The Raster Based Simulation of Pedestrian Injury Crashes using Artificial Neural Network in GIS (Case Study: Mashhad City)

This paper evaluates different MLPs (multi-layer perceptron neural networks) whilst considering 23 pedestrian injury accident parameters (in four categories) for time period of 3 years (1391-1393). The data belongs to the city of Mashhad in Iran. For this purpose, after preparing required raster layers in GIS, the influence of each parameter has been assessed through evaluation of different structures of MLP. It is worth noting that in the most works undertaken, the vector based modeling has been considered to determine the effect of spatial parameters; and the flexibility and the accuracy of raster-based prediction has been neglected. In the present study, authors have considered demographic characteristics and urban form factors in addition to the common variables in the simulation of pedestrian injury accidents. Furthermore, a total of 38 independent variables in four classes have been assessed. At first, the input raster data with the pixel size of 20 meters were created by employing geo_statistical processing in GIS and then MLPs algorithms have been programed using MATLAB language. A set of raster data containing 111,537 pixels has been analyzed using MLP changing the neurons, the layers, the transfer functions, the training functions and different combination of training and testing sets. Performance of 25 different neural network structures were validated by MSE and R-squared. The results indicate that a model with three-layer, 70 neurons, the RADBAS transfer function and Levenberg_Marquardt training function (= 0.95 and MSE = 0.0020) has the most efficient and optimum design for representing the relationship among the number of pedestrian injury accidents and the effective input parameters. By implementing this model to the study area of the city of, it was revealed that the residential density, the street network and the first level of social status characteristics have the highest impact on the pedestrian crashes. However, the sidewalks having more than 10 meters width, the fourth type of road and the second level of social status have the least impacts