A Comparison of Connected-Vehicle Roughness and Traditional Pavement Condition Index

Abstract

Accurate, scalable pavement condition monitoring is essential for effective asset management, yet traditional methods of collecting metrics like the International Roughness Index (IRI), Pavement Condition Index (PCI), and Pavement Surface Evaluation and Rating (PASER) can be inefficient, expensive, and subjective. Recent efforts by Original Equipment Manufacturers have introduced crowdsourced approaches that estimate IRI at scale using connected vehicles (CVs). This study analyzes one month of CV-estimated IRI (IRI_CVe) data and compares it with manually collected PCI data from Marion County, Indiana, in 2024. The study includes four roadway classes: primary arterial, secondary arterial, primary collector, and local street, with 562, 147, 426, and 2402 centerline miles of data, respectively. IRI_CVe coverage was nearly complete for arterial and collector roads (93–100%) but was limited for local streets (37%). Threshold optimization revealed that the “needs maintenance” IRI category (IRI > 170 in/mi) correlates most strongly with PCI values below 50. The study found that 68%, 65%, 70%, and 59% of the roadway segments had PCI and IRI classifications in agreement. Spatial and categorical comparisons suggest some systematic biases between the metrics across roadway types, reflecting how they measure different dimensions of pavement condition. The results demonstrate near-term applications of IRI_CVe data for quality control in PCI-based asset management and support practical guidelines for integrating complementary pavement assessment metrics.

1. Introduction

An effective pavement asset management strategy can substantially reduce the long-term maintenance expenditures across a given roadway network [1]. However, data-driven investment and maintenance decisions are limited by the quality, coverage, and granularity of the dataset used, and traditional methods of data collection are often time- and cost-inefficient. For reference, in 2024, nearly $2.3 billion was budgeted for roadway funding in the state of Indiana alone, and $215 million was budgeted in Marion County [2]. Yet despite these investments, many agencies inventory pavement conditions no more than once annually, and they often have insufficient resources to survey the entirety of their roadway network.

In the last decade, the sensors on connected vehicles (CVs) have enabled a paradigm shift in monitoring traffic congestion. More recently, assets such as pavement markings and condition can now be assessed via onboard sensors. One of the datasets that has emerged is the ability to estimate the International Roughness Index (IRI) at network scale. However, as agencies begin to adopt this technology, they will need to know how to interpret the CV-estimated IRI (IRI_CVe) data with existing data sources like the Pavement Condition Index (PCI) to support their asset management strategy.

1.1. Problem Objective

Therefore, the objective of this study is to analyze the relationship between manually collected Pavement Condition Index (PCI) surface condition data and crowdsourced IRI_CVe measurements and compare these results with the existing literature. This comparison will be used to determine the place that IRI_CVe has in asset management relative to traditional methods. Additionally, IRI_CVe will be evaluated for its ease of integration into existing pavement management practices and for its utility as an independent quality control tool.

1.2. Literature Review

PCI is a widely used pavement assessment metric that measures the surface condition of the roadway and ranges from 0 (failed) to 100 (excellent). The American Society for Testing and Materials (ASTM) D6433 document outlines a standard method to determine the PCI of a given segment [3]. It is based on visual inspection of distress types and severities, and although traditionally computed manually, software systems such as PAVER can automatically compute the PCI score once the unit’s condition data are provided [4]. New methods of PCI data collection use computer vision methods such as convolutional neural networks (CNNs) to identify pavement distress in a more efficient and objective manner [5,6]. These methods attempt to mitigate the variance introduced by human rater biases [7,8]. The images are often still collected manually by surveys; however, a growing number of studies have begun exploring the use of crowdsourced dashcam imagery for scaling these information-extraction methods [9,10].

While PCI captures visible surface-level deterioration, it does not directly measure the ride quality like IRI. The International Roughness Index, or IRI, measures the vertical displacement of a vehicle over the longitudinal movement, usually with units of in/mi or m/km [11]. IRI is sensitive to vertical surface irregularities but may not capture some cracking and degradation that does not affect ride quality. It is most commonly collected with high-speed inertial profilers [12], but recent studies have measured IRI through alternative methods such as image processing [13] and Light Detection and Ranging (LiDAR) [14]. A newer approach uses crowdsourced CV data to derive estimated IRI in near-real time. This IRI_CVe data is estimated by combining multiple onboard sensors through a Kalman filter and aggregating CV readings geospatially and temporally. In contrast to surveyed IRI which is often collected annually or biennially, IRI_CVe measurements are provided daily and in smaller segments of ~75 ft. Mathew et al. leveraged this data source to analyze spatiotemporal trends on Indiana highways [15], though their study used coarse 0.1 mi and 1 mi route segments.

Validation studies done with a comparison of IRI and IRI_CVe data have found moderate-to-strong correlations with R² values of 0.35–0.79 [16,17,18,19]. However, it must be acknowledged that because IRI_CVe inherently exhibits complex structural interdependencies with various pavement distresses, its predictability via non-standardized methods is often limited. Consequently, crowdsourced data from non-standard equipment may require continuous validation and calibration against traditional benchmarks to account for these dependencies and ensure data reliability. Although IRI_CVe has advantages over traditionally collected IRI such as producing more robust values via aggregation over multiple crowdsourced measurements, it also has unique biases. For example, drivers may avoid wheel paths that are highly deteriorated, potentially making some measurements look better than what a profiler survey would produce.

Past studies have attempted to model the relationship between IRI and PCI. In Washington, DC, Arhin and Noel found very weak correlations (R² = 0.008–0.073) when stratifying by functional class and pavement type [20]. Other studies also found that the IRI-PCI relationship is weak overall. Pramesti et al. found that PCI cannot necessarily explain the IRI after analyzing the relationship over 100 km (62.13 miles) of roadway [21].

Stronger correlations were found with more sophisticated models. Hasibuan & Surbakti modeled the relationship using exponential regression, yielding an R² of 0.59 [22]. Park et al. also found an R² of 0.59 using a transformed power model on the Long-Term Pavement Performance (LTPP) dataset [23]. More recently, Piryonesi & El-Diraby analyzed the LTPP database and found a low aggregate correlation between PCI and IRI with an R² of 0.31, but a much stronger correlation was found for within-group data stratified by functional class, with some reaching up to R² > 0.7 [24]. It is worth mentioning that these studies using the LTPP database derived PCI from the available LTPP distress data, which may not align perfectly with the definitions of ASTM D6433. A dual model approach yielded an overall R² of 0.89 using additional variables such as traffic and weather [25]. Additional studies found similarly strong correlations, yet with limited scope and environment diversity [26,27,28]. Adeli et al. used an exponential regression model on rural road segments in Iran, yielding R² values of 0.75, 0.76, and 0.59 for the PCI categories of good, fair, and very poor, respectively. The study’s results included that the PCI-IRI relationship weakens as the pavement distress severity increases [29].

Even though some correlations were found between IRI and PCI, these past studies were limited and often done in a narrow scope. The majority of these studies found a moderate or weak relationship between the two metrics, which indicates that it is difficult to establish a direct relationship. Those that had stronger correlations tended to use more complex models and often made use of additional predictor variables such as traffic volume or weather patterns.

All comparative studies to date have relied on traditional data collection methods for IRI data, such as inertial profilers, rather than connected-vehicle-derived data. Additionally, a common limitation across prior studies is the lack of either spatial granularity or network scale. For example, Adeli et al. analyzed only 6000 segments, each 100 m (~0.06 miles) long, while Mathew et al. assessed thousands of miles of interstate highways using coarser segment lengths of 0.1 to 1 mile. Few have evaluated the relationship between IRI_CVe and manually collected PCI, particularly at the local level or across different roadway classes. This study addresses these gaps by leveraging high-resolution (~0.02 mi) CV data to compare IRI and PCI across a comprehensive local roadway network.

2. Materials and Methods

2.1. Study Location

This study analyzes the local road network in Marion County, Indiana, which includes the city of Indianapolis, the largest urban center in the state (Figure 1). The county has 8307 miles of roadway, the most of any county in Indiana [1]. Its size, diversity of roadway classes, and availability of both IRI and PCI data make it a suitable sample for evaluating opportunities to integrate IRI_CVe data into local agency pavement management practices.

2.2. US Roads Dataset

This study uses the US Roads dataset, which is publicly available free of charge on the Google BigQuery Marketplace [30]. The data was extracted from the US Census Bureau’s Topologically Integrated Geographic Encoding and Referencing (TIGER) database and contains 5189 miles of roadway in Marion County. This network was then segmented and buffered into a collection of roadway segments that serve as a consistent spatial reference, as will be described in greater detail in the Data Processing Section.

2.3. Connected-Vehicle Roughness Data

The IRI data in this study comes from the onboard sensors of a fleet of connected vehicles whose anonymized data was made commercially available through a third-party vendor. The data are provided as road segments 50 to 85 ft in length, with daily granularity. Each segment’s daily estimated IRI value is the result of a 60-day moving average and is measured in in/mi, which represents the vertical displacement of the vehicle in inches per mile of longitudinal movement. This long-term average is utilized to obtain a more statistically robust IRI estimate, particularly on low-volume routes such as local streets. On these roadways, lower CV trajectory counts can lead to higher estimation variance due to a smaller sample of vehicle trajectories. By aggregating multiple passes over a 60-day window, the methodology minimizes the impact of transient factors to provide a stable roughness metric.

The IRI estimation is done with a multi-sensor fusion approach via a Kalman filter which combines the measurements of onboard Original Equipment Manufacturer (OEM) sensors such as the Inertial Measurement Unit (IMU), Global Positioning System (GPS), and tire pressure sensors [31]. The output of this Kalman filter is used to estimate a longitudinal profile of the roadway, which is then used for roughness calculation, and these experienced roughness values are normalized per vehicle to produce IRI. The IRI measurements were categorized as good (<95 in/mi), acceptable (95–170 in/mi), and needs maintenance (>170 in/mi) using the thresholds defined by the Federal Highway Administration (FHWA) [32].

2.4. PCI Data

Approximately one third of the PCI data used in this study was manually collected each year between 2022 and 2024 and integrated into a common dataset for the Indianapolis network as of November 2024. The dataset represents the most accurate estimation of the Indianapolis road network condition in 2024 as reported by the City of Indianapolis, although not all measurements were collected in 2024. The routes were segmented by intersection, meaning every section of roadway between intersections was evaluated independently. No specific distress data used in the PCI calculation were recorded in the dataset, and only the PCI ratings were available. Each segment has a PCI rating from 0 to 100, and these ratings were later categorized as good (86–100), satisfactory (71–85), fair (56–70), poor (41–55), very poor (26–40), serious (11–25), and failed (0–10), as per the ASTM standard outlined in their D6433 document [3].

Additional attributes such as roadway class and surface type were made available by this same data source. The roadway classifications, shown in Figure 2c, correspond to the FHWA National Functional Classification system [33].

2.5. Data Processing

To produce consistent and uniformly distributed route segments for analysis, the US Roads dataset was pre-processed through several stages. First, many routes had duplicated or overlapping geometry, often due to road name changes. The deduplication was done through an algorithm that repeatedly removes all overlaps from the network, combines their metadata, and then re-adds them to the network a single time until there are no overlaps remaining. After deduplication, each route was buffered by 60 ft on each side and segmented into 0.02-mile segments following a linear-referencing methodology discussed in detail in [34]. A 60 ft buffer was chosen for coverage of potentially noisy GPS measurements, while still being small enough that capturing measurements from nearby roads is not an issue. Similarly, the 0.02-mile (105.6 ft) segment length was chosen for its high spatial resolution to capture the detail provided by IRI_CVe’s ~75 ft measurements on shorter local roads. These route segments serve as the primary spatial reference for IRI and PCI comparison.

Again following the methodology in [34], the IRI and PCI data sources were then joined with the route segments through geospatial mapping and aggregation. This includes the following processes. First, each IRI measurement was assigned to the overlapping route segment with the smallest ID, which was generated in the direction of travel during segmentation. Then the daily IRI measurements were aggregated by month with a median, and another median was taken at the segment level to obtain one representative sample per month per segment. Similarly, the PCI data was mapped to the route segments by using the PCI rating with the PCI segment that has the most overlap with the given buffered route segment. This approach differs from the IRI mapping approach due to the lower spatial resolution of the PCI data. Because the PCI data is recorded from intersection to intersection, the length is often more than one 0.02 mi segment and applying the IRI data mapping approach would leave gaps in the data availability. Therefore, by taking the maximum overlap, the longer PCI rating measurements can be assigned to multiple of these segments if it spans the majority of more than one segment’s length.

The existing literature makes clear that modeling and analysis should be stratified by surface type due to their distinct degradation characteristics [20].

Table 1. Miles of roadway with both IRI and PCI data by roadway class and surface type.

Roadway Class	Asphalt	Concrete	Total
Primary Arterial	516.1	1.8	548.1
Secondary Arterial	140.2	0.2	141.4
Primary Collector	384.8	0.6	387.4
Local Street	844.6	12.2	870.8
Total	1900.7	14.7	1963.6

shows a data summary indicating that the concrete surface type only has under 15 miles of data. Therefore, the concrete surface data were not used in this study. All analysis in the following sections uses only the ~1900 miles of asphalt data. In Table 1, the total column includes all surface types, including where this field is unknown, and therefore may not equal the sum of asphalt and concrete exactly.

For a consistent temporal comparison, the analysis used November 2024 IRI data which aligns with the most frequently recorded month in the 2024 year of PCI data. Approximately 65% of the 2024 Indianapolis PCI data was recorded in November, with the rest being recorded a few months earlier during the fall and late summer. The resulting IRI_CVe and PCI data for central Indianapolis can be visually compared in Figure 2a and Figure 2b, respectively. Of particular note, Figure 2a highlights how IRI_CVe data cover a broad cross-section of road classes including interstates, while the PCI data are entirely composed of the local network managed by the City of Indianapolis.

2.6. Threshold Optimization

Threshold optimization was conducted on the merged PCI and IRI dataset, using segments that have known values for both metrics. The objective was to identify the PCI threshold that most closely correlates with the segments categorized as “needs maintenance” by their IRI value. The optimal PCI threshold is intended to support descriptive agreement analysis within the study area and is not presented as a generalizable or predictive classification rule. Maximizing Youden’s statistic (True Positive Rate—False Positive Rate) yielded a PCI threshold of 50.01 (rounded to 50), meaning PCI values less than 50 best correlate with IRI values greater than 170 in/mi (the IRI “needs maintenance” category). Notably, this threshold lies at the midpoint of the PCI scale.

Although this threshold is mathematically optimal in Youden’s statistic, three commonly used agency condition thresholds were also tested: 40, 55, and 70 [3,35,36,37]. We found that the threshold of 55 yielded nearly the same the Matthews Correlation Coefficient (MCC) and Area Under the Curve (AUC) values as the PCI = 50 threshold, while the PCI = 40 and PCI = 70 thresholds resulted in a significantly degraded classification performance in these metrics.

2.7. Summary of Processed Data

Figure 3 and Figure 4 present the miles of roadway in each IRI category filtered and grouped based on four known roadway classes: primary arterial, secondary arterial, primary collector, and local street. Since the roadway class is sourced from the same dataset as the PCI values, these figures reflect only segments with known PCI data. Figure 3 shows that the IRI_CVe data achieves nearly perfect coverage for the three larger roadway classifications (at 99%, 100% and 93%), while local streets lag behind with 37% of the PCI data. This disparity is expected, as lower-traffic roads generate fewer connected-vehicle observations, whereas PCI is manually surveyed once annually regardless of traffic volume. In total, there are 1900.7 miles of asphalt roadway with both IRI_CVe and PCI data used in this study. This data was spatially mapped to 95,035 segments at a 0.02-mile segment length.

3. Results

3.1. PCI-IRI Relationship

To examine the relationship between PCI and IRI, scatterplots and boxplots were generated using route segments with values for both metrics. In Figure 5, each point represents one route segment’s plotted PCI and IRI values. The figure indicates a weak inverse relationship between PCI and IRI, with notable outliers in the upper-left and lower-right quadrants representing segments where the two metrics provide conflicting information. For example, a low PCI indicates a poor roadway condition and a low IRI indicates a smooth surface. Segments that exhibit both may warrant manual inspection to verify the data accuracy.

A large cluster of segments around PCI = 87 is visible in Figure 5, which may reflect category bias, where raters assign the minimum score qualifying for the “good” ASTM PCI category, which ranges from 86 to 100 PCI. If this cluster is truly the result of category bias, then it indicates a larger systematic problem with manually reported data: that it is subject to human error and inconsistency.

Figure 6 verifies findings from prior studies [29] that the PCI-IRI relationship weakens as the pavement condition gets worse. The boxplots show increasing interquartile ranges as PCI decreases, reflecting the larger IRI variation. In contrast, the “good”, “satisfactory”, and “fair” categories have tighter distributions, indicating a stronger agreement between the two metrics for higher PCI values.

The Cumulative Frequency Distributions (CFDs) of the IRI and PCI data are provided in Figure 7, where subplots (Figure 7a–d) compare the IRI and PCI for a given roadway class, and subplots (Figure 7e–f) compare the roadway classes within each metric.

In plots (Figure 7a–d), PCI CFDs are more linear than those of IRI, indicating that the PCI ratings are more evenly distributed over their range.

Meanwhile, CFDs (Figure 7e–f) show a consistent ranking across three of the roadway classes: secondary arterial, primary arterial, and primary collector. Notably, the only roadway class that changed order was local street, which ranked the lowest in IRI but not for PCI. This suggests a potential negative bias in IRI_CVe data for lower-volume local streets.

3.2. PCI-IRI Agreement Analysis

Figure 8 shows the resulting threshold of PCI = 50 overlaid on the segment scatterplot, highlighting the regions of agreement and disagreement between PCI and IRI. Each segment falls into one of four outcomes:

• True Positive (TP): both metrics indicate poor condition.
• True Negative (TN): both metrics indicate good condition.
• False Negative (FN): PCI suggests good condition, but IRI indicates a rough surface.
• False Positive (FP): IRI suggests a smooth surface, but PCI indicates poor condition.

Figure 8. Confusion matrix IRI and PCI regions. Callouts i, ii, and iii refer to outliers analyzed with Figures 12–14.

Figure 8. Confusion matrix IRI and PCI regions. Callouts i, ii, and iii refer to outliers analyzed with Figures 12–14.

Callouts i, ii, and iii in Figure 8 identify three segments with metric disagreement which were later examined with Google Street View imagery for quality control in Section 3.3.

Figure 9 groups these outcomes by roadway class to assess their accuracy. The measure of metric agreement is defined here as the count of TP + TN segments divided by the total number of segments for that class. Primary collector performed the best with 70% agreement, and local streets performed the worst with only 59% agreement. Two likely explanations for the relatively poor local street agreement are the higher prevalence of intentional surface features like manholes that impact IRI but do not affect the PCI rating or the lower CV trajectory count leading to IRI_CVe estimates with higher variance. While the presence of surface anomalies or a low trajectory count may contribute to lower agreement, a precise quantification of this effect was not feasible due to the pre-aggregated nature of the dataset and a lack of a comprehensive geospatial inventory of these surface anomalies.

Although a fixed PCI threshold of 50 yielded the best correspondence with IRI classifications, more expressive models could significantly improve predictive accuracy.

For a more precise analysis of the IRI–PCI relationship, the Matthews Correlation Coefficient (MCC), Area Under the Curve (AUC), and coefficient of determination (R²) values were computed across combinations of roadway class as shown in

Table 2. MCC, AUC, and R² correlation strengths by roadway class.

Roadway Class	MCC	AUC	R2
Primary Arterial	0.37	0.74	0.18
Secondary Arterial	0.33	0.75	0.21
Primary Collector	0.40	0.77	0.21
Local Street	0.21	0.69	0.15
Total	0.30	0.71	0.15

.

The MCC ranges from −1 (complete disagreement) to 1 (perfect agreement), with 0 indicating no correlation. It is calculated using the confusion matrix outcomes described earlier (TP, TN, FP, and FN) that were computed with the PCI threshold of 50 and IRI threshold of 170. The AUC, in comparison, measures the separability of the two classes using the IRI_CVe and ranges from 0.5 (inseparable) to 1.0 (perfectly separable).

The results show that the primary collector roadway class agrees most closely with IRI, which is also supported by the findings from Figure 8. The MCC and AUC values indicate that Local Streets in particular are the least separable using the threshold and had the worst classification performance.

R² values were computed as the squared Pearson correlation under a simple Ordinary Least Squares (OLS) model. These values similarly suggest a weak overall correlation between PCI and IRI, though correlations generally improved when stratified by roadway class. The highest observed R² was 0.21 for secondary arterial and primary collector roadways.

In Figure 10, each of the possible prediction outcomes for the primary arterial roadway class was plotted on the Marion County map to find geospatial trends. Maps a, b, c, and d correspond to regions iv, i, ii, and iii in Figure 8, respectively.

False negatives (Figure 10a) are concentrated in central and high-traffic urban corridors, suggesting that PCI ratings in these regions may overestimate pavement quality relative to what is provided by the IRI_CVe data. Conversely, false positives (Figure 10d) appear more frequently in peripheral and suburban areas, potentially reflecting segments with surface distress that does not have a noticeable impact on ride quality. For example, an urban arterial segment with patching may have both high IRI and high PCI (a false negative) due to poor ride quality despite good surface integrity. On the other hand, an arterial road outside the city might have severe rutting yet a smooth ride, yielding a low PCI and low IRI (a false positive).

Mapping agreement outcomes as in Figure 10 offers agencies an effective method of identifying areas with metric disagreement which may be flagged for manual inspection. This enables more accurate condition assessment and better-informed investment and maintenance decisions.

Histograms of the agreement outcomes (Figure 11) illustrate how the PCI threshold of 50 optimizes the trade-off between false positives and false negatives for the primary arterial roadway class. Increasing the threshold reduces the number of false positives, but greatly increases false negatives, and lowering the threshold would have the opposite effect.

Additionally, Figure 11a shows a large number of false negatives around PCI = 87, adding to the earlier hypothesis of PCI category bias. These segments were likely assigned inflated PCI ratings to meet the ASTM “good” threshold (PCI 86–100), despite some having poor ride quality. As a result, many of the segments were misclassified because of the disagreement between their IRI and PCI values.

3.3. Quality Control Application

When both PCI and IRI are available, segment-level quality control checks can be applied at scale. This is done by identifying segments with significant between-metric disagreement and flagging them for further evaluation or field inspection. For example, in Figure 8, callouts i, ii and iii identify three such road segments, which are corroborated with Google Street View imagery of their locations in Figure 12, Figure 13 and Figure 14, respectively. To facilitate Street View inspection at scale, the centroids of each outlier segment were calculated and formed into Google Maps links. The data were then combined into a single interactive map using Google My Maps to provide spatial context and improve accessibility. While the tool could support field visits if necessary, the Street View imagery was sufficiently up to date in this case, making in-person inspection unnecessary.

Figure 12 is an example of a false-negative asphalt segment located on Intech Blvd., which was rated a 66.7 PCI (“fair” category), but with an extremely high IRI of 650. Visual inspection using Google Street View found large transverse cracks across both lanes, which is contributing to the high IRI value. This discrepancy was identified for four consecutive segments over approximately 300 ft on this route. In this case, it is likely that the IRI_CVe value is overestimating the degradation of the roadway. Due to the crowdsourced nature of IRI_CVe data, these values are often more robust than manual collection methods; however, for low traffic segments, the number of measurements is lower, leading to a less predictable estimate with higher variance.

Figure 13 highlights another discrepancy between IRI and PCI on Ohio St. with a 79.7 PCI rating (“satisfactory” category), yet an IRI of 551.2 in/mi, which is far beyond the FHWA “needs maintenance” IRI threshold of 170. The corresponding Street View imagery shows severe distresses, including a large longitudinal midline crack. Again, this segment warrants re-evaluation to ensure that the PCI rating does not overlook important surface distress.

Conversely, Figure 14 shows an asphalt segment that was given a PCI of 8.9 (“failed” category), while its IRI measurement is within the “acceptable” range at 115.7 in/mi. The Street View image shows no sign of surface deterioration, implying that the PCI rating may have been too low. From inspection of the image, the street appears to have been recently paved. Therefore, the discrepancy is likely attributable to the PCI data being collected prior to the road being repaved in 2024.

These examples emphasize the value of using IRI_CVe data and third-party Street View imagery to perform quality control on manually rated PCI data. Although only three examples were shown in detail, Figure 8 shows that there are far more discrepancies that can be inspected.

4. Discussion and Conclusions

This study compared two widely used pavement assessment metrics, the International Roughness Index (IRI) and the Pavement Condition Index (PCI), to evaluate their agreement and respective biases using high-resolution connected-vehicle (CV) data in Marion County, Indiana. While traditional IRI measurements are typically limited to sparse, profiler-based surveys conducted on high-volume routes, this work utilized IRI_CVe at a granular segment resolution of 0.02 miles. By analyzing 1900.7 miles (95,035 segments) of asphalt roadway, this study represents one of the largest comparative assessments of CV-based roughness against manual agency records to date. The findings highlight a valuable use-case combining emerging crowdsourced technology and established pavement management practices. The weak overall correlation (R² = 0.15) observed here is consistent with previous findings by Arhin and Noel [20] and Piryonesi and El-Diraby [24], confirming that these metrics capture fundamentally different surface characteristics. IRI_CVe captures ride quality, whereas PCI focuses on observable distress and structural integrity. Additionally, the observation that correlation strength varies by roadway class (Table 2) mirrors the within-group stratification trends seen in the LTPP database [24]. Bridging the gap between IRI and PCI through high-resolution spatial analysis allows for a more comprehensive understanding of network health than either metric provides in isolation. Despite this weak linear correlation, the classification performance, quantified by the MCC and AUC values in Table 2, demonstrates that IRI_CVe effectively aligns with agency PCI thresholds for identifying maintenance needs. This supports the idea that the value of IRI_CVe lies in its utility as a network-level screening and classification tool rather than as a direct linear predictor of PCI ratings.

Furthermore, this study evaluates these metrics within the practical constraints of the agency asset management program. Using a 2024 agency inventory, which inherently includes temporal drift from rolling surveys, reflects the real-world conditions under which quality control must occur. To address the higher IRI_CVe estimation variance typically associated with low-volume routes, this study utilizes a 60-day long-term average to obtain a more statistically robust estimate for localized asset management decisions. This work demonstrates that the presented methodology is a viable tool for flagging potentially outdated or biased entries in existing agency databases, thereby increasing the reliability of pavement management systems that rely on these data.

• Classification performance was highest for secondary arterials and primary collectors, as evidenced by the MCC and AUC values (Table 2). While the overall linear correlation (R² = 0.15) was weak, this is consistent with how these metrics capture different dimensions of pavement degradation. The strength of the classification metrics confirms that IRI_CVe is a reliable indicator for categorical “needs maintenance” screening, even though it does not linearly track the full range of PCI values.
• Threshold optimization identified PCI ≤ 50 as best aligning with the IRI “needs maintenance” category (IRI > 170 in/mi). While this value maximizes the metric agreement within the Marion County dataset, it is intended as a calibration for localized data validation (Figure 9, Figure 10 and Figure 11) rather than a universal predictive rule, which would require diverse inter-agency datasets to minimize the risk of overfitting.
• Systematic biases were observed in both metrics. For instance, a high false-negative rate for PCI values around 87 indicated possible category bias in the dataset (Figure 11), and these segments can be re-evaluated according to the ASTM standards to correct for it if needed. Therefore, agencies can mitigate problems from biased data by combining both metrics: using IRI for large-scale screening and quality control and relying on PCI for diagnosis and treatment planning due to its ability to account for distress type and severity.
• Finally, and perhaps most promising for immediate implementation, by integrating crowdsourced IRI data into a common database that has traditional PCI data, an agency can very efficiently perform important quality control checks that complement, rather than replace, the use of the PCI data. Figure 12, Figure 13 and Figure 14 provide examples of a small subset of the significant number of discrepancies visible in Figure 8. These inconsistencies may be caused by factors such as systematic bias, human error, or data uncertainty, but all can be detected and addressed through the use of IRI_CVe data.

These findings support the complementary use of PCI and IRI_CVe in pavement asset management workflows. Future research incorporating IRI_CVe with vehicle trajectory counts and intentional surface feature locations could further characterize the variance of CV-based metrics on low-volume roads, providing an additional layer of statistical quality control to refine crowdsourced data integration into agency decision-making. Other future research directions include exploring similar comparisons with metrics like PASER, considering temporal trends to compare degradation rates, or scaling up the analysis with statewide data to compare agency-level variation in data quality.