Abstract: This study proposes a GeoAI (Geospatial Artificial Intelligence) framework that transforms traffic information collected in image format into a spatially interpretable spatiotemporal network structure to predict traffic congestion. The study area is Atlanta, Georgia, USA, where a spatiotemporal network was constructed based on traffic image data and road network information collected at 10-minute intervals. Previous studies have focused on developing and applying GeoAI models that demonstrate high predictive performance by utilizing various variables such as traffic volume, average speed, weather information, and road characteristics (Akhtar & Moridpour, 2021). However, these approaches are limited in variable selection and reflecting regional characteristics, presenting challenges in predicting congestion within complex urban road networks. Accordingly, this study explored the optimal variable configurations for traffic congestion prediction and compared and evaluated the performance of various GeoAI models. The analysis results revealed that the GC-LSTM (Graph Convolutional Long Short-Term Memory) model, which incorporates historical traffic congestion patterns and traffic accident information as variables, exhibited the highest predictive performance. However, prediction accuracy showed a spatially heterogeneous trend depending on the location, with prediction errors increasing in areas with high time-series volatility. In particular, the optimal variable configurations differed according to regional characteristics in major arterial roads penetrating the downtown area, such as the northeastern intersection of I-285 and I-85 and the eastern intersection of I-285 and I-20. These findings suggest that selecting a GeoAI model suitable for regional characteristics, in addition to optimizing variable configurations, is essential for improving the accuracy of traffic congestion prediction. Furthermore, by emphasizing the importance of customized variable design that considers the temporal uncertainty and spatial heterogeneity of traffic data, this study provides empirical evidence that can contribute to data-driven policy establishment for solving urban problems.

Recommended citation: Lee, S., Hwang, C. S., & Seong, J. C. (2025, June). 시공간 네트워크 기반 GeoAI 프레임워크를 활용한 교통혼잡 예측: 애틀랜타 사례 분석 (Traffic congestion prediction using Spatio–Temporal Network–based GeoAI framework: A case study of Atlanta) [Paper presentation]. Proceedings of the Korean Geographical Society Conference Republic of Korea, 177.