Project 25008 VisMo-Health: An Explainable GeoAI Framework Linking Street-View Visuals, Human Mobility, and Post-Disaster Health Outcomes
Project Title
VisMo-Health: An Explainable GeoAI Framework Linking Street-View Visuals, Human Mobility, and Post-Disaster Health Outcomes
Bing Zhou, Assistant Professor, University of Tennessee, Knoxville
Project Description
VisMo-Health fuses two rich data streams—street-level imagery and anonymized mobility traces—to trace the full arc of a natural disaster’s impact on community well-being. First, a Vision Transformer–based model, augmented with Grad-CAM and Layer-wise Relevance Propagation, assigns fine-grained damage and environmental-quality scores to millions of Google Street View and Mapillary panoramas captured before and after hurricanes, floods, or wildfires. Parallel Graph Neural Networks interpret daily origin-destination flows from SafeGraph and similar datasets to quantify how residents’ activity spaces shrink, fragment, or recover.
The project then introduces a Daily Disaster Stress Index (DDSI) that combines visual damage severity with mobility disruption through an attention-based fusion network. SHAP and counterfactual explanations reveal why certain neighborhoods exhibit heightened DDSI and how specific improvements (e.g., debris removal, transit restoration) could reduce stress.
Finally, quasi-experimental Difference-in-Differences and spatial Bayesian hierarchical models link time-lagged DDSI exposure to Alzheimer’s admissions (6–24-month lags) and depression indicators (1–12-month lags), controlling for air quality, socioeconomic vulnerability, and spatial autocorrelation.
Tasks and Responsibilities
Experiment design; data visualization; deep learning model building and training; paper writing
Minimum Qualifications Graduate students in relevant fields, such as geography, spatial data science, computer science, urban planning, etc.
Terms of the Project
10-12 months
Key Words
Disaster, GeoAI, Dementia, Explainable AI, Spatial Analysis
Research Topics
How can pre- and post-disaster street-view imagery be automatically parsed to quantify neighborhood-scale damage, environmental change, and loss of green infrastructure?; In what ways do natural disasters alter daily human mobility patterns, and how quickly do these patterns rebound across different communities?;To what extent do combined exposures to visual damage and mobility stress predict subsequent changes in cognitive (e.g., Alzheimer’s-related hospitalizations) and mental-health outcomes (e.g., depression prevalence)?
Deliverables
Publishable papers, open-source tools, and open-source datasets.