Deflection Basin Parameters can be calculated from the FWD collected deflection values, which are shown in Figure 1 . Several researchers showed that the deflection basin parameters are effective tools to identify the possible distressed layer of the asphalt pavement. Chang et al. [11] and Horak et al. [12] developed threshold values for categorizing each pavement layer condition. Table 1 and Table 2 show the threshold values of the FWD Deflection basin parameter. For example, the asphalt layer is in very good condition if the SCI values are <4 mils. On the other hand, the range of values 4 - 6 mils, 6 - 8 mils, 8 - 10 mils, and greater than 10 mils indicate good, fair, poor, and very poor asphalt layer condition, respectively [11] . Similarly, the base layer is in very good condition if the BCI values are <2 mils. On the other hand, the range of values 2 - 3 mils, 3 - 4 mils, 4 - 5 mils, and greater than 5 mils indicate good, fair, poor, and very poor base layer condition, respectively [12] .

Significant research works developed statistical and machine learning models for the predicting the pavement surface condition while few research works focused on the pavement structural condition prediction model. Table 3 shows the list of papers which developed pavement structural condition prediction models utilizing Falling Weight Deflectometer (FWD) data. After conducting the literature review, only six relevant research papers were found in terms of predicting pavement structural condition.

Note: ANN = Artificial Neural Network, ARIMA = Autoregressive integrated moving average time-series analysis, ADT = Annual Daily Traffic, BDI = Base Damage Index, BCI = Base Curvature Index, BLI = Base Layer Index, DNN = Deep Neural Network, D0 = Maximum Deflection, FWD = Falling Weight Deflectometer, MLI = Middle Layer Index, LTPP = Long Term Pavement Performance, LLI = Lower Layer Index, IRI = International Roughness Index, ESAL = Equivalent Single Axle Load, SNeff = Effective Structural Number.

From the above literature review, it was found that only 2 research papers used Long Term Pavement Performance (LTPP) data for pavement structural condition prediction. Haridas et al. [14] developed Deep Neural Network (DNN) to predict Maximum Deflection (D₀), Base Layer Index (BLI), Middle Layer Index (MLI), and Lower Layer Index (LLI). This research used IRI, Pavement Age, Traffic, Precipitation, Atmospheric Temperature, Pavement Temperature as the predictor variables. Sollazzo et al. [15] predicted Effective Structural Number (SN_eff) using Artificial Neural Network (ANN). This research used Asphalt Layer Thickness, Base Layer Thickness, Total Pavement Thickness, IRI at year 0, IRI at a specific year, ESAL, Average Temperature, and Standardized Temperature as the predictor variables. In addition, both research works used Artificial Neural Network (ANN), which is subject to overfitting problem. In this case, the prediction performance cannot be reliable even if it provides high accuracy. To improve the existing research in the field of pavement structural condition prediction, more novel methodologies are required to be applied, which are theoretically superior.

There are several methods for pavement structural condition data collection. Out of these, the three most popular devices for structural data collection are the Falling Weight Deflectometer (FWD), Traffic Speed Deflectometer (TSD), and Rolling Weight Deflectometer (RWD). Non-destructive tests (NDTs) are essential for evaluating the pavement structural condition evaluation. The advantages of nondestructive testing (NDT) compared to destructive testing methods include quicker test completion times, simpler operation, lower costs of operation, a smaller required workforce, less disruptive procedures, and the ability to conduct tests at a greater number of locations. The technology used in this field has progressed from older devices such as the Benkelman Beam, Dynaflect, and Road Rater to the modern impulse loading system known as the Falling Weight Deflectometer (FWD). The Falling Weight Deflectometer (FWD) is the most popular and commonly used non-destructive testing (NDT) device. The FWD applies dynamic loads to a pavement surface, simulating the magnitude and duration of a single heavy-moving wheel load. The FWD loading system delivers a transient impulse load to the pavement surface. The pavement response (vertical deformation or deflection) at various distances from the loading plate is measured by a series of geophone sensors (usually seven). The geophones/sensors are located at 0, 8, 12, 18, 24, 36, 48, 60, and 72-in. spacing from the center of the load plate to measure the deflections. The geophones are capable of measuring with an accuracy of up to ±0.01 mils, while the minimum resolution for measuring deflection is ±0.04 mils. Many State Departments of Transportation (DOTs) integrated the assessment obtained from the FWD device into the broader network-level maintenance activities. The advantage of using the FWD in pavement assessment lies in its precision and the depth of information it provides. By accurately simulating real-world loading conditions and measuring the pavement’s response, it offers an accurate estimation of the pavement’s structural condition. This allows for more targeted and effective maintenance strategies [11] . Figure 1 shows the diagram of the FWD testing device. After obtaining the deflection measurements from the FWD device, structural condition indicator parameters can be calculated using these measurement values. Table 4 shows the equation of the FWD deflection parameters. Here, D0 is the maximum deflection value, which the sensor collects at the point of load application. D300/D12 is the deflection value collected by the sensor located at 300 mm/12-inch inches from the load application point. The unit of deflections is expressed in mils.

Huynh et al. [4] surveyed the transportation agencies of the USA regarding the integration of pavement structural condition into the pavement management system decision making process. Twenty-five agencies responded to this survey. Out of the 25 responses, it was found that 15 agencies depend on the Falling Weight Deflectometer (FWD) device for pavement structural condition evaluation. Five agencies use both FWD and Traffic Speed Deflectometer (TSD) to collect pavement structural condition data. Another question was asked to the agencies whether they are interested in incorporating the pavement structural condition into the pavement management system decision making process. Among the responses, 13% of the agencies out of 25 responses agreed to incorporate pavement structural conditions into their pavement management system decisions. On the other hand, approximately 48% of the agencies out of 25 respondents are interested in integrating pavement structural conditions into the pavement management system decision-making.

The data preparation process for this study followed a rigorous and systematic approach to ensure both the quality and relevance of the data used in the machine learning models. The primary data source was the Long-Term Pavement Performance (LTPP) dataset, which is a highly comprehensive collection of pavement performance data gathered from across the United States, covering a time span from 1989 to 2018. This extensive dataset provided a rich source of information for pavement condition analysis and predictive modeling. Figure 2 shows a detailed framework for the Pavement Structural Condition Prediction Model Development. Each step in the flow chart is contained within a rectangular box, and the steps are connected with arrows indicating the sequential flow of the process. The essential steps of the framework are FWD Data collection, LTPP FWD Data Processing, Important Variables Selection, Data Preparation, Model Development, and Evaluation, respectively. Data Collection is the initial step to gather data for the model. LTPP FWD Data Processing step involves taking the FWD deflection values and preparing them for further model development use. After processing the data, important variables that will be included in the model are selected from the literature review. The data preparation step includes integrating the FWD data with the pavement rehabilitation, traffic, climate, and structural information data based on the LTPP segment number and FWD data collection year for each state. A Python script was developed to automate the data preparation task. With the data prepared, the next phase was to develop the model, which involved selecting an appropriate modeling technique, training the model on the data, and tuning the model parameters. After the model was developed, it was evaluated to determine its performance, which will involve assessing how well the model predicts or fits new data and making any necessary adjustments. After evaluation, the model will be finalized, which will consider any refinements needed from the evaluation phase.

The target variable selected for prediction in this study was the Surface Curvature Index (SCI). The SCI is calculated by subtracting the deflection value recorded by the sensor located at a 12-inch distance from the load application point from the maximum deflection value. Previous research [1] [5] has demonstrated that the Surface Curvature Index (SCI) serves as a highly effective parameter for assessing pavement structural conditions. Motivated by this finding, this research adopted SCI as the target variable for predicting pavement structural condition, building on the foundation of earlier studies that recognized its significance in pavement structural performance evaluation. To predict the target variable (SCI), several predictor variables were chosen. These variables represent a balanced combination of traffic loading, structural properties, and climatic factors. Specifically, the predictors included the surface curvature index at year 0, pavement age, surface thickness, base thickness, KESAL (equivalent single axle loads in thousands), annual precipitation, and freeze index. Each of these variables plays a crucial role in influencing pavement behavior and deterioration. The data for these variables were meticulously extracted from multiple tables within the LTPP database, ensuring consistency and precision in the data collection process. To enable effective model development and validation, the dataset was strategically split into training (80%) and testing (20%) sets. This division allowed for robust model training and unbiased performance evaluation. As part of the data preparation, comprehensive descriptive statistics were computed for each variable.

Table 5 presents descriptive statistics for the variables used in machine learning model development, highlighting key features such as the Surface Curvature Index (SCI), Pavement Age, Annual ESAL, and environmental factors like Precipitation and Freeze Index. These statistics provide insights into the distribution of the variables, with minimum, 25th percentile, median, mean, 75th percentile, and maximum values reported. The data suggest significant variability across the variables, which are critical for developing predictive models for pavement performance.

Figure 3 shows the relationships between various factors related to pavement conditions. Key observations include a strong positive correlation between Surface Curvature Index and Surface Curvature Index at year 0 (0.80), indicating that initial structural conditions highly influence current conditions. Surface layer thickness has a moderate negative correlation with Surface Curvature Index (−0.39) suggesting thicker surface layers may be associated with reduced structural condition scores.

Random Forest algorithm is a powerful ensemble machine learning technique renowned for its prediction accuracy and ability to handle complex datasets. It operates by constructing multiple decision trees, each constructed from a random subset of the training data. This subset is chosen using a method known as bootstrapping, which allows for the same data point to be used in multiple trees, leading to a reduction in variance and an increase in the model’s robustness [5] . These predictions are then combined to form the final output of the Random Forest model. Random Forest (RF) can be used for both classification and regression problem. For a classification problem, the result is achieved through majority voting, where the prediction that most trees agree on is chosen as the final decision. For regression tasks, the model averages the numerical predictions from each tree to determine the final model result. Following this approach, Random Forest (RF) model enhances the predictive accuracy by mitigating the errors of individual trees but also helps in preventing overfitting, which is a common problem of the machine learning model particularly Artificial Neural Network (ANN). In brief, the Random Forest algorithm provides a robust predictive performance [20] .

XGBoost is an ensemble technique based on decision trees. It combines multiple weak tree models into a strong learner by sequential training, optimizing an objective function at each step. The XGBoost model develops a prediction model by combining multiple weak learners. Each model learns from the previous model and builds a strong model by adjusting weights in a sequential manner. XGBoost has the capability to automatically handle missing data, which can be useful for infrastructure condition monitoring when there are significant missing values [21] . XGBoost is more robust to outliers and easier to tune hyperparameters than Artificial Neural Networks (ANNs) or Support Vector Machine (SVMs). XGBoost can be used for both regression and classification problems. The eXtreme Gradient Boosting (XGBoost) method is an improvement of the Gradient Boosting algorithm. The theoretical foundation of XGBoost can be expressed mathematically as follows [5] :

${\hat{y}}_i={\displaystyle \underset{k=1}{\overset{t}{\sum }}{f}_k\left(x_i\right)}$ (1)

where t is the set of regression trees, and f_k is a regression tree in the set. The main idea of the XGBoost algorithm is that each update is based on the prediction results of the previous model. By adding a new tree f_k to fit the residual error between the predicted value of the previous tree and the actual value, a new model is formed, and the new model is used as the basis for the next model learning. Mathematically, this process can be stated as follows:

${\hat{y}}_i={\hat{y}}_i^{\left(t-1\right)}+f_t\left(x_i\right)$ (2)

where ${\hat{y}}_i^{\left(t-1\right)}$ is the predicted value at time t − 1, and $f_t\left(x_i\right)$ is the residual fitting value by the newly added regression tree, with x_i being the input data. To obtain a prediction as close as possible to the true value of y_i, the following objective function is minimized by the XGBoost algorithm:

$ob{j}^{\left(t\right)}=l\left(y_i,{\hat{y}}_t^{\left(t\right)}\right)+{\Upsilon }^{{\rm T}}+\frac{1}{2}\lambda {\displaystyle \underset{j=1}{\overset{T}{\sum }}{w}_j^2}$ (3)