Analysis Of Motorcycle Crash Severity In Ghana By Machine Learning

The number of registered motorcycles in Ghana is almost one-quarter of total registered motor-vehicles.Mostly in the rural northern region of Ghana,motorcycle riding has long been a conventional and cheap means of transportation.Recently,motorcycle use has become increasingly popular in cities as an alternative economic mode of transportation in congested road networks.Motorcycles crash regularly occur on shared roadways in Ghana,and their related injuries and deaths are significant problems of road traffic safety in Ghana and have seen an upsurge in recent years.Fatalities associated with motorcycle crashes in Ghana are presently second in rank to the fatalities of the pedestrians.This call for a need to research into the factors that cause motorcycle crashes in Ghana.Globally,motorcycle crash analysis is a well-known research area.However,motorcycle crash severity is under-researched in Ghana.Thus,no study investigates factors affecting motorcycle crash severity outcomes in Ghana.There is extensive and growing literature on the classical statistical models for the prediction of the severity of motorcycle crashes.Traditional statistical models have fundamental assumptions and pre-defined correlations that,if flouted,can generate incorrect results.This study employed machine learning based algorithms to predict motorcycle crash severity to address the weaknesses of statistical models.Machine learning based techniques are non-parametric models without the presumption of relationships between predictors and responses variables.Different machine learning based algorithms were evaluated and compared with one another in this study.Motorcycle crash dataset between 2011 and 2015 was extracted from the National Road Traffic Crash Database at the Building and Road Research Institute(BRRI)in Ghana.The dataset was classified into four injury severity categories: fatal,hospitalized,injured,and damage-only.Classification algorithms of machine learning were selected for this study because the target variable(motorcycle crash severity)is having four possible categories outcome(fatal,hospitalized,injured,and damage).Based on their accuracy,five well-known classification algorithms were selected to model the severity of injury in a motorcycle crash after an experiment with several classification algorithms.The selected classification algorithms were: Multilayer Perceptron(MLP),Rule Induction(PART),Classification and Regression Trees(CART),J48 Decision Tree Classifier,and Instance-Based learning with parameter k(IBk).This study also empirically investigates the application of four popular ensemble strategies(AdaBoosting,Bagging,Random Forest,and Majority Vote Combiner)in order to improve the classification accuracy of ordinary single classifiers.These algorithms were implemented in WEKA(Waikato Environment for Knowledge Analysis).These machine learning algorithms were validated using 10-fold cross-validation technique.Also,the relative importance analysis of the attribute was conducted to determine the impact of these attributes on injury severity outcomes.The results obtained using machine learning techniques revealed that the MLP classifier achieved an accuracy of 80.84% for prediction of motorcycle crash severity,with classification precision of 0.867,0.802,0.791 and 0.736 for fatal,hospitalized,injured and damage,respectively.For the PART,the accuracy achieved for prediction of motorcycle crash severity was 82.10%,with classification precision of 0.905,0.794,0.818,and 0.846 for fatal,hospitalized,injured,and damage,respectively.For CART,the prediction accuracy was 82.34%,with classification precision of 0.901,0.800,0.816,and 0.858 for fatal,hospitalized,injured,and damage,respectively.For J48,the prediction accuracy was 81.55%,with classification precision of 0.905,0.787,0.810,and 0.888 for fatal,hospitalized,injured,and damage,respectively.Finally,using the IBk,the prediction accuracy was 82.41% with classification precision of 0.902,0.799,0.819,and 0.856 for fatal,hospitalized,injured,and damage,respectively.Therefore,IBk algorithms show the overall best agreement with the experimental dataset out of the five machine learning algorithms,for its global optimization and extrapolation ability.The results of the ensemble strategies show that the AdaBoost improved the performance metrics of MLP,PART,J48,and IBk.Bagging improved the performance metrics of MLP,PART,J48,and IBk.However,in the case of CART,neither AdaBoost or Bagging improved its performance metrics.Also,the performances of random forest and majority vote combiner strategies on the dataset are superior to the single classifiers.Out of four strategies of ensemble employed in this study,random forest outperformed all other techniques used for the improvement of the performance of individual classifiers.Peak hours,location type,riding under influence,age,gender,time of collision,and settlement type were found to be the critical determinants of motorcycle crash injury severity in Ghana.