GeoAI Foundation Models for Environment Understanding: A Case Study on Road Safety Analysis

Abstract

Traffic safety remains a global public health challenge, and the physical road environment plays a decisive role in accident occurrence. Conventional approaches based on tabular crash records and engineered road features capture only a part of this visual environment, and most data-driven models remain correlational, offering limited insight into the causal mechanisms through which features shape traffic safety. This thesis asks how Geospatial Artificial Intelligence (GeoAI) foundation models can improve the extraction, reasoning, and evaluation of environmental road safety factors, and addresses the question through two studies built on a shared data infrastructure across Chicago, Los Angeles, and Madison. The first study develops a knowledge-guided Retrieval-Augmented Generation system that instructs Vision-Language Models to reason about driving risk over multi-source street-view, satellite, and map imagery; evaluated against both crash-rate ground truth and human perception rankings, the few-shot configuration with retrieved domain knowledge improves aggregated accuracy and produces reasoning texts that are more diverse and richer in domain-specific content. The second study turns from perception to explana-tion through a pipeline that integrates zero-shot semantic segmentation, SHAP-based interpretable modeling, and Generalized Propensity Score weighted causal inference on Chicago intersections, identifying Sidewalk Ratio, Building Obstruction Ratio, Emergency Space, and Drivable Area Ratio as statistically significant risk-reducing factors and Visual Openness and Visible Obstacle Density as the only robustly risk-increasing indicators. Taken together, the two studies produce complementary evidence on the same set of inter-sections, demonstrating that GeoAI foundation models, when combined across perceptual and causal modalities, can bridge the gap between automated streetscape assessment and the evidence-based decision needs for urban transportation planning.