Modeling retroreflectivity degradation of traffic signs using artificial neural networks

Traffic signs are vital for communicating guidance, rules, warning, and other highway agency information for safe and efficient navigation through transportation networks. Signs must be clearly detectable, readable, and understandable to fulfill their intended purpose. Poor sign visibility, particularly during nighttime, is the leading cause of fatalities worldwide. Sign retroreflectivity is one of the key measures to evaluate sign visibility conditions. It gradually deteriorates over time with sign aging, exposure, and other environmental conditions necessitating periodic sign maintenance or, ultimately, replacement when the sign retro values fall below the minimum prescribed standards. In literature, studies have mostly used traditional statistical regression models to model sign retroreflectivity as a function of available explanatory variables. Further, these studies have proposed separate retro degradation prediction models for different sign sheeting grades and colors that limit their applications for other scenarios. To fill the research gap, this study compared the performance of the linear regression method with three different architecture of the neural network namely, Feed-Forward Neural Network (FFNN), Cascade Forward Neuran Network (CFNN), and Elman neural networkNeural Network (ELMNN) for signs retro prediction with an aim to optimize sign maintenance and replacement activities and to enhance road safety. All the Neural Network models were employed with varying combinations of training algorithms, activation functions, and model parameters. Sign retro data for 539 in-service signs along selected sections of two expressways (M-1 and M-2) near the capital city of Islamabad in Pakistan were collected through portable handheld retroreflectometer GR3. Data on other sign attributes like sign ages (0, 2, 5, and 10 years), sign orientation, observation angle, sign sheeting brand, grade, and color were also acquired. Feature-based sensitivity analysis was conducted to identify the relative importance and ranking of input predictor variables. Model prediction results expressed in terms of various statistical evaluation metrics root mean square error (RMSE), mean absolute percent error (MAPE), RMSE-observation standard deviation satio (RSR), coefficient of determination (R²), Willmott's index of agreement (WIA), Nash-Sutcliffe efficiency (NSEC), and percent bias (PBIAS) showed that all the NN models outperformed the regression technique. Comparing the NN models, ELMNN architecture with 21 neurons in the hidden layer for ‘tansig’ activation function and ‘trainlm’ training algorithm yielded better retro-prediction performance. Feature sensitivity analysis revealed that variables sign age, sheeting brand, color, and observation angle were the most sensitive variables in predicting the retro output. Findings of this study can guide the transport agencies and decision-makers for effective policy implications and sign management practices.