Unveiling Crash Severity Determinants: A Multisource and Multiresolution Modeling Approach

In this study, we sought to delve deeper into factors affecting crash severity by bridging the existing knowledge gap in traffic crash modeling and the lack of synergy among different data sources. Aiming for enhanced modeling accuracy and new insights from associated factors, we synergized data from multiple sources, including high-resolution CCS and WIM traffic data, crash data, weather data, and roadway geometrics.  The data was compiled with varying temporal resolutions, covering three exposure windows of 5, 10, and 15 minutes prior to the reported crash time. Our methodology entailed a meticulous data curation and fusion process. A hierarchical modeling approach was adopted, consisting of two sequential models. In the first model, we harnessed the power of a tree ensemble to predict the manner of collision, which acted as a contextual proxy for the second model, facilitating a more precise evaluation of crash severity. In the second model, we leveraged a customized transformer model designed for tabular data for improved performance and interpretability. The results unveiled noteworthy patterns. Variables like age, use of safety equipment, prevailing weather conditions, and traffic exposure stood out as significant determinants of crash severity.