Predicting Pavement Condition Index Using an ML Approach for a Municipal Street Network

The study was carried out on the pavement network of Skellefteå municipality, which is in the northern part of Sweden with a population of 72,000 inhabitants. Skellefteå municipality is situated in climate zone 5, the coldest among Sweden’s five climate zones, characterized by its typically cold and prolonged winters and short summers (Trafikverket 2011). Furthermore, in this climate zone, the frost depth in the pavement normally exceeds 2 m, while the air freezing index ranges from approximately 550 to 2,200 degree days per year (Erlingsson and Saliko 2020). The municipality has a winter maintenance strategy where ploughing and sanding are the main maintenance activities/operations to enhance safety and adequate friction on the pavement surface, while salting is limited and is only applied to prevent ice formation on street pavement surfaces during transition periods in autumn and spring.

Pavement data used in the study (Fig. 1) were retrieved from Skellefteå municipality, which has a street pavement network of about 410 km in length. However, the retrieved data are limited since the pavement data are stored in different databases. Pavement surface condition assessments were conducted by the municipality in 2014 and 2018, using the windshield surveys approach at the network level. Pavement assessments were outsourced to the same contractor, and experts rated the distresses during postprocessing, using global positioning system (GPS)-coordinated images taken during the windshield survey. In essence, the assessment of pavement condition was entirely subjective, relying on the experiences of experts. The pavement network is divided into four street classes based on their functionality: main streets, collector streets, local/residential streets, and industrial streets. Similarly, the pavement network is divided into about 3,950 segments based on their condition, which means that the length of the segments is not constant (most segments are around 100 m in length). In the survey, the severity and density of six pavement surface distresses were collected to assess the pavement performance, which included surface unevenness $SU$ (longitudinal and transversal directions, excluding rutting); alligator cracking $ACR$; cracks $CR$ (longitudinal, transversal, frost, edge, joint reflection, etc.); rutting $RUT$; raveling $LAR$ (loss of coarse aggregates), including asphalt hardening; and miscellaneous surface defects $SD$, e.g., patching, potholes, edge deformation.

The Skellefteå municipality utilizes the PCI as the performance index, which ranges from 0 to 100. A PCI rating of 100 indicates excellent condition, whereas a rating of 0 reflects the worst surface condition of the pavement segments. In this context, experts utilize Table 1 to assess pavement condition, incorporating severity and density levels of various pavement distress types based on their experiences. This table was developed locally using guidelines from the pavement condition assessment handbook (Wågberg 2003). Similar handbooks have been developed both in the United States and in other European countries (Miller and Bellinger 2014; Ragnoli et al. 2018). According to Table 1, the severity levels indicate the degree of the distress type, while density levels indicate the extent of pavement surface affected, aligning with the corresponding difficulty levels, which serve as indicators of the level of complexity in addressing the distress type. In simpler terms, as the difficulty levels increase, addressing pavement distress becomes challenging, indicating a higher level of pavement deterioration. This is reflected through deducted points in the PCI rating computation of a pavement segment.

The pavement conditions are classified into four categories—green, yellow, red, and black—based on the PCI ratings, as described in Table 2. These categories also serve as indicators for the timing of maintenance activities. A similar approach has been adopted in the United States (CMAP 2024; Laton 2020; ODOT 2006). The municipality uses the following equations (Estholm 2019; Shahin 2005) to calculate the PCI rating or pavement condition of street segments:where $WDV$ is the weighted deduct value, $0.63636\times (x-10)-0.001715\times (x-10)\times (x-120)$, where $x$ is the total deduct value, which is the sum of all deduct values of distresses on the pavement section.

An example of pavement condition assessment for a street (Fig. 2) is presented in Fig. 3. The pavement data in Fig. 3 highlights the geometry, distress types, severity, and density of distress for this particular street segment, categorized under the black category, indicating the need for reconstruction.

The selection of pavement segments in the model data was based purely on the availability of maintenance history for the pavement segments in the municipality’s PMS database. Pavement maintenance activity data have been stored in the maintenance database since 2002. The maintenance history for most of the network includes the name of the street and functional class; the type and layer thickness of material used; the length and width of the pavement; traffic data; and the cost of maintenance measures. The quality of information stored in the database has been improved over the years.

In the data preprocessing, maintenance data held in the municipality’s PMS database for street pavements between 2002 and 2018 were retrieved. During the process, it became apparent that some street segments were inadequately documented, with issues such as inconsistent PCI ratings and missing traffic data. Consequently, nearly 22% of the street network was excluded from the study. The maintenance history of pavement segments includes maintenance year, maintenance treatment, and maintenance material. The maintenance data are further categorized concerning the cost of treatment. The PMS databases include pavement thickness data on most of the reconstructed pavements since 2010. Consequently, the majority of streets with a well-documented pavement history are nonresidential streets and are in the green category. The pavement age of residential streets that lack maintenance history was estimated based on the maintenance/rehabilitation history and PCI rating of the neighboring streets or neighborhoods, and the approximate year of construction of streets in the municipality. In this regard, the black category streets with no information about maintenance history were considered about 35–40 years old, the red category 25–30 years old, and the yellow category 15–20 years old.

The frequently applied maintenance treatments in the municipality are thin overlay (1–2.5 cm) of asphalt concrete and milling followed by overlaying with 2–3 cm of asphalt concrete (Anderson-Skold et al. 2022). Other maintenance treatments include crack sealing, patching, and reconstruction of streets.

The traffic database includes traffic volume, heavy traffic, and the speed limit on street segments. However, the municipality does not collect traffic data for all streets; as a result, some of the street segments were missing traffic data.

Climate data have not been included in the analysis since they does not vary significantly in the municipality, i.e., from one neighborhood to another as per the recent climate data.

After examining the pavement data, it was decided to use traffic data, surface distress data, and PCI rating for pavement assessment in 2014 in the models, together with the maintenance history up to 2018. In other words, the pavement condition observed in 2014 was used as a baseline to predict the PCI rating in 2018. The number of street segments is 3,024 with a total length of 320 km, as shown in Table 3. Moreover, it can be seen that street classes 1, 2, and 4 were included in one group (MCI streets or nonresidential streets) and class 3 in another group (residential streets). A comparative analysis of each model’s performance was carried out for both groups of streets. Table 3 shows that the street pavement data used in the models represent 75% of the total street network in length.

As described in Table 4, several variables have been utilized in the study. Age refers to the time since the last period of maintenance or rehabilitation that a street section has received, i.e., the age of the wearing surface (asphalt concrete). The criteria for assigning the deduct values for pavement distress during the pavement condition assessment are provided in Table 1. Moreover, traffic volume on MCI streets is separated into the variables light traffic and heavy traffic. All the variables are combined in several ways to analyze the performance of models for both groups of streets, as presented in Table 5.

In data preprocessing, the scaling of the original data was carried out before running the models. For the training RF model, two tuning parameters were specified: ntree and mtry. ntree represents a number of variables, which are randomly collected and sampled at each split time, whereas mtry is the number of branches that will grow after each time split. More specifically, different values were tested for both ntree and mtry, ranging from 250–1,000 for ntree, and 2–10 for mtry. For training neural networks, two hyperparameters were specified: size and decay. Size is the number of units in the hidden layer, while decay is the regularization parameter to avoid overfitting. More specifically, different sizes were tested, ranging from 1 to half the number of inputs, and decay from 0.1–0.5.

Note that for residential streets, traffic data were excluded from the models since traffic volumes are quite low. In contrast, models for MCI streets include a variable related to the amount of traffic data. Unfortunately, approximately 5% of the MCI streets are missing traffic data values. The missing traffic data were imputed by using the $k$-nearest neighbor (KNN) algorithm, a common method for addressing missing data issues (Murti et al. 2019; Sallaby and Azlan 2021). In this regard, the number of neighbors used in the imputation, $k$, was set as the closest integer to the square root of the number of observations (i.e., street segments).

Performance metrics are essential for evaluating the quality and effectiveness of predictive models. The models’ goodness of fit and predictive power were evaluated (Ali et al. 2023; Karballaeezadeh et al. 2020) using the following performance measures.

Mean absolute error (MAE) measures the average of the residuals:

Root-mean-square error (RMSE) measures the standard deviation of residuals:

The coefficient of determination ($R^2$) represents the proportion of the variance in the dependent variable, which is explained by the model:where $N$ is the number of street segments, $y$ is the street segment, $\widehat{y}$ is predicted value of $y$, and $\overline{y}$ is a mean value of $y$.

Note that low values for MAE and RMSE are good and imply higher accuracy of a model. However, high values close to 1 or negative 1 are good for $R^2$. RMSE gives larger penalization to high prediction errors, while MAE treats all errors the same.

All algorithms and comparisons were conducted using the R programming language. The caret2 (classification and regression training) package was utilized for algorithms, while the VIM3 (visualization and imputation of missing values) package was employed for the imputation of missing values. The seed for a random number generator was set to make the results reproducible by other researchers; specifically, the list of random numbers was started at position 123.

The effectiveness of the models was evaluated using a cross-validation technique. Therefore, the data were split into training and testing to avoid testing data that had already been used for training the models. Data were split into 10 folds and the process was repeated 10 times. In each cross-validation iteration, the data were randomly divided into 10 subsamples of approximately equal size whereby each subsample was used once as testing data and the remaining data were used for training. By repeating the process 10 times, 100 values of MAE, RMSE, and $R^2$ were obtained. Furthermore, during the model testing phase, the predicted and observed PCI were plotted and determined $R^2$ and Spearman’s rank correlation coefficients.

Finally, to explain the prediction of models, the importance (weighting) of variables was calculated by using different explainable ML techniques. In this regard, the absolute value of the $t$-statistic was used for each model parameter for LR models, Shapley additive explanations (SHAP) values for RF, and combinations of the absolute values of the weights for NN.

It can be seen in Table 6 that the RF model built on variables $A$ and $D$ performs better in the prediction of the status of street segments, followed by the NN, LR, and LR simple models. This order of performance is the same in MAE, RMSE, and $R^2$ evaluation matrices in the 10-fold cross-validation phase. By comparing the percentage difference of MAE values, the RF model is 6%–24% better, while in the case of RMSE values, the RF model is 6%–21% better than the NN, LR and LR simple models. Similarly, by comparing $R^2$ values, one can see that the RF model performance is roughly improved by 13%–43% compared to the rest of the models.

It can be observed in Fig. 4 that the order of the models according to their performance is the same in terms of MAE, RMSE, and $R^2$, and that the difference in performance is significant. In Fig. 5, predicted versus observed ratings of PCI were plotted to assess the correlation between the two variables. The correlation seems to be linearly positive in all models, which produced a range for $R^2$ from 0.30 to 0.59 in the model testing stage. Moreover, it is observed that predictions from the RF model are closer to a regressed diagonal line compared to those from the NN, LR, and LR simple models.

However, it is observed that predictions are close but do not necessarily follow a straight line; therefore, the relationship may be described by some monotonic function. To study this further, Spearman’s correlation was calculated to assess the pairwise degree of association between predicted and observed values. In Table 7, it can be seen that the Spearman’s correlation between predicted and observed values from RF is stronger than for the other models.

Furthermore, the influence of each input variable was calculated to determine the importance of variables in the output of the model. In this context, Table 8 highlights that variable $A$ holds the highest importance across all models, while the ranking order of the other variables varies between models. Notably, variable $SD$ has the lowest contribution for LR and RF, while it is ranked as the second-least important for NN.

From Table 9, it is evident that the RF model outperformed the NN, LR, and LR simple models in the 10-fold cross-validation phase. Moreover, it can be seen that the RF model built on $A+T$ variables performed best, with $\mathrm{MAE}=7.45$, $\mathrm{RMSE}=11.3$, and $R^2=0.50$. Subsequently, the RF model constructed with $A+D+T$ variables and $A+D$ variables followed in performance. The RF ($A+T$) model performed better with a range of 10%–55% and 1%–30% in the case of MAE and RMSE evaluations, respectively. Similarly, by comparing the $R^2$ evaluation values, it can be seen that the performance of the models decreases from 2% to 74% with respect to the RF ($A+T$) model.

Fig. 6 shows that the difference in performance is significant in terms of MAE, RMSE, and $R^2$ between those that perform well and those that perform badly, but the difference is not significant for those in the middle. Furthermore, it can also be seen that the order of the models by their performance is similar in terms of MAE and RMSE but is slightly different in terms of $R^2$.

To assess the model’s accuracy, predicted versus observed values were plotted during the model testing stage. From Fig. 7, it can be seen that different combinations of variables produced a wide range for $R^2$ from 0.14 to 0.79. Furthermore, it can be observed that predictions from RF built on $A+T$ variables and on $A+D+T$ variables are closer than for other models to a regressed diagonal line. However, it was noted that the predictions might be close but do not follow a straight line; the relationship, therefore, might be described by some monotonic function. The Spearman’s correlation coefficient was also calculated for this group of streets. Table 10 shows that the RF model predictions based on $A+T$ and $A+D+T$ variables have a stronger relationship with observed values compared to the other models. It can be seen that the values of Spearman’s coefficient indicate a monotonic relationship between predicted and observed PCI ratings across all models.

For each model, the influence of each input variable was calculated to determine the importance of each variable. Table 11 shows that $A$ is the most important variable for all the models built on $A+D$; however, the order of the other variables differs between the models. For example, $ACR$ is less important for both NN and RF but is quite important for LR.

However, Table 12 shows that for the RF built on the $A+T$, the $LT$ variable is more important than $A$. Furthermore, Table 13 highlights that the $A$ variable is again the most important one and $ACR$ is a less important variable for both NN and RF but not for $LR$. Variables $LT$ and $HT$ are quite important for RF but are less important for both LR and NN; the order of importance also changes.

The performance of models for residential streets varies when analyzing the model evaluation matrices using 10-fold cross-validation (Table 6), and $R^2$ (Fig. 5), Spearman’s coefficient value (Table 7), and the order of variable importance (Table 8) during the testing stage. RF models performed best throughout the aforementioned evaluations. RF models explain the data better, with an $R^2$ value of 0.5 using the 10-fold cross-validation. In other words, the RF model accounts for half the variance in the data, reflecting the complexity of the problem to the model’s ability to learn. Similarly, the best RMSE value is about 14 for RF, which is better compared to the other models. During the model testing stage, the highest $R^2$ value between the predicted and observed PCI ratings is 0.59, while the highest monotonic relationship, as indicated by Spearman’s coefficient, is 0.74.

High Spearman’s coefficient values indicate a strong positive monotonic relationship between the variables. The improvement in the Spearman’s coefficient increases with decreasing $R^2$ values, from LR simple to the RF model. This could be due to the lower sensitivity of Spearman’s coefficient to outliers. It could be argued that either the RF model might be useful in managing the outliers or it is less sensitive to outliers in predicting the future status.

In the case of residential streets, pavement age and pavement distress were utilized as input variables, while pavement age and maintenance history are highly important variables for predicting the PCI rating of streets. Similarly, pavement distress data are important in the timely maintenance of pavement. However, $SU$, $LAR$, and $CR$ are the most important variables followed by pavement age for residential streets in predicting the PCI rating. $SU$, $LAR$, and $CR$ are equally related to pavement age and cold climate. Climate is not one of the variables, but it has a great influence on the deterioration of pavements in the region. Regular collection of pavement distresses would help in improving the performance of the models. It may be interesting to use this model in densely populated municipalities by adding traffic information.

The performance of models for MCI streets, like for residential streets, varies in the model evaluation matrices using 10-fold cross-validation (Table 9 and Fig. 6), and $R^2$ (Fig. 7), Spearman’s coefficient value (Table 10), and the order of variable importance (Tables 11–13) in the testing stage. Thus, variables are utilized in 10 different ways to determine the aforementioned evaluation. The results show that the performance of models varies with the combination of input variables.

The pavement age and traffic ($A+T$) pair of variables performed well on MCI streets to forecast the pavement condition in the RF algorithm (Fig. 7). The difference in the performance of $A+T$ and $A+D+T$ is small, but it is interesting that by adding distresses as a third variable, the RF model accuracy decreases slightly. Conversely, the performance of the LR simple, LR, and NN models improves, though the RF model performs even better. Nevertheless, pavement age and traffic information are sufficient for predicting the PCI rating of MCI streets in the current study without nullifying the importance of regular assessment of pavement networks. Regular assessment of the street network would improve optimization of the models. RF models explain the variation in data better, with an $R^2$ value of 0.5 in the 10-fold cross validation. The best RMSE value is about 11, while the best correlation between the predicted and observed values is 0.79 for RF models with pavement age and traffic as input variables.

During the model testing stage, a high Spearman’s correlation was observed between the predicted and observed PCI ratings in all the models. The correlation between the predicted and observed values is strongly positive monotonic, particularly in the case of RF with $A+T$ and $A+D+T$, respectively. The improvement in the Spearman’s coefficient is similar to residential models, i.e., increasing with decreasing $R^2$ values. Similarly, the result of the RF model suggests that RF is better at managing the outliers due to splitting the data into several decision trees.

The order of importance of input variables varies among models (Tables 11–13). Pavement age is highly important in all the models, followed by $CR$ and traffic. Understandably, the more aged the pavement is, the higher the risk of $CR$ distress appearing on MCI streets due to climate and traffic. Interestingly, traffic is more significant than distress in RF models in predicting the PCI rating. This significance is in contrast to the rest of the MCI models.

The Skellefteå municipality employs the pavement condition category as a framework for organizing their annual maintenance activities. Tables 14 and 15 provide a comparative summary of model prediction accuracy across various street condition categories and overall accuracy for residential and nonresidential streets, respectively.

Notably, the residential RF model appears to be the most effective overall (54% accuracy), followed closely by the NN model (48%). Each model demonstrates varying strengths across different pavement condition categories. For instance, NN performs best in predicting the green category streets, LR simple has a notably high accuracy in the yellow category, and RF achieves the highest accuracy in predicting both the red and black categories. The choice of model for residential streets may depend on specific priorities, such as accurately predicting different pavement condition categories or achieving a high overall accuracy.

In the case of nonresidential streets, the RF ($A+T$) model achieves the highest overall accuracy at 84%, followed closely by RF ($A+D+T$). However, the accuracy diminishes with lower pavement condition categories, potentially tied to data imbalance and subjective pavement condition assessments. NN models show varying overall performance across variable combinations, suggesting sensitivity to the input variables. The choice of the best-performing model depends on the specific street condition category and the combination of input variables considered. RF, especially with $A$ and $T$ variables, tends to perform well across categories and is particularly strong in the green and yellow categories.

The variance in overall prediction accuracy between residential and nonresidential models is marginal, not exceeding 16% between the most and least accurate models. While the models do not achieve high accuracy in predicting PCI ratings, they possess sufficient capability to enhance maintenance approaches. Enhancements in model accuracy may be achievable with additional data and the incorporation of condition data obtained through automated surveys.

The RF models stand out as preferable for planning maintenance activities, but simple models could be effective in the case of limited budget and resources. Additionally, a combination of models could provide a diverse set of strategies for managing the street network. The models developed in the study are useful for optimizing maintenance schedules, improving the timing of maintenance treatments, and ultimately reducing both municipal maintenance costs and road user expenses in the long run.

For effective implementation, other municipalities would need a comprehensive pavement database and regular pavement assessments for PMS, which can be challenging due to resource constraints. Addressing these challenges would require investing in a data collection system, training personnel, and integrating new technologies for ongoing monitoring. Despite these hurdles, ML techniques can enhance resource efficiency, reduce road user costs, and improve the level of service by enabling better maintenance decisions and optimized resource allocation.