Threshold value = 3 (fair condition) was specified for triggering maintenance interventions when gravel road subgrade exposure due to gravel loss is between 10 – 25%. Moreover, the developed FFNN gravel loss condition (GVL) prediction model yielded R2 = 0.95 > 0.9 benchmark based on minimum MSE = 0.055 < 0.1. FFNN produced better prediction accuracy than OLR model for categorical condition rating of gravel loss (Excellent – 1 Good – 2 Fair – 3 Poor – 4 Bad – 5). From the model evaluation results, Ordinal logistic regression (OLR) model was developed to predict categorical condition rating for gravel loss conditions but was limited and could only predict excellent condition rating.
This is to ensure the compatibility and quality of data sets and make greater use of available data for improved statistical reliability in developing performance prediction models. The study experience has shown that it is essential to select a common set of gravel road performance deterioration parameters and use consistent methods to collect and record condition data. The author also carried out field visual condition survey on 37km Molepolole - Lentsweletau (Road B123) of Kweneng district in 2012 for validation of the developed models. The input data for the models were generated from the triennial condition survey for Botswana carried out in 2002, 20. In addition, an improved district GIS-based map was finally developed using linear referencing approach to display gravel loss conditions as a threshold to trigger optimal maintenance interventions. Feed Forward Neural Networks (FFNN) trained with Levenberg – Marquardt (L-M) modelling technique was explored that is increasingly being applied to solving nonlinear complex problems like prediction, optimisation, clustering, pattern classification and function approximation as a parallel system that has the capability to learn, generalise and forecast pavement condition accurately. Ordinal logistic regression approach can transform a nonlinear relationship to linear type by a transfer function up to any desired degree of accuracy.įurthermore, improved artificial intelligent gravel road performance models which best capture the effects of gravel loss condition influencing factors were developed as extension of knowledge in existing gravel road condition models. As improvement over previous attempts to develop gravel loss condition forecasting models using multiple linear regression approach, this research modified the binary logistic regression model by incorporating the ordinal nature of a dependent variable through defining the probabilities differently to develop improved gravel road performance models based on categorical rating and site specific data. Till date in literature, a comprehensive model that could predict gravel loss (GVL) condition accurately has yet to be developed for district road networks. Optimal maintenance interventions at the appropriate time to preserve the asset value are required. Gravel roads are majorly affected by deterioration which manifests as loss of gravel materials due to traffic, climate and environmental conditions. 30 percent of the 18,300km Botswana Public Highway Networks constitute districts gravel road networks (5499.6 km) which are significant in providing access to rural areas where the majority of the population lives. Botswana roads largest assets estimated value as at 2010 was 15 billion Pula (2 billion USD). Roads are expensive asset and should be properly maintained regardless of their class or function to enhance their performance.
Therefore, prediction of gravel road conditions in the future is the basis for optimal maintenance intervention threshold to prevent further deterioration with time. Such information can be applied to transportation planning, decision making processes and identification of future maintenance interventions. Performance modelling of gravel road conditions is required in order to predict their conditions in the future and provide information on the manner in which they perform.