Graph-Based Deep Learning and Multi-Source Data to Provide Safety-Actionable Insights for Rural Traffic Management

This study confronts the significant challenges inherent in Traffic State Estimation (TSE) for rural arterial networks, where sparse sensor coverage and complex, dynamic traffic flows complicate effective management and safety assurance. Traditional TSE methodologies, often dependent on single-source data streams, fail to accurately model the intricate spatiotemporal dependencies present in such environments. This fundamental limitation precipitates critical safety hazards, including pervasive over speeding and dangerous queue spillback phenomena at intersections. To address these deficiencies, we introduce a novel hybrid intelligence framework that synergistically combines a Graph Attention Temporal Convolutional Network (GAT-TCN) with advanced Kalman Filter variants, specifically the Extended, Unscented, and Sliding Window Kalman Filters. The GAT-TCN component is engineered to excel at learning complex, non-linear correlations across both space and time through multi-source data fusion. Empirical validation conducted on a real-world rural toll corridor demonstrates that our proposed model achieves a statistically significant superiority over conventional benchmarks, as rigorously quantified by substantial reductions in both Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). Beyond mere predictive accuracy, the framework delivers transformative safety enhancements by facilitating the proactive identification of hazardous events, enabling earlier detection of over speeding and queue spillback compared to existing methods. Consequently, this research provides a scalable and robust framework for proactive rural traffic management, fundamentally shifting the paradigm from achieving incremental predictive improvements to generating decisive, safety-actionable insights for infrastructure operators.