustainable urban planning. While some researchers have made efforts to predict the vulnerable areas early, they mostly rely on decennial census data, and fail to capture rapid changes occurring in an urban area. To address this, the study proposes a novel deep learning architecture that predicts future gentrification by analyzing historical and current building footprint data, which reflects rapid changes in the area's physical appearance. The proposed method makes two key contributions. First, it introduces a dual-scale transformer encoder that generates region-level embeddings by encoding building characteristics to capture spatial correlations within a region. Second, it employs a cross-attention transformer to detect temporal changes by analyzing how these embeddings evolve over time. We show the effectiveness of the proposed method by conducting thorough experiments. The result demonstrates that the method effectively predicts future gentrification status of neighborhoods, offering an alternative to census-based approaches.

 

Presenter: Tasmiah Tamzid Anannya (04123054002)

Venue: Graduate Seminar Room