Assessment of the level of road crash severity: Comparison of intelligence studies
Sina Shaffiee Haghshenas, Giuseppe Guido, Alessandro Vitale, Vittorio Astarita
Abstract
In measuring road safety, accident severity is a key concern. Crash severity prediction models inform researchers about the severity of a crash based on a variety of criteria. To date, an enormous amount of research has been done on crash severity, and several models have been suggested to forecast crash severity utilizing existing test data or simulated datasets created using regression or classification analysis. In this study, a new approach was developed to determine the level of road crash severity (LRCS) using a large amount of real-existing data (1627 cases) by applying machine learning methods to the roads of Calabria in southern Italy. This study has three main goals: building prediction models based on classification approaches with the highest accuracy; comparing the performance of two supervised learning methods, including artificial neural networks (ANN) and convolutional neural networks (CNN), for predicting the LRCS; as well as determining the role of each of the contributing parameters in the LRCS and presenting a ranking by performing a sensitivity analysis. Finally, based on the accuracy values, it has been found that there isn't a salient difference between the predicted models. But it should be noted that 68.4% accuracy for the testing dataset implies the CNN model was superior to the ANN model, which scored 61.74% accuracy. This is acceptable if the minor variation in modeling accuracy is desired. Also, the results of the sensitivity analysis showed that the number of vehicles and the road element had the highest and lowest effect rates on the LRCS, respectively.