Infrastructure asset management is the science and practice of driving the maximum value and levels of service from physical infrastructure systems (e.g. roads, bridges, water and wastewater). This is not limited to regular maintenance, however, and expands to assessing actual conditions, predicting future deterioration, estimating the risks of failure or reduced levels of service, considering the impacts on sustainability and using financial and life-cycle analysis tools to allocate budgets.

This field of practice is based on data-intensive analyses. Nonetheless, because of the limited investments and attention paid to this field, the available data is typically inadequate, and what is available is not reliable in most cases. This study therefore aims to explore and exploit advances in data science to address these two obstacles. First, can future deterioration be predicted through noncausal (circumstantial) attributes when more data is already available? In other words, can climate and generic roadway attributes (which are freely and reliably available) be used to predict future deterioration? Examples of such attributes are temperature ranges, freeze–thaw cycles, levels of precipitation, road functional class and traffic counts. Second, having linked deterioration to climatic attributes, this research examines whether the proposed approach could be used to study the impact of climate change on road infrastructure.

It is crucial to point out that the developed models are not causal. Instead of relying on a single equation or a causal relationship, machine learning algorithms predict outcomes by discovering patterns in data. In other words, what are the typical values of certain attributes that have been observed or co-occurred with historical levels of deterioration (not necessarily causing deterioration)? To this end, this study did not attempt to investigate the main causes of deterioration. Rather, in line with new data-driven decision-making practices, this study investigates which attributes increase the quality of pattern detection, without any claims on causation.

This study focused on analyzing performance indicators for asphalt pavements—particularly, pavement condition index (PCI) and the international roughness index (IRI). The data used was extracted from the Long-Term Pavement Performance (LTPP) database. The LTPP database is one of the largest, most reliable and well-maintained datasets in the domain. In analyzing LTPP data, the study examined the following machine learning algorithms:​ decision trees, logistic regression, naïve Bayes classifier, naïve Bayes classifier with kernel estimation, random forest, gradient boosted trees, k-nearest neighbours, linear regression and random forest regression.

The conducted analyses show that the following attributes can be used to predict asphalt pavement deterioration:​ initial condition, annual average of daily minimum (and maximum) air temperatures, number of freeze–thaw cycles, annual average daily traffic (AADT), annual precipitation, functional class of road, number of years since last remedial action, type of last remedial action, granular base equivalence (GBE) and pavement type. Among the most informative attributes were the initial condition, maintenance history (type and interval), traffic and GBE. The accuracy varies depending on algorithm, prediction interval, data size, data quality and the type of performance indicator. Therefore, the effect of each of these factors was studied one by one. Increasing the size of the training set increased the accuracy, and so did segmenting the data into smaller climatic regions. For instance, the mean accuracy of a gradient boosted trees algorithm for predicting the PCI in three years was 82 percent and reached 91 percent in a certain climatic region. The numbers for IRI prediction were 85 and 94 percent, respectively. An observation with regard to type of performance indicator was that simpler algorithms, such as linear regression or decision tree, had a higher performance when predicting the IRI.

Considering the range of variation in these predictive attributes, the impact of a sample climate change scenario was analyzed. The focus of the analysis was Ontario asphalt roads. However, a base (control) set of roads with a significant climate difference (in Texas) were also analyzed to put the Ontario results in perspective. In general, while precipitation is expected to increase, Ontario roads will have lower deterioration levels in the future, possibly due to the increase in temperature and the reduction in freeze–thaw cycles. Texas roads will experience a more aggressive deterioration.