A pipeline to enhance animal vehicle collision analysis in crash report dataset

Animal vehicle collisions (AVCs) are a global safety concern, requiring analysis and predictive models for understanding and mitigation. Police crash report data are one of the main sources of AVC data globally. However, they are prone to reporting policy change and other inconsistencies, particularly in rural areas, hindering the development of predictive models. Through development of a robust approach for data cleaning, quality control, feature selection, and contribution level identification, this study proposes a pipeline to address this shortcoming.

North Dakota crash data set is used as a case study due to high rates on AVC in this rural region and its diverse wildlife ecosystem. Theil's U association index, and chi-square tests were implemented in the pipeline to evaluate the proposed pipeline effectiveness. The pipeline detects and removes skewed proportion samples, while addressing data collection inconsistency, low variance, and duplicated features.

Pipeline imposed 3.5% sample size and 88.9% feature size reduction on the original crash data over 20 years. Observation on the modified dataset revealed year, day, and driver features had the lowest while hour, county, and speed limit had the highest statistical contribution to the AVC. Light, hour, and month were lumped in daily solar cycle and represented as a single temporal feature that can be used effectively to develop predictive model. Finally, presented pipeline increased spatiotemporal integrity while reducing the runtime by 92.46% for the association analysis.