Protocols for Network-Level Macrotexture Measurement (2021)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

64 This chapter discusses the analysis, results, and main findings of each of the three equipment comparison field experiments. All data from walking-speed and high-speed devices were treated with the adaptive outlier removal routine as described by Katicha et al. (2015) to use a consistent approach to outlier removal for all datasets. The false discovery rate method employed is a statistics-based approach that adapts to the given dataset to select an appropriate threshold that guarantees that no more than a certain percentage (10% in this case) of outliers are incorrectly identified. Due to the method’s statistical nature, large datasets should be used. Accordingly, consecutive rows of line- laser data were concatenated to create a dataset similar in size to those of the single-spot lasers for their outlier elimination. To minimize edge effects in the concatenation, a simple linear regression subtracted from the row of interest suppressed the slope and set the mean of the row to zero (see Equation 20):    [ ]            → . (20) 1,1 1,2 1, ,1 ,2 , 1,1 , h h h h h h h h n m m m n yields m n After outlier removal, the data streams were filtered to ensure only the wavelengths of interest for macrotexture were evaluated. A low-pass, lowest-order Butterworth infinite impulse response filter was designed for each device based on the sampling interval used according to the guidance in ASTM E1845. All long-wavelength trends (i.e., grade of the road, slope due to position of the device relative to the surface) of the 100 mm MPD base lengths were removed. This was accomplished by subtracting a first-order regression line from the profile of each base length, resulting in a zero-mean, zero-slope (detrended) base length. Finally, Mean Segment Depths (MSDs) were calculated for every 100 mm of longitudinal travel in the left wheelpath according to ASTM E1845. For the line lasers, MSDs values were calculated in the transverse direction and then arithmetic means were calculated to harmonize the reporting length of the line laser with the results of the single-spot lasers. For example, MSDs were calculated for each trans- verse line-laser scan, and all MSD values within the 100 mm of interest were averaged to compute the MPD. This means the data from more than 100 profiles from the line-laser devices were averaged for every MPD reported by the walking-speed device with the single- spot laser. C H A P T E R   4 Data Analysis

Data Analysis 65   4.1 Repeatability and Reproducibility The repeatability and reproducibility of most of the available devices were determined using the data collected in the first experiment at the Virginia Smart Road and verified with the data collected at MnROAD. 4.1.1 Methodology To overcome the limitations of other forms of device comparison (i.e., analysis of variance [ANOVA], correlation, and harmonization), the repeatability of each device and the agreement in measurement between device pairs (which also sheds light on an instrument’s bias and it is often referred to as reproducibility) were evaluated. Repeatability determines the extent to which a device can reproduce its previous results. In this study, this was quantified by calculating a device’s repeatability coefficient. A repeatability coefficient is superior to a coefficient of correlation, as the former provides a specific quantity (i.e., the difference in MPD measured in mm), whereas the latter only provides a proportion of the relationship (1.0 being a perfect relationship) between the repeat runs of the device. With the specific quantity provided by the repeatability coefficient, engineering judgment can be used to determine if the device has sufficient repeatability for the planned purpose. For example, a device with a repeatability coefficient of 0.09 mm may be deemed sufficient for measuring macrotexture, as this is within the resolution required to delineate between investigatory and intervention levels, which are often given in tenths of a millimeter. Conducting an ANOVA can show if means are equal for multiple runs of a device, where a high p-value corresponds to a failure to reject the null hypothesis, meaning that the means of repeated measurements taken are equal. However, this approach is limited when comparing devices one with another because a device with high variance in collected data tolerates a more substantial difference in means before rejecting the null hypothesis of equal means than a device with lower variance. Correlation analysis to compare devices can also be misleading, as it shows only the strength of the relationship between devices but does not quantify the differences (or, conversely, agreement) between devices. Harmonization studies have been used in the past to compare a device against a ground truth to force measurements to be more similar to another device. This is difficult for a pavement’s macrotexture, as the true value of macrotexture is unknown for the surfaces measured; in other words, there is no ground truth. Furthermore, these studies have been difficult to replicate under differing experimental conditions (Flintsch et al. 2009; Fuentes and Gunaratne 2010; Vos and Groenendijk 2009). The methods employed in this study to quantify repeatability and agreement are described in the next sections. Device Repeatability Repeatability determines the extent to which a device can reproduce its previous results. In this study, this was accomplished quantitatively by calculating a device’s repeatability coefficient. A repeatability coefficient is superior to a coefficient of correlation, as the former provides a specific quantity (i.e., the difference in MPD measured in mm), whereas the latter only provides a proportion of the relationship (1.0 being a perfect relationship) between the repeat runs of the device. With the specific quantity provided by the repeatability coeffi- cient, engineering judgment can be used to determine if the device has sufficient repeatability for the planned purpose. For example, a device with a repeatability coefficient of 0.05 mm may

66 Protocols for Network-Level Macrotexture Measurement be deemed sufficient for measuring macrotexture, as this is within the resolution required to delineate between investigatory and intervention levels, which is often given in tenths of a millimeter. The repeatability coefficient is derived from the device’s mean square error (MSE) from an ANOVA for several runs over the same pavement section, where MSE is essentially the variance of the device. Repeat measurements were made on the same pavement sections under identical conditions. In this study, this was achieved by measuring the same patch of pavement with static devices in the HVS and the same sections of the test track with walking-speed devices. From these measurements, the within-subject variance (MSE), or, standard deviation (SD), which is the square root of MSE, was calculated. Given the interest in the amount that measurements change between replicates, the differences between these SDs are of particular importance and can be calculated per Bartlett and Frost (2008) as shown in Equation 21: p= 2 . (21)SD of differences between measurements SD Given the expectation that 95% of the differences between measurements will be within two SDs of each other, Equation 21 then becomes the repeatability coefficient as defined by the British Standards Institution (British Standards Institution 1979): p p( ) =1.96 2 . (22)Repeatability coefficient c SDr As multiple runs were made with each walking-speed device, SD (the within-device standard deviation) was calculated by taking the square root of the MSE after performing an ANOVA. The interpretation of the test is that two measurements made on a section by a device should differ by no more than the repeatability coefficient 95% of the time, assuming a normal distri- bution of differences between measurements (Bartlett and Frost 2008). For the walking- and high-speed devices, the ANOVA was conducted for the devices with pavement section as the input and MPD as the model effect. For all the analyses, 1 m aggregated data collected over the pavement sections were averaged to a single average MPD for each pavement section for each of the three runs. Device Agreement Bland and Altman (1986) presented LOA as a superior device comparison method for cor- relation coefficients, as a strong correlation between devices does not necessarily guarantee strong agreement between them. This method explores the differences between measurements made by any two devices and quantifies these differences against the mean of the measure- ments made. Three runs of walking-speed data were collected over the 10 pavement sections as shown in Table 11. 1.96 , (23)LOA Sc= p where Sc = corrected standard deviation of differences = + +• • ,2 1 12 2 22S f S f SD SD = standard deviation of the difference between the mean of the runs for each road section by two devices compared, fi = 1 − 1 mi (m is the number of runs for each section), and S1 and S2 = variances of the devices (MSE determined from an ANOVA of the device with the pavement section as the model input and average MPD as the response).

Data Analysis 67   For the walking-speed devices, m = 3 runs per section; therefore, f1 = f2 = 0.67. For comparison of static devices, Sc is simplified to just SD. This is because only one measurement was taken at each measurement location along the pavement sections; therefore, the within-device variance cannot be included in the calculation. The boundaries for LOA are the mean of the differences for the devices over the sections ± the LOA. The LOA are plotted with the mean of the two devices’ measurements on the x-axis and the difference between the devices on the y-axis. 4.1.2 Static and Walking-Speed Equipment Data from static devices were treated using the methods implemented by the manufacturer’s software. All data from walking-speed devices were treated with the adaptive outlier removal routine as described by Katicha et al. (2015). The authors wanted to use a consistent approach to outlier removal for all walking-speed datasets. The false discovery rate method employed is a statistics-based approach that adapts to the given dataset to select an appropriate threshold that guarantees that no more than a certain percentage (10% in this case) of outliers are incor- rectly identified. Due to the method’s statistical nature, large datasets should be used. Accord- ingly, consecutive rows of line-laser data were concatenated to create a dataset similar in size to those of the single-spot lasers for their outlier elimination (see Equation 12). To minimize edge effects in the concatenation, a simple linear regression subtracted from the row of interest suppressed the slope and set the mean of the row to zero. Static Equipment Repeatability Data from static devices were treated using the methods implemented by the manufacturer’s software. The repeatability of two of the static devices was assessed under controlled condi- tions in the HVS chamber as previously discussed. The average output values for Device 12 were calculated by the device software. The various parameters were calculated individually for the 2,917 scan lines, and the arithmetic mean was made to represent the total pavement surface scanned. These average output values (one for each of the five replicates) of each parameter were used in the calculation of the repeatability coefficients given in Table 20. These coefficients represent the maximum difference that can be expected from the device on the same pavement sample 95% of the time. For example, if an average MPD of 0.80 mm is measured by the device for the pavement, Parameter cr MPD 0.0021 mm ETD 0.0017 mm RMS 0.0023 mm Ra 0.0021 mm Rq 0.0027 mm Rsk 0.0292 Rku 0.2042 Length 1.8909 mm Length Ratio 0.0182 Table 20. Repeatability coefficient of mean results output by Device 12.

68 Protocols for Network-Level Macrotexture Measurement 95% of measurements will be within the range of 0.798 to 0.802 mm based on these results. Average values represent how a road agency may gather and summarize data for a network of roads in their jurisdiction. Furthermore, this device allows greater granularity than the rest. An analysis of the full range of data available for all parameters of the five replicates was made. An example distribution of results is shown in Figure 40. Statistics for the remaining parameters are summarized in Table 21. Repeatability coefficients were then calculated for each of the five replicates across all scan lines. For example, a cr was calculated using the methodology explained above for the within-subject variability of the five MPD values calculated for scan line one. This was repeated for scan line two through scan line 2,917. The results of the calculations are summa rized in Table 22. This is useful when considering using fewer scan lines to represent a pavement sample. In this case, the averaging effect (which decreases variability) of increased scan lines is reduced. Device 6 has a circular measurement path made with a single-spot laser. The device was tested for repeatability in three different locations within the HVS. The first location was in the Figure 40. Histogram and boxplot of all RMS values gathered for the subject pavement sample (in mm). MPD ETD RMS Ra Rq Rsk Rku Length Length Ratio Mean 0.7969 0.8375 0.8317 0.6324 0.8544 -1.9634 4.8957 178.99 1.7209 SD 0.2165 0.1732 0.1760 0.1531 0.1858 0.6444 4.3443 25.27 0.2430 Max 1.5489 1.4391 1.2914 1.1007 1.3504 -0.3437 34.0445 330.13 3.1741 Q75 0.9202 0.9362 0.9559 0.7412 0.9820 -1.5045 6.7330 192.30 1.8489 Q50 0.8089 0.8471 0.8499 0.6599 0.8781 -1.892 3.7577 174.34 1.6762 Q25 0.6342 0.7073 0.6989 0.5083 0.7248 -2.3332 2.0146 160.93 1.5473 Min 0.3903 0.5122 0.4156 0.2917 0.4366 -4.9160 -1.1910 134.98 1.2978 Table 21. Summary statistics of parameters collected for five replicates.

Data Analysis 69   middle of the traffic pattern for the test wheel of the HVS. This location was most similar to the conditions in which the device would be used in the field. Five tests were conducted with Device 6, each time pressing the measurement button without moving the device. Results of the repeatability test for MPD and RMS (the two parameters output by the device) are presented in Table 23. Testing at the second location was accomplished by using the same marked location but picking up and replacing the device between measurements. This simulates the practice of repeat measurements over time. Maximum care was taken to place the device in the same location according to marks made on the pavement. However, the results showed that the repeatability of the device decreases as these small changes in measurement location occur. The third test location was in an area of the HVS that receives no traffic by the test wheel. This can be considered virgin pavement and demonstrates the small difference between virgin and trafficked pavement when compared to the difference between repeat tests when the device is and is not repositioned each time. Walking-Speed Equipment Repeatability The calculated coefficients of repeatability for the walking-speed devices are found in Table 24. The cr for all devices tested fell into a similar range (from 0.025 to 0.054 mm), meaning that on 95% of occasions the measurements on pavement types similar to those tested by the devices used will differ by no more than the repeatability coefficient. The coefficients shown in the table are for all pavement surface types tested. Static Equipment Agreement For the static devices, the pavement tested was the same as the pavement tested for the walking-speed devices. This was done to provide a variety of test surfaces to understand the agreement between devices on different types of pavement. Five locations were identified along the length of the section and marked by a 400-mm × 400-mm box in which the static measurement devices were placed. The results of the LOA analysis are found in Table 25. cr of MPD cr of ETD cr of RMS cr of Ra cr of Rq cr of Rsk cr of Rku cr of Length cr of Length Ratio Mean 0.0166 0.0133 0.0264 0.0127 0.0262 0.1516 1.123 14.01 0.1347 SD 0.0185 0.0148 0.0325 0.0125 0.0313 0.2226 2.007 12.57 0.1209 Max 0.2250 0.1800 0.2232 0.1119 0.2166 1.9568 24.305 95.47 0.9179 Q75 0.0197 0.0158 0.0335 0.0157 0.0330 0.1700 1.184 18.28 0.1758 Q50 0.0109 0.0088 0.0143 0.0086 0.0147 0.0714 0.422 9.97 0.0959 Q25 0.0067 0.0054 0.0066 0.0049 0.0070 0.0326 0.158 5.43 0.0522 Min 0.0011 0.0009 0.0005 0.0004 0.0006 0.0027 0.008 0.60 0.0058 Table 22. Summary of cr for each scan line of five replicates. Test Location cr of MPD Cr of RMS Traffic 0.05 0.17 Traffic: Device Replaced 0.14 0.20 Non-Traffic 0.02 0.15 Table 23. Summary of cr for Device 6 (mm).

70 Protocols for Network-Level Macrotexture Measurement Note that Device 12 (a single-spot device) was rotated 90° on some surfaces (denoted by a “T”) with directional texturing to demonstrate the effect of said texturing on parameter calculation. For example, the agreement between Device 12 and Device 8 improved for the longitudinal sections when Device 12 was placed with the scan lines traveling longitudinally along the pavement, positioning it to capture the same pavement profile as Device 8 (a static line-laser device) for each of its scan lines. Similarly, the agreement between Device 6 (a device that scans in a circular pattern) and Device 12 improved on the longitudinally textured pave- ment when Device 12 was set transversely (with the scanning line going across the grooves of the pavement texture), which was more similar to several of the arcs measured by Device 6. This effect is most easily seen when the means of each device are plotted against the differ- ence between the means of each device (see Figure 41). In the figure, notice the large band of agreement for Device 6 and Device 12 (Figure 41a) and the much narrower band of agreement for Device 6 and Device 12T (Figure 41b). The analysis was initially carried out for all pavement types tested; however, it was noted that the directional texturing of many of the PCC pavements would affect the macrotexture parameters calculated. Hence, the pavements were divided into four groups, as seen in Figure 42. As seen in the figure, random texturing covers all AC pavements and PCC1b and PCC1d, which were PCC pavements with a surface treatment with relatively small aggregate. The longitudinal sections comprised PCC pavements with longitudinal grooves and/or diamond grinding (PCC2 and PCC1f). The transverse pavements were those with transverse tines (SRB and PCC1a). Walking-Speed Equipment Agreement For the walking-speed devices, the mean of all 1 m MPDs for each run of a particular pave- ment section was calculated. The results are plotted in Figure 42 for each of the devices tested. The LOAs for each of the walking-speed devices were calculated as previously described. A summary of these results for the four pavement surface texture groups is found in Table 26. In most cases, the agreement between the two line-laser scanners (Device 7 and Device 11) Device 7 Device 9 Device 11 MSE 8.2 E-5 9.4 E-5 3.8 E-4 cr 0.025 0.027 0.054 Table 24. Summary of cr for the three tested walking-speed devices (mm). All Random Longitudinal Transverse Device Pair Mean of Difference LOA Mean of LOA LOA Mean of Difference LOA 6, 8 0.094 0.935 -0.111 0.374 0.926 0.709 -0.120 0.339 12, 8 -0.068 0.326 -0.067 0.333 -0.047 0.377 -0.096 0.274 12T, 8 0.135 0.922 -- -- 0.969 0.692 -- -- 6, 12 0.163 0.875 -0.045 0.217 0.972 0.706 -0.024 0.142 6, 12T -0.040 0.207 -- -- -0.043 0.245 -- -- Difference Mean of Difference Table 25. Summary of LOA analysis for static devices by type of texture (mm).

Data Analysis 71   (a) LOA, Device 6 and Device 12 (longitudinal orientation) (b) LOA, Device 6 and Device 12 (transverse orientation) Figure 41. Comparing LOA of Device 6 to LOA of Device 12, oriented in the longitudinal (a) and transverse (b) directions. was closer than the agreement between the line-laser scanners and the single-spot device. In most cases, however, the mean of the differences was wide enough to indicate that the measure- ments from the devices should not be used interchangeably for the surfaces tested. 4.1.3 High-Speed Equipment High-Speed Equipment Repeatability All data sets received the same outlier removal, detrending, and filtering treatments before the data were analyzed. Table 27 shows the calculated coefficients of repeatability for the data analyzed according to the procedures described in the previous sections. The coefficient of

72 Protocols for Network-Level Macrotexture Measurement Figure 42. Mean MPDs for walking-speed devices, calculated for each pavement section tested. All Random Longitudinal Transverse Device Pair Mean of Difference LOA Mean of Difference LOA Mean of Difference LOA Mean of Difference LOA 9, 11 -0.23 1.17 -0.12 0.20 -1.19 1.20 0.39 0.61 9, 7 -0.06 1.06 0.02 0.06 -0.88 1.25 0.50 0.88 11, 7 0.17 0.25 0.14 0.22 0.31 0.07 0.11 0.27 Table 26. Summary of LOA analysis for walking-speed devices by type of texture (mm). Device 1 Device 2 Device 3 Device 4 Device 5 Device 15 MSE 7.3 E-4 6.7 E-4 10.07 E-4 5.27 E-4 5.2 E-4 9.38 E-4 cr 0.075 0.072 0.088 0.064 0.063 0.085 Table 27. Summary of device repeatability (mm).

Data Analysis 73   repeatability for all devices tested fell within a similar range (from 0.063 mm to 0.088 mm), meaning that measurements on pavement types similar to those tested by the devices used will differ by no more than the repeatability coefficient (i.e., 0.063 mm) on 95% of occasions. This is thought to be sufficient given that the repeatability of sand patch tests has been found to vary by approximately 24% of the mean value of each test location in another study (Doty 1975), indicating that these high-speed, non-contact means are an improvement to manual means. High-Speed Equipment Agreement The mean of all 1 m MPDs for each run of a particular pavement section was calculated. The results are plotted in Figure 43 for each of the devices tested. The figure clearly shows that the results were similar for most of the surfaces tested. For example, PCC1e had relatively Figure 43. Mean MPD results for each section tested.

74 Protocols for Network-Level Macrotexture Measurement low MPDs, and Asphalt Section K (an open-graded friction course) had relatively high MPDs reported by all devices. It is also apparent in PCC1f (a longitudinally ground and grooved PCC section) that one device yielded much higher results than the other four devices tested for each of its runs. PCC2 showed a similar but less extreme trend. The next task was to quantify the agreement (or lack thereof) between devices. The LOAs were initially calculated for pairs of all devices and all sections of pavement. However, as it happened with the walking devices, the agreement of all devices against the transversely mounted line-laser scanner (Device 5) was poor compared to the pairwise com- parisons of the other devices. The difference in MPD for Device 5 is seen in Figure 43. Notably, all five runs are plotted for each device, but they are typically too close to distinguish at this scale as quantified by their repeatability coefficients in the previous sections. The plots in Figure 43 show differences between the longitudinally mounted texture surfaces, and the plots in Figure 44 show how much the single-spot lasers disagreed with the line-laser system. The narrow band of agreement of two single-spot laser devices (Figure 44a) and the wider range for the single-spot laser compared to the line-laser scanner (Figure 44b) are notable, but this does not mean that the readings made by either of the devices were necessarily poor; rather, the results demonstrate that the devices are not interchangeable for certain pavement types. The pavement surfaces were again grouped based on the orientation of the texture, as shown in Table 28. The “Random” texture (asphalt-like sections) columns show better LOA for the line-laser equipment (Device 5) with all other devices when compared to the LOA for all pave- ment surface types. The LOA dropped (improved) by an average of 74%. Because it was apparent that the engineered directional textures (i.e., grooves and tine marks) created poor agreement between the single-spot and line-laser devices, LOA analysis was then carried out on the PCC sections, first those with longitudinal texture (PCC2 and PCC1f, labeled “Longitudinal Only” in Table 28), and next those with transverse grooves or tine marks (the remaining PCC-surfaced sections, labeled “Transverse Only” in Table 28). As expected, the line-laser results again showed poor agreement with all devices for the “Longitudinal Only” sections (the line-laser devices can capture this texture and the other devices cannot); however, it is interesting to note that the best agreement between all devices was for the PCC sections with transverse texturing, with LOA ranging from 0.05 mm to 0.14 mm. Theoretically, the results from a transverse-mounted line-laser scanner should fall either into a peak or into a valley for each reading. The close agreement indicates that the line laser does not always run parallel to the transverse groove or tine and, therefore, captures peak and valley information similar to the single-spot laser systems. Table 28 again confirms that the two technologies (single-spot and line lasers) cannot be used interchangeably for all pavement types. Table 29 highlights the difference in means between any two devices to show where the greatest agreements and disagreements were for each pavement section. Device 3 was also found to have a consistent bias of approximately 0.1–0.2 mm (see the “Mean of Difference” column in Table 28) when compared to all other devices. The com- parison of measurements conducted at different speeds shed some light on the possible cause. Figure 45(a) shows that Device 3 produces measurements that seem to be dependent on the measurement speed. In reviewing the data for Device 2 and Device 3 (both equipped with 32 kHz sensors), the raw profiles captured for Device 3 were smoother with less aggressive peaks than those of Device 2, shown in Figure 45(b). This was before the data were processed with any software filtering. This smoother signal is thought to account for the difference in MPD between these devices.

Data Analysis 75   (a) Device Pair (Device 1 and Device 2) (b) Device Pair (Device 1 and Device 5) Figure 44. Bland-Altman plots for all sections tested for device pairs (a) Device 1 and Device 2 and (b) Device 1 and Device 5.

76 Protocols for Network-Level Macrotexture Measurement Texture Type All Random * Longitudinal Only * Transverse Only * Device Pair Mean of Difference LOA Mean of Difference LOA Mean of Difference LOA Mean of Difference LOA 1, 2 0.02 0.10 0.01 0.12 0.02 0.05 0.02 0.08 1, 3 0.13 0.19 0.20 0.12 0.07 0.08 0.06 0.14 1, 4 0.01 0.15 0.03 0.18 -0.07 0.07 0.02 0.08 1, 5 0.00 0.68 0.04 0.17 -0.82 1.29 0.15 0.10 1, 15 -0.03 0.18 0.02 0.12 -0.21 0.25 -0.04 0.04 2, 3 0.11 0.18 0.19 0.10 0.05 0.08 0.04 0.09 2, 4 0.00 0.16 0.02 0.20 -0.09 0.06 0.00 0.05 2, 5 -0.03 0.67 0.03 0.12 -0.84 1.27 0.13 0.06 2, 15 -0.04 0.18 0.01 0.10 -0.23 0.24 -0.06 0.09 3, 4 -0.12 0.21 -0.17 0.23 -0.14 0.06 -0.04 0.08 3, 5 -0.14 0.66 -0.16 0.14 -0.89 1.22 0.09 0.06 3, 15 -0.15 0.17 -0.18 0.11 -0.27 0.20 -0.09 0.15 4, 5 -0.03 0.63 0.01 0.25 -0.75 1.24 0.13 0.06 4, 15 -0.04 0.19 -0.01 0.23 -0.13 0.21 -0.06 0.09 5, 15 -0.01 0.55 -0.02 0.13 0.62 1.06 -0.18 0.11 * Comparisons with LOA > 0.15 mm are highlighted in bold. Table 28. Summary of LOAs (mm). Sections Device Pairs (in mm) 1, 15 2, 15 3, 15 4, 15 5, 15 1, 2 1, 3 1, 4 1, 5 2, 3 2, 4 2, 5 3, 4 3, 5 4, 5 SRB -0.02 -0.07 -0.17 -0.08 -0.23 0.06 0.15 0.06 0.21 0.09 0.00 0.15 -0.09 0.06 0.15 PCC2 -0.12 -0.15 -0.21 -0.07 0.24 0.03 0.09 -0.06 -0.36 0.06 -0.08 -0.39 -0.15 -0.45 -0.30 RRB -0.06 -0.15 -0.22 -0.14 -0.29 0.09 0.16 0.07 0.23 0.07 -0.01 0.14 -0.08 0.07 0.15 PCC1g -0.05 -0.05 -0.06 -0.06 -0.14 0.01 0.01 0.01 0.10 0.01 0.00 0.09 0.00 0.08 0.08 PCC1f -0.29 -0.30 -0.33 -0.20 1.00 0.01 0.04 -0.09 -1.29 0.03 -0.10 -1.30 -0.13 -1.33 -1.20 PCC1e -0.03 -0.05 -0.01 -0.02 -0.14 0.02 -0.02 -0.01 0.12 -0.04 -0.03 0.10 0.01 0.13 0.12 PCC1d -0.02 0.02 -0.22 0.05 0.01 -0.04 0.20 -0.07 -0.03 0.24 -0.03 0.00 -0.27 -0.24 0.04 PCC1c -0.03 -0.02 -0.05 -0.01 -0.17 -0.01 0.02 -0.02 0.14 0.02 -0.01 0.15 -0.04 0.12 0.16 PCC1b 0.11 0.01 -0.18 0.21 -0.10 0.10 0.29 -0.09 0.21 0.19 -0.20 0.11 -0.38 -0.08 0.31 PCC1a -0.04 -0.02 -0.04 -0.02 -0.13 -0.02 0.01 -0.02 0.10 0.03 0.01 0.12 -0.02 0.09 0.11 L2 0.05 0.05 -0.14 0.01 0.00 0.00 0.19 0.04 0.05 0.19 0.04 0.05 -0.15 -0.14 0.01 K 0.06 0.04 -0.14 0.05 0.07 0.02 0.21 0.01 -0.01 0.19 -0.01 -0.03 -0.20 -0.22 -0.02 J -0.01 -0.05 -0.22 -0.20 -0.05 0.04 0.21 0.19 0.04 0.18 0.16 0.01 -0.02 -0.17 -0.15 I 0.01 0.02 -0.16 -0.05 -0.06 -0.01 0.17 0.06 0.06 0.18 0.07 0.07 -0.11 -0.10 0.00 H 0.00 0.00 -0.16 -0.06 -0.03 0.00 0.16 0.06 0.03 0.16 0.06 0.03 -0.10 -0.13 -0.03 HWB -0.03 -0.05 -0.12 -0.08 -0.19 0.01 0.08 0.04 0.16 0.07 0.03 0.14 -0.04 0.07 0.11 D2 0.00 0.01 -0.17 -0.03 -0.03 0.00 0.17 0.03 0.03 0.18 0.04 0.03 -0.14 -0.14 0.00 C 0.00 0.01 -0.18 -0.03 0.03 -0.01 0.18 0.03 -0.04 0.19 0.04 -0.03 -0.15 -0.22 -0.06 Table 29. Summary of differences of mean MPD values between device pairs.

Data Analysis 77   (a) (b) L2 PC C 1a Section 10 20 30 40 50 60 70 AV G M PD Y = 1.205 - 0.002248*X 0.6 0.7 0.8 0.9 1.0 1.1 1.2 0.6 0.7 0.8 0.9 1.0 1.1 1.2 Y = 0.6768 - 0.00082*X H ei gh t ( m m ) H ei gh t ( m m ) Measurement number - Device 2 Measurement number - Device 3 Figure 45. Results from the constant-speed experiment showing (a) MPDs for all runs of Device 3 and (b) raw data profiles of two single-spot lasers (Device 2 and Device 3).

78 Protocols for Network-Level Macrotexture Measurement One concern with performing network-level macrotexture data collection is that uncontrolled factors (e.g., vehicle wander and a driver’s ability to reproduce a vehicle’s path) will distort the data. Recall that Device 15 was mounted to the same vehicle and directly in-line with Device 1. The LOA for Device 1 and Device 15 were extremely similar to those of all other single-spot laser devices except for the longitudinally textured sections, where single-spot lasers have been shown to be ineffective. This means that MPD measurements are not significantly affected by vehicle wander and macrotexture measurement path, at least under the controlled conditions of the experiment. 4.1.4 Verification of High-Speed Equipment Repeatability and Agreement Table 30 compares the repeatability of the devices on the tested MnROAD surfaces. In general, the values fell within the range (or better) than those obtained in the first comparison; however, two exceptions (highlighted in red) are analyzed in detail. Figure 46 shows the average MPD Device 1 Device 3 Device 4 Device 5 Device 6 Device 7 Device 10 MSE (mm) 5.5 E-5 9.3 E-4 5.3 E-4 1.3 E-3 3.4 E-4 2.2 E-3 1.8 E-4 (mm) 0.02 0.08 0.06 0.10 0.05 0.13 0.04 Device 1 = WDM TM2; Device 3 = VTTI; Device 4 = Ames Accutexture; Device 5 = LTPP; Device 6 = SSI placed at 45º; Device 7 = Ames Optocator; Device 10 = SSI placed transversally. Table 30. Summary of device repeatability (Butterworth filtered, in mm). Line Lasers on Long Grooved PCC Figure 46. Comparison of MPD results for the various devices on all the surfaces.

Data Analysis 79   values for the various sections, devices, and runs. It is clear, again, that the line lasers do not produce accurate texture measurements on the longitudinally textured concrete, as highlighted in the figure. Table 31 shows the LOA calculations for the various device pairs. As was observed in the first comparison experiment, the agreement was highly affected by the type of sensor used and by the texturing pattern of the tested surface. The comparisons again highlighted the problems of using single-spot lasers to measure the MPD on Section 9, which had pronounced longitudinal grooves. 4.1.5 Main Findings The main findings of the repeatability and reproducibility analysis are as follows: Static and Walking-Speed Devices • The lowest MPD repeatability coefficient, 0.0021 mm, was obtained with Device 12 (reference beam) on the pavement tested in the HVS. These values are quite good, as the threshold and investigatory levels for pavement macrotexture for several agencies (Austroads AP-T290-15 2015; New Zealand Transport Agency 2013) are given with one decimal place of precision. All other parameters calculated showed acceptable repeatability. • The MPD repeatability coefficient for Device 6 (C.T. Meter), a single-spot laser device that follows a circular path), was found to be 0.05 mm in the trafficked area of the HVS when Type of Sensor * Device Pair Texture Type All Random ** Longitudinal Only ** Transverse Only ** Mean of Difference Mean of Difference Mean of Difference Mean of Difference LOA LOA LOA LOA Li ne – Li ne 1, 6 0.01 0.14 0.04 0.08 0.04 0.06 -0.11 0.05 1, 10 0.07 0.22 0.12 0.24 0 - 0 0.05 6, 10 0.08 0.24 0.09 0.31 -0.02 - 0.11 0.05 SS L – Li ne 1, 3 0.03 0.63 -0.08 0.22 0.35 0.89 -0.19 0.11 1, 4 0.09 0.67 -0.04 0.06 0.44 0.99 -0.14 0.07 1, 5 0.11 0.52 0.05 0.1 0.37 0.74 -0.14 0.03 1, 7 0.06 0.57 -0.03 0.14 0.36 0.77 -0.17 0.08 3, 6 -0.02 0.59 0.12 0.22 -0.31 0.86 0.08 0.16 3, 10 0.18 0.31 0.25 0.18 -0.13 - 0.19 0.14 4, 6 -0.07 0.63 0.08 0.06 -0.4 0.95 0.03 0.12 4, 10 0.11 0.31 0.17 0.28 -0.16 - 0.14 0.1 5, 6 -0.1 0.46 -0.01 0.13 -0.32 0.71 0.03 0.08 5, 10 0.07 0.24 0.09 0.2 -0.13 - 0.14 0.07 6, 7 0.05 0.52 -0.07 0.14 0.32 0.73 -0.06 0.12 7, 10 0.13 0.29 0.18 0.26 -0.12 - 0.17 0.11 SS L – SS L 3, 4 0.06 0.16 0.04 0.19 0.09 0.16 0.06 0.05 3, 5 0.08 0.18 0.13 0.14 0.01 0.23 0.05 0.08 3, 7 0.03 0.16 0.05 0.16 0.01 0.21 0.02 0.05 4, 5 0.02 0.21 0.09 0.09 -0.07 0.28 0 0.04 4, 7 -0.03 0.17 0.01 0.12 -0.08 0.25 -0.04 0.03 5, 7 -0.05 0.13 -0.09 0.12 0 0.13 -0.03 0.05 * SSL = single-spot laser; Line = line laser. ** Comparisons with LOA > 0.15 mm are highlighted in bold. Table 31. Summary of LOA (Butterworth filtered, in mm).

80 Protocols for Network-Level Macrotexture Measurement repeatedly tested without moving the device. When the device was picked up and replaced (to replicate field conditions) the repeatability coefficient increased to 0.14 mm. • The MPD repeatability of each of the walking-speed devices tested was found to be less than 0.05 mm. • The MPD LOA of all static devices on the 10 pavement types tested at the Virginia Smart Road were 0.2–0.4 mm for randomly-textured surfaces, 0.3–0.7 mm for longitudinally textured pavements, and 0.1–0.3 mm for transversely-texture sections (i.e., tines). This means that in most cases, the devices should not be used interchangeably if, in the engineer’s judg- ment, the differences in measurements should not exceed these ranges. • The LOA for all of the walking-speed device MPDs were 0.06–0.2 mm for randomly-textured sections, 0.07–1.2 mm for longitudinally-texture sections (i.e., grooves or diamond grinding), and 0.3–0.8 mm for transversely tined PCC sections. In general, agreement was better when comparing the two line-laser sensors with one another than some combination of the line lasers and the single-spot device with the line laser. The interpretation of these results is that in some cases, the devices may be used interchangeably (i.e., those cases where the agreement is better than within 0.1 mm MPD) and in other cases, they should not be used interchangeably. High-Speed Devices • All devices tested showed good repeatability for all surfaces tested with coefficients of repeatability ranging from 0.063–0.088 mm in the first equipment comparison. Similar values were obtained for most of the equipment tested in the verification experiment, except in the case of the single-spot lasers, which showed higher variability on the longitudinally textured surfaces (0.10–0.13 mm). • As expected, low agreement was found for the line laser when compared to all other single- spot laser devices on all sections in a LOA analysis. The wide band of agreement is likely because the line laser measures the peaks and valleys of longitudinal texturing, whereas single-spot lasers cannot. Agreement improved significantly (lack of agreement fell by more than half ) when considering only surfaces similar to asphalt (i.e., SMA, OGFC, and HFST) in the LOA analysis. • Some problems were observed with some of the older, 32 kHz single-spot lasers. One of the devices provided a smoother raw data signal than the other single-spot lasers tested, and appears to have a bias in calculated MPD of approximately 0.1 mm when compared to all other devices. Speed was also found to be a significant factor in determining a pavement’s macrotexture for some of these devices. • Readings from single-spot and transversely mounted line lasers should not be used interchangeably on longitudinally textured surfaces such as tined, grooved, or ground PCC surfaces. 4.2 Validation Results The validation experiment included test speed and laser exposure time in the comparison test matrix and measurements with a reference beam to collect static reference texture data using a high-resolution laser. All the evaluations were based on the computed MPD. 4.2.1 Reference Measurements TTI researchers collected reference texture measurements with the LAPS at 5 m stations along the test wheelpath of each section. After positioning the line laser to track the test wheel- path, researchers collected displacement readings over a 1 m test length, making two repeat LAPS runs on each 5 m station. For data collection, researchers angled the laser such that its

Data Analysis 81   footprint was oriented perpendicular to the direction of measurement, with a width of approxi- mately 50 mm at the height the laser was set. Displacement readings were collected at 0.1 mm intervals in the transverse and longitudinal directions. Figure 47, Figure 48, and Figure 49 show examples of the 3D illustrations of line- laser displacement readings collected over a sample length of 100 mm from each test surface. Because the laser footprint width was half of the standard 100 mm base segment length specified in ASTM E 1845, TTI researchers computed MPDs using the displacement readings collected in the longitudinal direction. For this analysis, researchers used the data in the center 100 longitudinal profiles, calculating mean segment depths at 100 mm intervals, and averaging the mean segment depths over the 1 m test length to determine the MPD for each longitudinal profile. Researchers then took the average of the MPDs computed using the center 100 longitu- dinal profiles to determine the 1 m MPD at each measurement station. Figure 50 compares the 1 m MPDs determined from the two repeat measurements made with the LAPS. It is observed that the macrotexture statistics between runs plot very close to the line of equality, indicating the excellent repeatability of the LAPS test data. The average of the absolute differences in 1 m MPDs between runs is 2 µm. 4.2.2 MPD Reproducibility Across Test Equipment The two participating manufacturers brought test vehicles equipped with two lasers for macro- texture measurements. Each company configured the laser pair to track the same wheelpath on any given run. Because texture readings were collected concurrently, the following sections examine the agreement in the computed MPDs between the lasers used on each test vehicle. For this purpose, researchers examined the differences between computed 1 m MPDs and determined the 95% confidence intervals (CIs) of these differences. Figure 47. Example 3D plot of LAPS data from ground and longitudinally grooved concrete section.

82 Protocols for Network-Level Macrotexture Measurement Figure 48. Example 3D plot of LAPS data from OGFC section. Figure 49. Example 3D plot of LAPS data from DGF2 section.

Data Analysis 83   Comparison of Single-Spot Lasers Figure 51 compares the CIs of the average MPD differences determined from data collected with the two SS lasers on the HMA and PCC sections. The horizontal axis in each figure iden- tifies the combination of test variables (surface type, test speed, and exposure setting) at which the given 95% CI was determined. Each CI has been plotted as a vertical segment whose end- points correspond to the upper and lower limits of the interval. The square within each vertical segment corresponds to the average MPD difference. CIs that include zero identify cases where the MPD difference is not statistically significant. As seen in the figure, in most cases, the differ- ences in MPDs between the lasers are significant at α = 0.05. Most of the differences are positive, indicating higher MPDs determined from Optocator measurements compared to the Acuity. To gauge the practical significance of the MPD differences between lasers, the researchers used a tolerance of ±0.1 mm, identified by the red horizontal lines in Figure 51. This tolerance corresponds to the separation between normal, investigation, and intervention levels of macrotexture used by some agencies. Comparison of Results from Line Laser Using Different Orientations One of the test vehicles was equipped with two SSI Gocator line lasers that were mounted in two configurations (LLT and LLA), referred to as “SSIT” and “SSIA” in this section. Figure 52 shows the CIs of the average MPD differences determined from SSIT and SSIA data on the HMA and PCC sections, respectively. On the HMA sections, the mean differences between SSIT and SSIA MPDs were not statistically significant for most cases. In addition, the mean MPD differences were mostly within ±0.1 mm. For the four cases where the average MPD differences were outside this tolerance, the test data were collected at the laser exposure setting of 40 µs. Figure 50. Comparison of 1 m average MPDs determined from repeat LAPS measurements.

84 Protocols for Network-Level Macrotexture Measurement (a) HMA sections (b) PCC sections Figure 51. The 95% CIs of differences between Optocator and Acuity MPDs.

Data Analysis 85   (a) HMA sections (b) PCC sections Figure 52. 95% CIs of differences between SSIT and SSIA MPDs.

86 Protocols for Network-Level Macrotexture Measurement On the longitudinally grooved concrete section, the differences between SSIT and SSIA MPDs were not statistically significant in 11 of the 12 comparisons made. In all cases, the average MPD differences were within ±0.1 mm. As expected, all average MPD differences were statistically significant on the transversely grooved concrete section, and they fell outside the ±0.1 mm tolerance band. The differences were negative, reflecting the lower MPDs determined from the SSIT data for this section. As noted previously, the displacement readings along the SSIT scan will not show as many differ- ences in elevations compared to the SSIA scan, where the laser footprint cuts across the ridges and grooves of the transversely grooved surface. Thus, the computed MPDs from the SSIA data were expected to be larger than the corresponding values from the SSIT data. 4.2.3 Reproducibility Compared to Reference Beam This section evaluates the ability of the various macrotexture measuring devices to reproduce the measurements obtained with the reference beam. Comparisons of MPDs Computed from LAPS and Single-Point Texture Measurements Figure 53 and Figure 54 show 95% CIs of the mean differences between MPDs computed from the LAPS and Acuity and Optocator measurements, respectively. As shown, the mean Figure 53. The 95% CIs of differences between Acuity and LAPS MPDs.

Data Analysis 87   differences between LAPS and Acuity MPDs vary over a narrower range compared to the differ- ences between LAPS and Optocator MPDs, highlighting the benefits of the higher-frequency laser. The following observations were noted from the results presented: • In general, the Acuity MPDs showed better agreement with the corresponding LAPS indices than the Optocator MPDs. Although almost all the mean MPD differences were statistically significant, more of these differences fell within or closer to the ±0.1 mm tolerance band for the Acuity laser than for the Optocator. For the DGF1, HMAP, and stone-matrix asphalt (SMAF) surfaces, the mean MPD differences for the Acuity laser were all within the ±0.1 mm tolerance band. • The differences in MPDs between the LAPS and single-point texture lasers were relatively greater on the concrete and OGFC sections. On these sections, higher mean MPD differences were associated with the Optocator, particularly at the 55 mph test speed. • On the surfaces where the mean differences between the Acuity and LAPS MPDs were more than 0.1 mm, the normal exposure setting generally gave the least mean MPD differences, particularly on the concrete sections. On the transversely grooved section, the mean MPD differences at 25 and 40 mph were within the ±0.1 mm tolerance band at normal exposure. • Considering the findings from the statistical analysis of main and two-way interaction effects, and the comparisons between LAPS and single-point texture laser MPDs, the Acuity laser at the normal exposure time setting was found to be more suited to field measurements of macrotexture than the Optocator. Figure 54. The 95% CIs of differences between Optocator and LAPS MPDs.

88 Protocols for Network-Level Macrotexture Measurement Comparisons of MPDs Computed from LAPS and Line-Laser Measurements Figure 55 and Figure 56 show 95% CIs of the mean differences between MPDs computed from LAPS and SSIA, and SSIT measurements, respectively. The following observations were noted from the results presented: • The number of cases where the mean MPD differences were statistically significant was about the same for both the SSIA and the SSIT configurations. On the HMA sections, the mean differences between SSIA and LAPS MPDs were statistically significant in 38 of the 60 comparisons. The corresponding number for the SSIT configuration was 36 comparisons. On the PCC sections, all the mean MPD differences were statistically significant for both the SSIA and the SSIT configurations. • Although a sizeable majority of the mean MPD differences was statistically significant on the HMA sections, the 95% CIs of the mean MPD differences overlapped with the ±0.1 mm tolerance band in 55 of the 60 comparisons made between SSIA and LAPS MPDs. The corresponding number for the SSIT configuration was 54 comparisons. Except for the OGFC, the SSIA mean MPDs agreed to within 0.1 mm of the corresponding LAPS indices on the other four HMA sections. On the other hand, mean differences between SSIT and LAPS MPDs above 0.1 mm were observed with the OGFC, HMAP, and SMAF sections. Among these three types of surfaces, four cases had mean MPD differences above 0.2 mm. These four cases were associated with SSIT test data collected on the HMAP and OGFC sections at the laser exposure setting of 40 µs. • On the PCC sections, none of the 95% CIs overlapped with the ±0.1 mm tolerance band. Figure 55. The 95% CIs of differences between SSIA and LAPS MPDs.

Data Analysis 89   Comparisons of MPDs Computed from LAPS and Walking Texture Measurements At the RELLIS experiment, the TM2 operator made three repeat runs on each pavement section. Because the TM2 produced repeatable results in the previous experiments, TTI researchers also compared the TM2 MPDs with the corresponding indices determined from LAPS data to evaluate its usefulness as a reference texture-measurement device. Figure 57 shows the results from these comparisons. The results shown for the PCCT section are based on test data from runs made where the TM2 operator followed a zigzag path along the test section, as illustrated in Figure 58. This path was determined from GPS data included with the TM2 measurements and has been color-coded in the figure to indicate the variation in computed MPDs along the path the oper ator followed. Because the TM2’s line-laser orientation was fixed perpendicular to the travel direction, the equipment operator also walked the TM2 to collect measurements with the line laser angled relative to the transverse tines (see Figure 58). The 95% CIs of the mean differences between the TM2 and LAPS MPDs generally shows good agreement between both sets of measurements. Researchers noted the following observa- tions from Figure 57: • On the DGF1, SMAF, PCCL, and PCCT sections, the mean MPD differences were not statistically significant. The 95% CIs included zero on these sections. Figure 56. The 95% CIs of differences between SSIT and LAPS MPDs.

Figure 57. The 95% CIs of differences between TM2 and LAPS MPDs. Figure 58. Test path tracked by TM2 operator on PCCT section (third run).

Data Analysis 91   • The 95% CIs of the mean MPD differences overlapped with the ±0.1 mm tolerance band on all test sections, except for the PCCT section. • Among the sections tested, the CIs were within the ±0.1 mm tolerance band on DGF1, DGF2, HMAP, SMAF, and PCCL. • On average, the TM2 MPDs were 0.12 mm lower than the LAPS MPDs on the OGFC section. Given the good agreement between the TM2 and the LAPS on the DGF1, DGF2, HMAP, SMAF, and PCCL sections, comparing the Ames and SSI MPDs with the corresponding TM2 indices was expected to yield similar results as those presented from the earlier comparisons between MPDs from high-speed runs and the LAPS on these sections. On the OGFC section, the previous comparisons showed higher MPDs from the Ames and SSI test runs relative to the corresponding LAPS indices. In this regard, the Acuity and SSIA MPDs were, on average, greater than the corresponding LAPS MPDs by 0.22 mm and 0.14 mm, respectively. Because the TM2 slightly underestimated the LAPS MPDs, comparing the Ames and SSI MPDs with the corresponding TM2 indices on the OGFC section would likely show less agreement relative to results from the earlier LAPS comparisons. On the PCCT section, the mean MPD difference between the TM2 and the LAPS (0.04 mm) was close to the overall mean difference between the SSIA and the LAPS diagonal MPDs (0.03 mm) across the range of test speeds and exposure settings used to collect SSIA data. Because the TM2 operator tested along a different path, the comparison between the TM2 and LAPS MPDs was based on the mean MPDs of the respective paths tested. This comparison was more sensitive to the macrotexture variability along the path tested as opposed to paired comparisons of corresponding MPDs determined over the same path. For this reason, the CI in Figure 41 on the transversely grooved concrete section is wider than the CIs from earlier comparisons of the SSIA and LAPS MPDs (Figure 36). Because the TM2 followed a different path, the CI shown in Figure 57 for the PCCT section is not as meaningful, in the opinion of the researchers, as the CIs determined on the other sections. In the researchers’ opinion, the conclusion to be drawn from this comparison is that the mean MPD over the path the TM2 followed differed from the mean MPD over the path the LAPS tested by 0.04 mm. 4.2.4 Main Findings Based on the results from the comparative evaluation of macrotexture measurement devices, the main findings and conclusions from the comparison experiment are the following: • Statistically significant effects of test speed, laser exposure setting, and laser type on the MPDs determined from single-point texture measurements were primarily observed with the PCCT and OGFC sections. • Between the two single-point texture lasers, however, the effect of test speed on the Acuity laser MPDs was not statistically significant. In addition, the Acuity laser MPDs showed better agreement with the corresponding LAPS indices compared to the Optocator MPDs. These findings confirm that the higher-frequency Acuity laser is better suited for field measurements of macrotexture than the Optocator. The findings also support setting this laser to normal exposure mode when collecting field measurements. • The statistical analysis of test speed–exposure time interaction on the SSI test data revealed no significant effect of test speed on computed MPDs from data collected at the auto expo- sure setting. Some significant differences were found for the data collected at the other exposure settings. These findings again support using the auto exposure mode to collect macrotexture measurements with the same Gocator line lasers used in this study. • Except for the OGFC, the SSIA mean MPDs agreed to within 0.1 mm of the corresponding LAPS indices on the other HMA test sections. On the other hand, the mean differences

92 Protocols for Network-Level Macrotexture Measurement between SSIT and LAPS MPDs above 0.1 mm were observed not only on the OGFC but also on the HMAP and SMAF sections. On the HMAP and OGFC sections, mean MPD differences above 0.2 mm were obtained from SSIT test data collected at the 40 µs exposure setting. • Comparisons between the Gocator line laser and the LAPS MPDs on the concrete sections highlighted the need to account for the directional nature of the concrete surface macro- texture when collecting and processing test data on these surfaces. In this regard, researchers used the 3D data from the LAPS to resolve significant differences between computed MPDs on the PCCL and PCCT sections tested at RELLIS. Between the 0° and 45° laser footprint orientations (with the angle relative to the lateral axis), the findings from an analysis of the test data favored the 45° angle. This orientation will capture the longitudinal/transverse ridges or grooves resulting from concrete surface tining, grooving, or grinding. • The Acuity single-point and SSIA line-laser readings gave comparable MPDs on the trans- versely grooved concrete section. However, the Acuity laser MPDs underestimated the SSIA MPDs by about 0.6 mm, on average, on the longitudinally grooved concrete surface. This finding confirms that line lasers should be used for measurement of macrotexture on these surfaces. • Comparisons of MPDs determined from the TM2 and LAPS test data generally showed good agreement. Among the seven sections tested, the 95% CIs were within the ±0.1 mm tolerance band on the DGF1, DGF2, HMAP, SMAF, and PCCL sections. On the OGFC section, the TM2 MPDs were 0.12 mm lower, on average, than the corresponding LAPS MPDs. In comparison, the SSIA MPDs were, on average, 0.14 mm higher than the corresponding LAPS MPDs. Among the HMA sections, the MPD comparisons generally showed less agreement on the OGFC. • Modifying the TM2 to allow adjustment of the line-laser footprint angle would permit the operator to collect angled line-laser readings directly along the wheelpath of transversely tined or grooved concrete sections. Given the generally good agreement between the TM2 and LAPS MPDs, this modification would significantly improve the utility of the TM2 as a device for verifying high-speed macrotexture measurement systems. 4.3 Accuracy on Engineered Surfaces The effects of speed and exposure time were further studied with measurements on the reference plates. 4.3.1 MPD Comparisons For the test setup described in Chapter 3 under “Engineered Surfaces,” the MPD values were calculated using the test data from all devices collected on all plates for each speed and exposure time test matrix combination. Table 32 summarizes the mean MPD values for Plate 1. In general, the MPD results tended to be negatively correlated to vehicle speed and exposure time. This was not always the case, however, and was explored further in the forest plot analysis. Table 32 shows the MPD results for each of the devices tested at each of the travel speeds and exposure settings laid out in Table 17. Results are given for both raw (unfiltered) profiles and for profiles that were first filtered according to the parameters given in ASTM E1845 (2015). For the single-spot laser profiles, filtering was completed on the entire measured profile (which included several meters of lead-in and lead-out data before and after each plate). For the line-laser profiles, each 100 mm (approximate) base length was first unfolded by 2.5 mm (the wavelength of the low-pass filter used) by mirroring the start and end. This was done to

Data Analysis 93   avoid distortion of the beginning and end of the profile. Mirrored portions were removed after filtering was complete. For the longest exposure time, MPD decreased with speed. This was observed to a lesser extent (or not at all) as the exposure time was made shorter. Filtering lowered the MPD of all devices at all speeds and exposures by 16% (for short exposures) to 22% (for longer exposures). These trends can be seen in Figure 59 and Figure 60. Table 33 shows the MPD values for the other two plates. The impacts of speed and exposure were significantly lower for these plates. Raw data profiles of the first 100 mm of the high-speed measurements of the single- spot laser are shown in Figures 61 through 63. As evidenced in Figure 61, the plate with the small waveform (Plate 1) shape proved difficult to profile. This difficulty was exacerbated by increased speed and exposure time. The inaccuracy of profile reproduction increased as the vehicle speed and sensor exposure time increased. For the fastest test speed (55 mph or 88.5 km/hr) and longest exposure time (30–40 µs), the profile was quite degraded. The signal was smoothed by an averaging effect of the long exposure time (both slope and peak data were captured in a single sample), resulting in lowered MPD value. The profiles shown in Figure 62 and Figure 63 demonstrate a more stable measurement of the plate profiles. However, small “lobes” can be seen in the raw data profile near the edges of the peak and valley plateaus. This can be attributed to either light dispersion or hardware filtering at the sharp corners and represents a shortcoming of this measurement approach for sharp features. The effects of these edges were averaged into the surrounding profile by low-pass filtering, as shown in the respective figures. Figure 64 further illustrates the effect of the laser head orientation on the displacement read- ings obtained from testing the VTTI plate. This plate has a higher groove depth-to-groove width ratio compared to the other large plates (Plates 5 and 6). Exposure Speed (mph) Plate 1 Mean MPD (mm), Raw Plate 1 Mean MPD (mm), E1845 Filter SSL LLL LLT SSL LLL LLT Auto Mean - 1.12 1.18 - 0.93 0.96 25 - 1.16 1.17 - 0.94 0.98 40 - 1.11 1.18 - 0.93 0.95 55 - 1.10 1.19 - 0.93 0.96 Long Mean 0.88 0.72 1.34 0.74 0.60 1.03 25 0.92 0.99 1.32 0.79 0.84 1.03 40 0.90 0.73 1.36 0.75 0.61 1.03 55 0.81 0.45 1.33 0.69 0.34 1.04 Medium Mean 1.04 1.05 1.23 0.89 0.90 0.98 25 1.04 1.15 1.23 0.89 0.95 0.98 40 1.05 1.05 1.24 0.89 0.90 0.98 55 1.04 0.95 1.21 0.89 0.84 0.99 Short Mean 1.03 1.15 1.21 0.88 0.94 0.96 25 1.03 1.16 1.21 0.87 0.95 0.95 40 1.05 1.15 1.18 0.89 0.94 0.95 55 1.03 1.13 1.23 0.87 0.94 0.96 SSL = single-spot laser; LLL = line laser with longitudinal orientation; LLT = line laser with transverse orientation. Table 32. Summary of MPD values measured by high-speed devices for Plate 1.

94 Protocols for Network-Level Macrotexture Measurement Figure 59. Example MPD results—single-spot laser. Figure 60. Example MPD results—LLL.

Data Analysis 95   Figure 64 shows that, depending on the laser head orientation, outlier readings were measured on the side slope where the receiver did not capture the transmitted beam from the laser head. To minimize the observed outliers found from examining the initial reference plate readings, researchers reconfigured the LAPS to orient the line-laser footprint perpendicular to the transverse grooves, as shown in the figure. Researchers then retested the reference plates with this line-laser orientation. The data from this retest did not show any outliers, and they were subsequently used to determine the reference plate MPDs. 4.3.2 Detailed Analysis To obtain more granularity on which plate and speed/exposure combinations cause unequal means, forest plots were used. Forest plots are graphical depictions of given center values and a range. This project sought to examine the difference between the reference (LAPS) device and the field device. Therefore, the center values used for the forest plot were used to indicate the difference in means between the two devices for a given plate, speed, and exposure combi nation. Because the reference device was used in a laboratory setting without varying the speed or exposure time, one set of six MSDs—one from each segment shown in Figure 34—were used to represent each plate for the LAPS. Six MSDs were likewise used for each high-speed device and speed/exposure combination. These six measurements from the high-speed devices were averaged together via arithmetic mean to compare against the average MSD (the MPD) from the LAPS. E xp os ur e Speed (mph) Plate 5 Mean MPD (mm), Raw Plate 5 Mean MPD (mm), E1845 Filter Plate 6 Mean MPD (mm), Raw Plate 6 Mean MPD (mm), E1845 Filter SSL LLL LLT SSL LLL LLT SSL LLL LLT SSL LLL LLT A ut o Mean - 3.92 3.92 - 3.94 3.94 - 5.06 5.20 - 5.07 5.21 25 - 3.95 3.90 - 3.96 3.92 - 4.83 5.20 - 4.83 5.20 40 - 3.91 3.96 - 3.93 3.98 - 5.17 5.12 - 5.18 5.14 55 - 3.90 3.90 - 3.92 3.92 - 5.18 5.27 - 5.18 5.29 Lo ng Mean 3.93 3.93 3.93 3.88 3.94 3.96 5.19 5.17 5.16 5.16 5.18 5.19 25 3.89 3.95 3.99 3.87 3.96 4.01 5.14 5.16 5.11 5.13 5.16 5.13 40 3.92 3.93 3.91 3.87 3.94 3.94 5.22 5.16 5.12 5.18 5.16 5.15 55 3.98 3.92 3.90 3.90 3.92 3.92 5.21 5.20 5.26 5.16 5.20 5.27 M ed iu m Mean 3.88 3.92 3.93 3.89 3.93 3.95 5.18 5.20 5.20 5.18 5.20 5.22 25 3.88 3.93 3.90 3.89 3.95 3.92 5.17 5.23 5.13 5.18 5.24 5.15 40 3.89 3.91 3.89 3.89 3.93 3.92 5.14 5.17 5.20 5.15 5.18 5.22 55 3.87 3.91 4.00 3.88 3.93 4.02 5.22 5.19 5.25 5.21 5.19 5.27 Sh or t Mean 3.88 3.96 3.89 3.89 3.97 3.90 5.20 5.23 5.17 5.20 5.24 5.18 25 3.87 3.97 3.90 3.89 3.98 3.92 5.16 5.22 5.11 5.17 5.21 5.12 40 3.88 3.97 3.90 3.89 3.98 3.91 5.18 5.21 5.13 5.18 5.22 5.14 55 3.90 3.94 3.86 3.89 3.95 3.88 5.26 5.28 5.28 5.26 5.28 5.30 SSL = single-spot laser; LLL = line laser with longitudinal orientation; LLT = line laser with transverse orientation. Table 33. Summary of MPD values measured by high-speed devices for Plates 5 and 6.

0 50 100 150 200 250 300 350 400 450 -1.5 -1 -0.5 0 0.5 1 1.5 Te xt ur e H ei gh t( m m ) Plate 1, Short Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 -1.5 -1 -0.5 0 0.5 1 1.5 Plate 1, Short Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 -1.5 -1 -0.5 0 0.5 1 1.5 Plate 1, Short Exposure, 90 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 450 -1.5 -1 -0.5 0 0.5 1 1.5 Te xt ur e H ei gh t( m m ) Plate 1, Medium Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 450 -1.5 -1 -0.5 0 0.5 1 1.5 Plate 1, Medium Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 450 -1.5 -1 -0.5 0 0.5 1 1.5 Plate 1, Medium Exposure, 90 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 Datapoint -1.5 -1 -0.5 0 0.5 1 1.5 Te xt ur e H ei gh t( m m ) Plate 1, Long Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 450 Datapoint -1.5 -1 -0.5 0 0.5 1 1.5 Plate 1, Long Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 450 Datapoint -1.5 -1 -0.5 0 0.5 1 1.5 Plate 1, Long Exposure, 90 km/hr, filter = ASTM Filter raw filtered Figure 61. Single-spot laser Plate 1, first 100 mm, all exposures and speeds.

Figure 62. Single-spot laser Plate 5, first 100 mm, all exposures and speeds. 0 50 100 150 200 250 300 350 400 -4 -3 -2 -1 0 1 2 3 4 Te xt ur e H ei gh t( m m ) Plate 5, Short Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 -4 -3 -2 -1 0 1 2 3 4 Plate 5, Short Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 -4 -3 -2 -1 0 1 2 3 4 Plate 5, Short Exposure, 90 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 -4 -3 -2 -1 0 1 2 3 4 Te xt ur e H ei gh t( m m ) Plate 5, Medium Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 -4 -3 -2 -1 0 1 2 3 4 Plate 5, Medium Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 -4 -3 -2 -1 0 1 2 3 4 Plate 5, Medium Exposure, 90 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 Datapoint -4 -3 -2 -1 0 1 2 3 4 Te xt ur e H ei gh t( m m ) Plate 5, Long Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 Datapoint -4 -3 -2 -1 0 1 2 3 4 Plate 5, Long Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 400 Datapoint -4 -3 -2 -1 0 1 2 3 4 Plate 5, Long Exposure, 90 km/hr, filter = ASTM Filter raw filtered

Figure 63. Single-spot laser Plate 6, first 100 mm, all exposures and speeds. 0 50 100 150 200 250 300 350 -6 -4 -2 0 2 4 6 Te xt ur e H ei gh t( m m ) Plate 6, Short Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 -6 -4 -2 0 2 4 6 Plate 6, Short Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 -6 -4 -2 0 2 4 6 Plate 6, Short Exposure, 90 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 -6 -4 -2 0 2 4 6 Te xt ur e H ei gh t( m m ) Plate 6, Medium Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 -6 -4 -2 0 2 4 6 Plate 6, Medium Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 -6 -4 -2 0 2 4 6 Plate 6, Medium Exposure, 90 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 Datapoint -6 -4 -2 0 2 4 6 Te xt ur e H ei gh t( m m ) Plate 6, Long Exposure, 40 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 Datapoint -6 -4 -2 0 2 4 6 Plate 6, Long Exposure, 65 km/hr, filter = ASTM Filter 0 50 100 150 200 250 300 350 Datapoint -6 -4 -2 0 2 4 6 Plate 6, Long Exposure, 90 km/hr, filter = ASTM Filter raw filtered

Data Analysis 99   The ranges used on the forest plots in Figure 65 and Figure 66 are representative of the 95% CI for the measurements made. To obtain this interval, standard errors for each com- bination of the plates and speed/exposures were calculated for each device using a pooled estimate obtained from the 12 MSDs from both the reference measurement and the high-speed device. The 95% CI was then calculated by adding (for the upper interval) or subtracting (for the lower interval) the product of the standard error and 1.96 (given that the desired confidence level was 95%). This provided a 95% confidence that the mean MPD difference at a given speed/exposure combination for a given plate will fall somewhere on the line plotted in Figure 65. In this way, the research team could visualize the variance of a given device. In Figure 65, the ranges that contain zero are colored black. This can be interpreted as no difference between the reference and field measurements, given the 95% confidence that the field measurement will fall somewhere on this line. Because the field measurements can be lower or higher than the zero-line drawn, statistically, the difference between the two devices is zero. The central green zone on each plot represents a difference in means of 0.1 mm. Some agencies have established normal, investigation, and intervention levels of macrotexture, and these levels are separated by 0.1 mm MPD. As such, it is useful to see if the difference between reference and field measurements is within this threshold to assess its practical significance. In Figures 65 and 66, center values (differences in means) that fall within this green zone indicate that the mean field measurements were within this tolerance of the mean reference measurements. Figure 64. Effect of laser head orientation on VTTI plate test readings.

Plate 1, no filtering Difference of Mean MPDs (mm) -0.5 -0.3 + / - 0.1 0.3 0.5 1_LLT_Short_55 1_LLT_Short_40 1_LLT_Short_25 1_LLT_Med_55 1_LLT_Med_40 1_LLT_Med_25 1_LLT_Long_55 1_LLT_Long_40 1_LLT_Long_25 1_LLT_Auto_55 1_LLT_Auto_40 1_LLT_Auto_25 1_LLL_Short_55 1_LLL_Short_40 1_LLL_Short_25 1_LLL_Med_55 1_LLL_Med_40 1_LLL_Med_25 1_LLL_Long_55 1_LLL_Long_40 1_LLL_Long_25 1_LLL_Auto_55 1_LLL_Auto_40 1_LLL_Auto_25 1_SSL_Short_55 1_SSL_Short_40 1_SSL_Short_25 1_SSL_Med_55 1_SSL_Med_40 1_SSL_Med_25 1_SSL_Long_55 1_SSL_Long_40 1_SSL_Long_25 Plate 5, no filtering Difference of Mean MPDs (mm) -0.50 -0.25 + / - 0.1 0.25 0.50 Plate 6, no filtering Difference of Mean MPDs (mm) -0.5 -0.3 + / - 0.1 0.3 0.5 Figure 65. Forest plots of differences between reference and field measurements—unfiltered.

Plate 1, E1845 filtering Difference of Mean MPDs (mm) -0.5 -0.3 + / - 0.1 0.3 0.5 1_LLT_Short_55 1_LLT_Short_40 1_LLT_Short_25 1_LLT_Med_55 1_LLT_Med_40 1_LLT_Med_25 1_LLT_Long_55 1_LLT_Long_40 1_LLT_Long_25 1_LLT_Auto_55 1_LLT_Auto_40 1_LLT_Auto_25 1_LLL_Short_55 1_LLL_Short_40 1_LLL_Short_25 1_LLL_Med_55 1_LLL_Med_40 1_LLL_Med_25 1_LLL_Long_55 1_LLL_Long_40 1_LLL_Long_25 1_LLL_Auto_55 1_LLL_Auto_40 1_LLL_Auto_25 1_SSL_Short_55 1_SSL_Short_40 1_SSL_Short_25 1_SSL_Med_55 1_SSL_Med_40 1_SSL_Med_25 1_SSL_Long_55 1_SSL_Long_40 1_SSL_Long_25 Plate 5, E1845 filtering Difference of Mean MPDs (mm) -0.50 -0.25 + / - 0.1 0.25 0.50 Plate 6, E1845 filtering Difference of Mean MPDs (mm) -0.5 -0.3 + / - 0.1 0.3 0.5 Figure 66. Forest plots of mean differences, reference field measurements—with E1845 filtering.

102 Protocols for Network-Level Macrotexture Measurement Plate 1 As seen in Figures 65 and 66, the transverse line laser did not reliably reproduce the plate with the smallest waveform (Plate 1). The single-spot laser and the longitudinally oriented line lasers were both effective at reproducing the waveform at short, medium, and auto (for the LLL) exposure settings at the fastest speeds tested (with the exception of LLL, medium exposure time). Many of the single-spot laser measurements for Plate 1 contained zero within their CI, meaning the differences noted between the reference measurements and the high-speed device were effectively zero. The LLT, on the other hand, did not reproduce the waveform of Plate 1 at any speed or exposure setting. It should be noted that line lasers inherently have a longer exposure time than single-spot lasers. The line laser’s shortest exposure was similar to the single-spot laser’s longest exposure. This caused an averaging of peaks and slope readings, which resulted in a lower MPD reading. This effect also is seen generally as speed increased (and the same exposure time covered more distance on the ground), resulting in lower MPDs; however, the opposite effect can occur if laser light is reflected off the narrow valleys of Plate 1 and bounced between the subsequent peaks, causing an erroneous reflection back to the sensor. As a result, high-speed measurements may be higher than the laboratory reference measurements. For the single-spot laser and the line laser with a longitudinal orientation, speed had little effect on the results, with the exception of the longest exposure times. Plate 5 This plate had deeper groove depths than Plate 1 and valleys that matched the designed peak; thus, the LLT high-speed measurements were higher than the LAPS. The worst cases were observed with the long exposure settings. Speed was observed to have little effect on the difference of means between the reference measurements and the LLT. Of the devices tested, however, the LLT had the greatest observed variances, as shown by the CI bands in Figure 65. Most of the measurements made by all devices on Plate 5 contained zero within the 95% pooled CI, indicating that the lab reference measurements were the same statistically as the field measurements. Likewise, the differences of means between the lab and most field measurements were observed to be within the researchers’ tolerance of 0.1 mm, as indicated by the central dots within the green boundary area in the figure. A large majority of high-speed field measure- ments were higher than the lab reference measurements, indicating erroneous reflections back to the laser. The LLT was the only device to have two of the mean differences outside of the established 0.1 mm tolerance limit. Plate 6 This plate contained readings with the greatest variations, as demonstrated by the large CI bands in Figure 65. The scales on all plots in the figure are identical to allow for the most straightforward comparisons. As seen in the figure, Plate 6 had the deepest groove depth, and the greatest variations were seen using the auto exposure setting. This setting dynamically changes the exposure setting to anywhere within the exposure bands tested in this work to obtain an “optimal exposure” according to the device’s algorithm. However, larger variations may be seen due to sudden changes in surface color and reflectivity, initiating a change in exposure and causing readings in a spread of values that is larger than those of the fixed-exposure setting. A majority of device 95% CIs contained zero, meaning that the differences between the refer- ence and field measurements were not statistically significant. All the devices tested were able to accurately reproduce the waveform at the various speed/exposure combinations. This was true for most of the LLT readings; however, several of the differences of means fell very nearly outside of the 0.1 mm tolerance, and CIs were very close to not containing zero. This means that the LLT had the greatest difficulty of the devices tested in accurately reproducing the waveform. In cases where the LLT was close to the thresholds defined, the high-speed measurements were

Data Analysis 103   higher than the lab reference measurements. This was thought to be due to the lobes created on the signal at higher speeds and exposure settings. Effect of Filtering Figure 66 shows the effect of ASTM E1845-compliant low-pass filtering on the MPD results (visualized in Figure 61 and Figure 63). In general, the waveforms are smoother with filtering and appear to better represent the total peak height of the reference plate. For Plate 1, however, the raw waveform is reduced significantly, as the half-period of the wave (2.5 mm) is the same wavelength as the cutoff frequency of the filter designed for smoothing the signal. This results in a waveform that appears to be over-damped. However, if the waveforms taken by the reference equipment are treated with the same filtering, the two signals are comparable. Given its averaging effect, filtering tends to lower the variance of the results slightly, as is seen in all plots of Figure 66 when compared to Figure 65. The difference in means also has shifted slightly for all plates and devices. The field measurements have been made lower than the LAPS measurements because a sort of “double-penalty” is applied to the field measurements (first with the synthetic filter and next with the filtering that occurs as the greater ground covered for a given exposure tends to smooth the signal out). Filtering does, however, have the distinct advantage of removing the lobes created on the raw profiles from specular reflection at the peak corner points (see Figure 62 and Figure 63). 4.3.3 Main Findings The main findings of the accuracy analysis using engineered plates are as follows: • Reference surfaces such as those used in this experiment are effective at testing the accu- racy of laser triangulation devices at high speed. This is possible because the reference surfaces can be manufactured to known dimensions for comparison to device-measured values. Waveform periods should be selected as multiples of the base length for evaluation to ensure proper detrending can be performed for accurate results. – The base lengths selected for MPD analysis in this work were meticulously selected from the raw signal data, resulting in waveforms that were symmetric about a vertical axis so regression lines and calculated mean profile lines were unaffected. Plate shapes in the future should be designed such that the wavelength is a multiple of the selected base length (peak to peak) selected for evaluation. Such designs would greatly aid in automating the MPD calculation. – Plate base material should always be surface treated to reduce the specular reflections of the plate, especially at peak corner points. The plates in this experiment were coated in matte primer. This treatment worked well, but powdered coating could also be a viable alternative or potential improvement if the coating is more durable and results in improved proper reflection. • Vehicle speed and sensor exposure times were found to be significant factors in at least some of the measurements. – A deeper analysis of individual sensor and speed/exposure pairings through visualizing data on forest plots showed that as factors, speed and exposure are more critical for smaller waveforms, especially at higher speeds and longer exposure times. – Speed and exposure were not found to bring high-speed values out of the 0.1 mm tolerance established for this experiment. For the deepest plate (Plate 6), however, larger variations were observed. These variations were credited to the deepness and steepness of the machined channels and possibly to the sharp corners. For these reasons, it is suggested that a shape similar to Plate 5 be used with a tolerance of ±0.1 mm from a reference measurement.

104 Protocols for Network-Level Macrotexture Measurement – Filtering did not have any beneficial effect on the larger waveforms (Plates 5 and 6) and caused MPD values for Plate 1 to drop significantly. Therefore, filtering is not recommended for measurement of the reference surfaces. Care should be taken, however, to ensure that excessive outlier values are not contained in the profiles by carefully controlling the plate manufacturing process. If filtering is used on resulting data, the lab reference measure- ments also must be filtered to create similar waveforms for comparison, especially in the case of the smallest waveforms. 4.4 Macrotexture Characterization Parameters The objective of this portion of the project was to identify the best macrotexture parameters from a suite of existing and newly developed parameters (aimed at improving correlation with pavement surface properties) to predict a roadway’s frictional and noise characteristics without the use of additional specialized equipment for the collection of friction and noise data. The left wheelpath of the “uphill” travel lane of the Virginia Smart Road (see Figure 21) was measured with a single-spot laser device to capture the pavement surface macrotexture at 55 mph. The same path was measured with two continuous friction measurement devices and a test vehicle capable of measuring road noise. The friction and sound data were then distilled into the parameters for which prediction was attempted (predicted variables) from either a single macrotexture parameter (predictor variable) via single-variable linear regression or from several predictor variables via multiple linear regression. 4.4.1 Predictor Variables As mentioned in Chapter 3, one single-spot laser device was used to record a raw texture height profile of all road surfaces. Measurements were taken in the left wheelpath at 90 km/hr. All raw time domain data were converted to spatial data via linear interpolation based on texture height data, and distance pulses were gathered by a wheel encoder mounted to the driver-side rear tire. The minimum spatial sampling distance of 1⁄4 mm was used for all profiles, as determined by the following calculation (Equation 24): × × × = 88.5 1,000,000 3,600 1 100,000 0.25 . (24) km hr mm km hr s s samples mm sample Once a raw spatial-domain profile was obtained, the profile was distilled into a suite of macrotexture parameters that included the most popular parameters used in the United States— MPD and RMS—as well as less-widespread parameters obtained from the literature (i.e., skewness, kurtosis, and mean difference of elevation, or MDE). The research team also devised additional parameters with the intent of obtaining better correlations with the predicted variables. Table 34 summarizes all the parameters considered. Outliers were removed from the raw pavement profile following the methods described in Katicha et al. (2015). Filtering per ASTM E1845 (2015) was performed on each pavement surface profile before parameter calculation unless denoted to the contrary. The majority of the parameters used a base length of 100 mm for each calculation, and slope suppression via simple linear regression was performed on each base length for all parameters except for those derived by wavelet transformation. Note: In this report, a base length with the slope removed is referred to as a detrended base length.

Data Analysis 105   Traditional Measures MPD and RMS are well-documented macrotexture parameters; information on their calcula- tion can be obtained in the references given in Table 34. MDE takes the average difference of profile points as shown in Equation 25: ∑= − = − 1 , (25)1 1 MDE DE n ii n where DE = difference in elevation between point of interest on a detrended base length and the successive point, and n = number of points in the base length. Enveloping Profiles Enveloping profiles account for the free space between tire and pavement. Enveloping profiles can be used in a host of ways. For this research, the project team included for evaluation the Variable Number Parameter References 1 Mean Profile Depth (MPD) ASTM E1845 (2015) ISO 13473-1 (1997) 2 Root Mean Square (RMS) Wennink and Gerritsen (2000) ISO 13473-2 (2002) 3 Mean Difference of Elevation (MDE) Chou et al. (2017) 4–7 22–25 Enveloping Profiles Empirical Physical Effective Area of Water Evacuation (EAWE) Clapp (1983) Von Meier et al. (1992) Klein et al. (2004) Goubert (2007) Mogrovejo et al. (2016) 8–9 Geometric Statistical Methods Skewness (Rsk) Kurtosis (Rku) ISO 4288 (1996) ISO 4287 (1997) ASME B46.1 (2009) 10 Maximum Height (Max H) New 11–13 Percentile MPD (MPD95, MPD97, MPD99) New 14–21 Tire Contact Length (TCL) New 26–55 Wavelet Transformations (Wd,x)Various statistical measures of Haar details New; more information in: Zelelew et al. (2013) Leandri and Losa (2015) 56–60 Enveloping Profile MPD (MPDe) New; more information in: Goubert and Sandberg (2018) 61 Profile Length Ratio (PLR) New 62–79 Peak Data Parameters Mean Peak Height Above Zero (MPGZ) Mean Peak Above Zero Separation (MSEPGZ) Mean Prominence Separation Ratio (MPMSR) Mean Prominence Above Zero (MPROMGZ) Mean Width of Peaks Above Zero (MWGZ) Mean Prominence to Width Ratio (MPWR) Mean Peak Width Mean Peak Separation Ratio (MWMSR) Number of Peaks Above Zero (NPGZ) New Table 34. Summary of predictor parameters.

106 Protocols for Network-Level Macrotexture Measurement EAWE proposed by Mogrovejo et al. (2016). This is a single parameter that places the envelop- ing profile of a tire over the measured pavement surface and calculates the area between the two (Klein et  al. 2004). This method requires the use of tire stiffness coefficients (d*) to determine the extent of tire rubber penetration into the pavement surface. Table 35 lists the coefficients used in the analysis for the NCHRP research. Skewness (Rsk) and kurtosis (Rku) are the third and fourth statistical moments. Skewness can be used to capture positive or negative texture behavior. Negative skewness values indicate a negative texture (i.e., more troughs than peaks). Kurtosis can generally be used to describe the peakedness of a roughness profile (i.e., how severe the peaks and troughs are). In image process- ing, kurtosis can be used to describe the uniformity of the grayscale distribution. The calcula- tions for skewness and kurtosis (presented in Chapter 2 as Equations 15 and 16 and repeated here as Equations 26 and 27) are: ∑ ( ) = = , and (26) 3 1 3 R y n R sk ii n q ∑ ( ) = = . (27) 4 1 4 R y n R ku ii n q Maximum height is a simple measure of the distance between the maximum peak height in a detrended base length and the minimum valley distance in the same detrended base length. Percentile MPDs are calculated by finding not the maximum height in a half-base length, but rather a value near the maximum height. This approach is helpful if many outliers are suspected in a profile. Pavement texture heights in a detrended half-base length are sorted, and the nth percentile value is taken and averaged with the nth percentile value of the second half-base length. For this analysis, the 95th, 97th, and 99th percentiles were used. Tire contact length (TCL) measures the percentage of pavement that is in contact with the tire. This measurement is important for two reasons. First, it describes the percentage of the pavement that is taking advantage of the adhesive properties (a function of microtexture) of the pavement. Secondly, it helps to describe the percentage of pavement that is being used to evacuate water beneath the tire contact area. TCL is calculated by first deriving an enveloping profile using the Klein et al. (2004) method with tire stiffness coefficients given in Table 35. Next, the pavement profile points that coincide with the enveloping profile (to within a specified tolerance) are counted, and the ratio of coincident points to base length is taken as the TCL. Wavelets can be used as a method to approximate the characteristics of a measured wave- form. Wavelets are localized in both time and frequency, whereas the Fourier transformation is only localized in frequency. A short-period mother wavelet is selected and then scaled and Tire Stiffness d* Soft Tire 1E-2 Medium-soft Tire 1E-3 Medium-stiff Tire 1E-4 Stiff Tire 1E-5 Table 35. Tire stiffness coefficients used in tire enveloping profile procedure.

Data Analysis 107   shifted along the signal to decompose the signal into the wavelet transformation. The research team used the Haar mother wavelet, which is a square-wavelet made up of two complementary components—the difference in magnitude between two points (the details) and the average value (the amplitude data) of the two points. The wavelet decomposition is performed numerous times (levels) to the signal to gain insights about different scales of data decomposition that may shed light on wavelengths of interest for pavement surface properties. Each successive level is a wavelet decomposition of the previous level. The decomposition is complementary because the details and amplitude information of two successive levels can be used to reconstruct the previous level. The various levels of the details of the wavelet decomposition are independent (orthogonal) to one another and are therefore not collinear. The research team decomposed the raw data profile into 10 levels. As each of the data points is spaced at the minimum spatial distance obtainable by the measurement device (0.25 mm), the 10-level decomposition provides information up to the 0.25 × 210 = 256 mm scale. The details of each wavelet decomposition level were then processed to gather the following predictor parameters: RMS, Rsk, and Rku. This resulted in 30 predictor parameters derived from the three parameters calculated for each of the 10 levels of the wavelet decompositions. The enveloped profile MPD (MPDe) proposed by Goubert and Sandberg (2018) is the MPD taken of an enveloping profile. For this research, the enveloping profile was calculated using the method described by Klein et al. (2004), and the tire stiffness coefficients are given in Table 35. Next, an MPD was calculated for the enveloping profile according to the procedures in ASTM E1845 (2015). However, no filter was performed on the relatively smooth enveloping profile. Another parameter is the profile length ratio (PLR). The PLR is taken as the ratio between the length of the measured profile calculated via the sum of Euclidian distances between two data points in the base length and the base length. A perfectly flat and smooth surface would have a PLR of 1.0, and PLR increases with increasing texture height (see Equation 28). ∑ ( )( )= − + ∆+= − . (28) 1 2 2 1 1 PLR h h x base length i ii n Various peak data parameters were developed for this work using various peak height char- acteristics. Peaks were found using various capabilities of the “findpeaks” function in MATLAB (Mathworks 2016). Within the function, the minimum peak prominence can be specified. The findpeaks function measures the prominence of the peak due to its intrinsic height and location characteristics relative to other peaks, which helps to identify peaks that stand out by themselves. A low but isolated peak can have a greater prominence than a similar peak that is near another peak at a similar height. This is assessed by finding all peaks (any point with a height greater than the points immediately preceding and following the point) and then extending a horizontal line in the forward and reverse directions of the signal from each peak until the line crosses the signal (or reaches the end of the signal). The minima (see labels A and B in Figure 67) are then found within these two intervals of the horizontal line. The higher minimum value is used as a reference level, and the peak height above this reference level is calculated to establish the peak prominence. Any peak with a prominence greater than the user-specified value is then identified. The analysis for this report used peak prominence values of 0.1 and 0.25 mm. The base length used for all peak parameters was 100 mm, and each base length was detrended (slope removed and mean of base length set to zero) by subtracting its linear regression from itself. From this detrended profile, the following parameters were calculated for the minimum prominence levels specified: • MPGZ = mean peak height for peaks above the zero-mean line; • MSEPGZ = separation (in mm) between peaks in a base length;

108 Protocols for Network-Level Macrotexture Measurement • MPMSR = MPGZ MSEPGZ ; • MPROMGZ = mean of prominence for all peaks > zero; • MWGZ = mean of peak width (equal to half the peak prominence) for all peaks > zero; • MPWR = MPROMGZ MWGZ ; • MWMSR = MPGMWGZ MSEPGZ ; and • NPGZ = count of peaks above zero-mean line in a base length. After all parameters (3 predicted parameters and 79 predictor parameters) were calculated for each section of the road, data measures spaced less than 1 m were averaged to a length of 1 m via arithmetic means. Next, distances were aligned via linear interpolation. Minor differences in measured distance occur due to experimental error, such as distance measurement sensor calibration issues or interruptions to distance measurement, such as if the measurement device bounced and momentarily lost contact with the surface. The distances of the predicted variables were brought to match those of the predictor parameters (which were all the same, as the same non-contacting displacement device was used to measure the original pavement surface pro- file from which the variables were derived). In this way, the values are still valid 1 m samples (or, the interpolated value between two valid 1 m samples) but have been scaled to match the 1 m values recorded by the pavement profiler. 4.4.2 Single-Variable Regression After bringing all parameters to the same scale so equal numbers of samples could be evaluated, pairwise multivariate linear regression was performed on all datasets. In pair- wise multivariate linear regression, each possible combination of variables is correlated via least-squares regression and Pearson correlation coefficients (ρ) are calculated for each pair. More than 2,600 rows of 1-m data for each of the 82 variables (3 predicted variables and 79 predictor variables) were correlated, resulting in a 6,724-cell matrix of correlation coeffi- cients. Strengths of relationships could be visualized with this matrix and its corresponding scatterplot matrix. More importantly, the predicted variables’ correlation coefficients could be examined to find the best predictor parameters to predict the variables of interest. To simplify the results, the absolute values of the correlation coefficients were sorted, and the 10 largest values were taken as the best single-variable parameters to predict the parameter of interest. The results are summarized in Table 36 and Table 37 for the random-textured surfaces Figure 67. Peaks with minimum prominence values of 0.25 mm.

Data Analysis 109   SCRIM GT OBSI EAWE (filter, d* = 1E-2) -0.64 PLR (no filter) -0.69 Wd, RMS (Level 7, no filter) 0.34 PLR (no filter) -0.61 EAWE (filter, d* = 1E-2) -0.69 Wd, RMS (Level 9, no filter) 0.34 TCL (filter, tol. = 0.1, d* = 1E- 2) 0.61 Wd, RMS (Level 1, no filter) -0.68 Wd, RMS (Level 10, no filter) 0.34 EAWE (no filter, d* = 1E-2) -0.61 Wd, RMS (Level 2, no filter) -0.68 NPGZ (filter, P = 0.1) -0.34 Wd, RMS (lvl 1, no filter) -0.60 TCL (filter, tol. = 0.1, d* = 1E-2) 0.68 Wd, RMS (Level 8, no filter) 0.34 Wd, RMS (lvl 2, no filter) -0.59 EAWE (no filter, d* = 1E-2) -0.68 Wd, RMS (Level 6, no filter) 0.33 Wd, RMS (lvl 3, no filter) -0.59 Wd, RMS (Level 3, no filter) -0.68 RMS 0.33 Wd, RMS (lvl 4, no filter) -0.58 Wd, RMS (Level 4, no filter) -0.67 MPD95 0.32 MDE -0.57 MDE -0.67 MPD97 0.32 MPMSR (filter, P = 0.25) -0.56 MPMSR (filter, P = 0.25) -0.66 Max H 0.32 MPD -0.42 MPD -0.57 MPD 0.31 RMS -0.49 RMS -0.61 RMS 0.33 SCRIM 1 SCRIM 0.80 SCRIM 0.07 GT 0.80 GT 1 GT 0.02 OBSI 0.07 OBSI 0.02 OBSI 1 Tol. = Tolerance. Table 36. Single-variable Pearson correlation coefficients, random texture. SCRIM GT OBSI MWMSR (filter, P = 0.1) -0.22 TCL (no filter, tol. = 0.1, d* = 1E-4) 0.51 EAWE (filter, d* = 1E-2) 0.26 MPMSR (filter, P = 0.1) -0.20 MPDe (no filter, d* = 1E-4) -0.50 NPGZ (filter, P = 0.1) 0.24 MPMSR (filter, P = 0.25) -0.20 TCL (no filter, tol. = 0.1, d* = 1E-5) 0.48 MWGZ (filter, P = 0.1) -0.24 MWGZ (filter, P = 0.1) 0.20 MPMSR (filter, P = 0.25) -0.47 EAWE (filter, d* =e1E-2) 0.23 Wd, Rku (Level 4, no filter) 0.19 MPDe (no filter, d* = 1E-3) -0.46 MWGZ (filter, P = 0.25) -0.23 MSEPGZ (filter, P = 0.1) 0.18 TCL (no filter, tol. = 0.1, d* = 1E-3) 0.43 NPGZ (filter, P = 0.25) 0.23 Wd, Rku (Level 5, no filter) 0.18 MPDe (no filter, d* = 1E-5) -0.43 PLR (no filter) 0.21 MWGZ (filter, P = 0.25) 0.17 MWMSR (filter, P = 0.1) -0.43 MSEPGZ (filter, P = 0.1) -0.21 Wd, Rku (Level 6, no filter) 0.17 EAWE (filter, d* = 1E-5) -0.41 MPMSR (filter, P = 0.1) 0.21 NPGZ (filter, P = 0.1) -0.17 MWMSR (filter, P = 0.25) -0.41 TCL (filter, tol. = 0.1, d* = 1E-2) -0.21 MPD -0.13 MPD -0.38 MPD 0.16 RMS -0.12 RMS -0.31 RMS 0.17 SCRIM 1 SCRIM 0.14 SCRIM -0.09 GT 0.14 GT 1 GT -0.22 OBSI -0.09 OBSI -0.22 OBSI 1 Tol. = Tolerance. Table 37. Single-variable Pearson correlation coefficients, transverse texture.

110 Protocols for Network-Level Macrotexture Measurement (the HMA, OGFC, SMA, or the EP-5 or Cargill Safe Lane surface treatments) and transversely textured surfaces (the PCC surfaces which had been tined or grooved perpendicularly to the direction of travel), respectively. This kind of single-variable analysis is beneficial to agencies desiring to minimize the amount of calculations needed to predict friction or noise data from a pavement surface profile. Correlations between the predicted variable column and the MPD and RMS, as well as the other predicted variables, also are provided at the bottom of each table for reference. The best parameters to predict friction from the SCRIM or the GT from a pavement surface profile are the same for randomly-textured surfaces. Small differences in order based on cor- relation coefficients exist due to minor variations in the data. In general, the best performers for the single-variable correlation are those that account for the tire’s enveloping profile (i.e., EAWE and TCL) and the PLR, which is a measure of how much texture is provided by the pavement in comparison to a perfectly smooth surface. The RMS values of several levels of the details of the Haar wavelet decompositions also perform well, especially when compared to the RMS values given near the bottom of the table on a filtered, detrended profile. The MDE (a differencing algorithm similar to the details of the Haar wavelet decomposition on the raw pavement profile) and the MPMSR (a ratio parameter between the mean peak height of peaks with a prominence of at least 0.25 mm and their spacing) round off the list with correlation coefficients greater than 0.5. Notably, in the research for this project, the sign of correlation coefficient was positive in some cases and negative in others. For example, EAWE had a nega- tive sign, and TCL was positive. This likely reflects an inverse relationship between the two parameters. If the TCL is high, the tire will be in contact with more of the surface, leaving less space for an effective area of water evacuation. Correlation between the randomly-textured pavement surface profile and noise (as charac- terized by the overall sound pressure level measured by OBSI) is best with variations of the statistical measure of RMS or peak information such as MPD. The maximum height of peaks above the lowest valley or number of peaks above the mean profile (NPGZ) line of a 100 mm base length also outperforms the rest of the variables for noise prediction. The correlation coefficients, however, were low in this research for all measures evaluated, with no more than a 34% strength of relation. It is interesting to note that each of the top 10 predictor parameters in Table 36 outperforms the most commonly used pavement macrotexture characterization parameters of MPD and RMS; however, correlation coefficients are still relatively small. The reader should also note that the Pearson correlation coefficient between the SCRIM and GT (two devices that both characterize pavement friction via a rubber tire in contact with the road and lubricated by a thin film of water) is 0.8. Despite differences in operating principles (e.g., slip speed, normal loads used, water flow rates), one can expect any given predictor parameter to have a maximum correlation in a similar range given that the two devices should, in theory, measure wet- weather friction. The correlations for all predictor variables for SCRIM and OBSI were all low for the trans- verse textured pavements; however, Table 37 still proves useful, as it shows each of the top 10 predictor variables outperforms the most commonly used parameters to describe a pavement’s macrotexture (MPD and RMS). This comparison shows that improvements can be made to these common predictors to enable better pavement management in the future. Stronger rela- tionships were found for GT data by several proposed macrotexture parameters, with the best being those that account for the deformed shape of the vehicle’s tire into the pavement texture: TCL, MPDe, and peak analysis-based parameters that account for the separation of peaks above a certain prominence or the widths associated with these peaks.

Data Analysis 111   4.4.3 Multiple Regression Each of the aforementioned predictor parameters brings its own strengths to the single- variable linear regression. For example, some parameters characterize the peaks of the base length evaluated, whereas others give various measures of the enveloped tire profile on the surface, and still others evaluate the statistical measures of the base length at various wavelet- decomposed scales. The next objective of this work was to determine if the power of the model developed to predict friction or noise from surface profile measurements could be increased. This could be done by taking into consideration several variables to relate pavement surface properties to the desired objective predicted variable. Multiple linear regression results in additional coefficients for the linear model. A similar approach was taken by Sohaney and Rasmussen (2013) on their analysis of 3D macrotexture data measured by a walking-speed, semi-autonomous vehicle. This type of data stream is not yet available for high-speed, network-level analysis, however; therefore, in this work, single-spot laser triangulation data were used, as these data are commonly available. In the analysis for this report, the additional parameters of pavement geometry (grade, crossfall, and curvature) were also added to the predictor parameters outlined in Table 34 for the multiple linear regression. These measurements could be collected with the same vehicle gathering the pavement surface profile by the addition of an inertial unit. Selection of Variables One approach to arrive at a multiple regression model would be to include all predictor variables in the model. This approach would, however, pose several significant problems for the analysis. Chiefly, the complexity of the final model and the need for calculation of additional predictor parameters would greatly increase, possibly without any benefit to the final model. Too many parameters will most often result in a model that is overfitted. The predictor variables also could bring multicollinearity issues to the model, thereby falsely inflating confidence in the final model. To select the best parameters for the final model and avoid overfitting, a least absolute shrinkage and selection operator (LASSO) model was developed using JMP software. LASSO is a minimization problem similar to basis pursuit denoising as used in signals processing. The approach (in the least-squares case, as was used in the following analysis and depicted in Equation 29) forces regression coefficients to be less than a fixed value (t), as shown in Equation 29 (Tibshirani 1996). In so doing, some coefficients are forced down to zero, which removes them from the final model. This approach is also similar to ridge regression; however, ridge regression only reduces the magnitude of the regression coefficients. It does not set some to zero (remove them) based on predetermined criteria. ∑ ∑( )− β − β     β ≤ β β = = min 1 , (29) , 2 1 1N y x such that ti o iTi N jj p o where N = number of cases (observations), p = number of predictor variables, y = outcome (predicted variable), x = the covariate vector of predictor variables for the ith case, and 1N = a vector of ones.

112 Protocols for Network-Level Macrotexture Measurement Variance Inflation If the predictor variables are not independent (i.e., collinear), several regression parameters will be found to be solutions to the minimization problem described above. For this work, to avoid collinearity, variance inflation factors (VIFs) were considered. VIFs were calculated in a pairwise fashion for all parameters selected to be the LASSO analysis. VIFs are calculated as inverse correlations of the ith pair of variables, as seen in Equation 30: = − 1 1 , (30) 2 VIF R i i where Ri2 = coefficient of variation of the ith pair of variables. In general, VIFs below 4 are considered appropriate (Hair et al. 2011). VIFs between 5 and 10 indicate a relatively high correlation between predictor variables, and values above 10 will likely produce poor estimates due to stronger multicollinearity. VIF values for variables selected by the LASSO analysis were sorted, and the variable with the highest VIF was removed. VIFs were then again calculated for the model, and the process was repeated until all remaining variables had VIFs lower than 4. After valid parameters were selected for the model, the dataset was divided for cross- validation by designating 8 of 10 data points to a training dataset, the remainder being reserved for validation. Models were formed using the training dataset and cross-validated using the validation dataset. Table 38 and Table 39 summarize the results for the random and transversal textures, respectively. Results The tables show that the use of multiple regression increased Pearson correlation coefficients by 10%–20% for random-textured pavements when compared to single-variable regression Tol. = Tolerance. SCRIM GT OBSI Intercept Wd, Rku (Level 3, no filter) NPGZ (filter, P = 0.25) MPMSR (filter, P = 0.1) MWMSR (filter, P = 0.1) Grade 145.06 -2.26 -1.49 -177.1 -59.33 2.19 Intercept Rsk TCL (filter, tol. = 0.1, d* = 1E-5) Wd, Rku (Level 3, no filter) Wd, Sku (Level 3, no filter) Wd, Sku (Level 7, no filter) MPMSR (filter, P = 0.25) NPGZ (filter, P = 0.25) MWMSR (filter, P = 0.1) Grade 1.48 -0.05 -0.08 -0.02 0.05 0.04 -1.94 -0.01 -0.73 0.01 Intercept TCL (filter, tol. = 0.1, d* = 1E-2) MWMSR (filter, P = 0.1) Grade 103.30 0.08 -4.27 0.40 Model ρ (training) 0.75 0.77 0.41 Model ρ (validation) 0.75 0.76 0.38 SCRIM Mean 79.09 0.74 102.70 SCRIM St Dev 6.76 0.07 0.96 Model RMSE (training) 4.52 0.05 0.88 Model RMSE (validation) 4.48 0.05 0.89 Table 38. Summary of model coefficients, r, and RMSE values, random texture.

Data Analysis 113   analysis. The friction-related predicted values (SCRIM and GT) both included the peak-related predictors of NPGZ, MPMSR, and the kurtosis of the details of a Haar wavelet transform as common predictors. The GT data benefited from taking into account the tire’s interaction with the surface in the form of TCL. The noise-related parameter of OASPL (as measured by OBSI) included the peak-related parameter dealing with the ratio of the mean peak width to the mean peak spacing (MWMSR) and the influence of the deformed tire into the pavement surface (TCL). Each of the predicted parameters benefited from including the road geometry measure of grade. The ρ and root-mean square error (RMSE) of the training and validation sets were very close to one another, indicating good cross-validation of the model. Prediction of SCRIM and noise (OASPL) data as measured by the OBSI proved again to be difficult on transversely-texture pavements. Pearson correlation coefficients increased dramati- cally compared to single-variable regression coefficients; however, ρ remained at or below 0.5 for these two parameters. Prediction of GT from a pavement surface profile (as measured by ρ) rose by 35% when compared to single-variable regression. Taking tire envelopment (TCL and EAWE) into account, as well as aggregate width and separation information, resulted in higher ρ for transversely textured pavements. Road geometry information such as grade, cross fall, and curvature also was found to be significant in the model. RMSE and correlation coefficients of the training and validation sets were very close to one another, indicating good cross-validation of the model. 4.4.4 Aggregation of Predictor Variables Inclusion of multiple parameters to describe a pavement’s macrotexture resulted in an appre- ciable increase in the ability to predict all three predictor variables (SCRIM, GT, and OBSI). These models were built and validated on 1 m data taken from the Virginia Smart Road. The 100 mm pavement macrotexture measurements were aggregated to 1 m data via arithmetic means to match the minimum reporting length of the devices to be predicted. However, most state transportation officials do not store or analyze data based on 1 m data. Linear referencing and required storage space for the data begin to become problematic for tens of thousands of kilometers of road data. Furthermore, roads are not managed on a meter- by-meter basis. It would be rare for a project to be initiated based on a single 1 m variable, and Tol. = Tolerance. SCRIM GT OBSI Intercept Wd, Rku (Level 4, no filter) Wd, Rku (Level 10, no filter) Wd, Rsk (Level 1, no filter) MWGZ (filter, P = 0.1) Grade Curvature 71.54 0.68 1.84 -1.50 0.41 -0.30 545.10 Intercept TCL (filter, tol. = 0.1, d* = 1E-5) EAWE (filter, d* = 1E-2) MWMSR (filter, P = 0.25) MWGZ (filter, P = 0.25) Crossfall Curvature 1.06 0.09 0.0002 0.001 -0.47 0.01 -10.72 Intercept TCL (filter, tol. = 0.1, d* = 1E-3) Wd, Rsk (Level 4, no filter) MWMSR (filter, P = 0.25) Grade Crossfall 101.20 -0.81 -1.01 3.44 0.08 -0.19 Model ρ (training) 0.31 0.69 0.53 Model ρ (validation) 0.31 0.65 0.59 SCRIM Mean 81.02 0.74 104.1 SCRIM Standard Deviation 2.75 0.04 1.11 Model RMSE (training) 2.64 0.03 0.94 Model RMSE (validation) 2.64 0.03 0.97 Table 39. Summary of model coefficients, r, and RMSE values, transverse texture.

114 Protocols for Network-Level Macrotexture Measurement road projects are always carried out on a much larger scale, not through the replacement of a single meter of pavement. Therefore, it was of particular interest to aggregate the data (predicted and predictor variables) in this experiment to test the models developed. Predicted and predictor variables were aggregated by taking arithmetic means of the data to arrive at data at 1 m, 3 m, 10 m, and 20 m intervals. The proposed multivariable linear models were then applied to the aggregated datasets of predictor variables, and correlation coefficients were cal- culated on the predicted variables for the SCRIM, GT, and OBSI. The results of this aggregation are given in Table 40 and Table 41. The data presented in Table 40 and Table 41 represent the correlation coefficients and RMSE to SCRIM, GT, and OASPL measured by OBSI that agencies can expect if the models proposed in Table 38 and Table 39 are employed at the reporting interval maintained by the agency for their road profile data. As is the case for all the data presented in this work, these values apply to road networks made up of pavement surfaces similar to those represented in this experiment. In general, all correlation coefficients improve with greater aggregated distance data. This is because local changes in macrotexture on smaller scales are attenuated by surrounding values in the aggregation process. In this research, the RMSE values decreased with larger aggregations for SCRIM and OBSI, whereas values remained essentially constant for the GT. This disparity may be explained by the relatively large scale of SCRIM and OBSI values, which were two orders of magnitude larger than GT values. 1 m 3 m 10 m 20 m SC R IM Model ρ 0.75 0.79 0.84 0.86 Model RMSE 4.52 4.10 3.65 3.45 SCRIM Mean 79.09 79.11 79.06 79.17 G T Model ρ 0.76 0.79 0.88 0.94 Model RMSE 0.05 0.04 0.03 0.03 GT Mean 0.74 0.74 0.74 0.74 O BS I Model ρ 0.41 0.47 0.55 0.58 Model RMSE 0.88 0.76 0.64 0.60 OASPL Mean 102.67 102.68 102.68 102.68 Table 40. Summary model performance, various distance aggregations, random texture. 1 m 3 m 10 m 20 m SC R IM Model ρ 0.29 0.39 0.45 0.54 Model RMSE 2.63 2.01 1.74 1.54 SCRIM Mean 81.02 81.02 81.01 81.07 G T Model ρ 0.66 0.73 0.80 0.88 Model RMSE 0.03 0.03 0.02 0.02 GT Mean 0.74 0.74 0.74 0.74 O BS I Model ρ 0.54 0.60 0.70 0.76 Model RMSE 0.94 0.81 0.65 0.58 OASPL Mean 104.12 104.12 104.12 104.11 Table 41. Summary model performance, various distance aggregations, transverse texture.

Data Analysis 115   4.4.5 Main Findings The analysis of the various macrotexture parameters led to the following findings: • The top 10 single-variable predictor parameters for random, transverse, and longitudinally textured pavements correlated better with friction and/or noise than the most commonly used predictor parameters (MPD and RMS). – For randomly-textured pavement (i.e., asphalt), the best single-variable predictors of friction (as measured by the SCRIM and GT) are those that account for the tire’s enveloping profile (i.e., EAWE and TCL) and the PLR, which is a measure of how much texture is provided by the pavement in comparison to a perfectly smooth surface. The maximum Pearson correlation coefficients for randomly-textured pavements for the SCRIM and GT were 0.64 and 0.69, respectively. – For transversely textured pavements, correlations were lower; however, several of the tested parameters still correlated better with friction and/or noise than the commonly used measures MPD and RMS. • For randomly-textured surfaces, the use of multiple regression increased Pearson correlation coefficients by 10%–20% when compared to single-variable regression analysis. – GT and SCRIM data were better predicted when the researchers included both peak data (including NPGZ and MPMSR) and kurtosis of the details of a Haar wavelet transform as common predictors. The GT data benefited from taking into account the tire’s interaction with the surface in the form of TCL. – The noise-related parameter of OASPL as measured by OBSI included the peak-related parameter dealing with the ratio of MWMSR and the influence of the deformed tire in relation to the pavement surface (TCL). – Each of the predicted parameters benefited from including the road geometry measure of grade. – The correlation coefficients and RMSE of the training and validation sets were very close to one another, indicating good cross-validation of the model. • Aggregation of pavement macrotexture data into longer reporting lengths is favorable for both the confidence in model coefficients and effective management of road networks. In summary, the analysis showed that several existing and newly developed macrotexture parameters were shown to outperform the de facto standard of MPD or RMS in single-variable linear regression models; however, alternative parameters to MPD and RMS should undergo continued testing on a greater variety of pavement surfaces before adoption as standards. Furthermore, correlation coefficients also improved by using multiple linear regression (care was taken to avoid multicollinearity by controlling the VIF), but additional testing should be conducted to determine if the improvement justifies the additional complexity. 4.5 Operational Conditions The last analysis evaluated the impact of various operational conditions on the measured macrotexture parameters. 4.5.1 Various Constant Speeds For the constant-speed tests described in Table 13, an ANOVA was conducted for both pavement sections with speed as a continuous variable model input and MPD as the response. For the variable-speed tests, an ANOVA was conducted with the four various acceleration conditions as the model input and MPD as the response. In both cases, using a significance level

116 Protocols for Network-Level Macrotexture Measurement of 0.05, a failure to reject the null hypothesis would demonstrate no effect of variable speeds, and rejection of the null hypothesis would demonstrate an effect of the variable speeds. To analyze the effect of speed (24 km/h to 105 km/h in 16-km/h increments) on the devices tested, an ANOVA was performed for each device on the two sections tested. A significance level of 0.05 was selected, above which the analysis failed to reject the null hypothesis that speed does not affect the mean 1 m MPD calculated. The results are presented in Table 42. Recall that Section L2 is a dense-graded HMA, and PCC1a is a transversely tined PCC section. As Table 42 shows, the ANOVA results for Device 3 and Device 5 rejected the null hypothesis that speed has no effect on the MPD calculated. Device 3 and Device 5 have lower sampling frequencies of 32 kHz and 5 kHz, respectively. Device 5 is a line laser and captures over 300 points transversely 5,000 times per second. Figure 45(a) shows a clear linear trend with MPD negatively correlated to speed. The range of the data, however, is entirely different for the two pavement types tested. For Section L2 (dense-graded HMA), the range of MPDs through the speeds was from 1.24 down to 0.99 mm, for a spread of 0.25 mm. For the transversely tined section (PCC1a), the data ranged from 0.68 down to 0.61 (a spread of 0.07 mm), which fell within the range of good repeatability for Device 1. Because the analysis had rejected the null hypothesis that means are equal for Device 3 and Device 5, statistical evidence showed that speed does affect the reading of average MPD for these devices. However, as the range of data for the PCC section was within the range of repeatability coefficients for both devices, speed may not have a practical effect on these devices on PCC. As the range of data for the asphalt sections was outside of the range of repeatability coefficients for both devices, the conclusion was that speed does have an effect on these devices (both with lower sampling frequencies) on asphalt pavements. Device 5 had the same negative correlation as Device 3. It also shared similar range charac- teristics, with a spread of only 0.054 mm (MPDs from 0.60 down to 0.54 mm) for the PCC1a section. For the Asphalt Section L2, the MPDs ranged from 1.32 down to 1.20 mm. This spread (0.129 mm) was larger than the repeatability of the device. When the raw profiles were inves- tigated, it was noted that the profiles were smoother and more closely resembled a flat line as speed increased. This effect could be due to an averaging effect on the signal with increased vehicle speed and constant exposure time on the device’s sensor. Averaging peaks and valleys Device Section p-value 1 L2 0.3854 2 L2 0.5966 3 L2 0.0009 4 L2 0.3977 5 L2 0.0001 1 PCC1a 0.0607 2 PCC1a 0.2206 3 PCC1a 0.0001 4 PCC1a 0.4951 5 PCC1a < 0.0001 Table 42. Summary of ANOVA results for the constant-speed experiment.

Data Analysis 117   forms flatter surfaces, as can be observed in Figure 68. This indicates that not only does speed have an influence, but also that surface type tends to either magnify or reduce this effect. 4.5.2 Variable Speed The effects of acceleration and deceleration (speeding up and braking) described in Table 13 were evaluated similarly to the constant-speed experiment. Table 43 presents the results of an ANOVA with four acceleration conditions as the model input and MPD as the response. Only the device with a line laser rejected the null hypothesis that acceleration does not affect the calculated MPD. Box plots of the processed data collected by this device are shown in Figure 69. Processing included outlier removal, low-pass filtering, profile detrending, and calculation of MPD. Although the null hypothesis was rejected, it should be noted that the data ranged from 1.23 down to 1.17 mm (a spread of only 0.06 mm), which fell within the range of good repeat- ability for the device. Because the results failed to reject the null hypothesis for the other devices, there is statistical evidence that the acceleration and deceleration maneuvers tested do not affect the reading of average MPD for single-spot devices. Furthermore, the good repeatability shown for the line laser indicates that the ANOVA may be penalizing the line-laser data due to its low variance. In other words, with low variability, a small deviation from the mean causes a rejection of the null hypothesis that all means are equal. Acceleration of the test device was found to be statistically significant for the line laser; however, given that the range of 1 m MPDs (0.06 mm) CS = Constant speed. Figure 68. Smoothing of line-laser readings as speed increases. Device p-Value 1 0.4653 2 0.6184 3 0.4306 4 0.8423 5 0.0051 Table 43. Summary of ANOVA for variable- speed experiment.

118 Protocols for Network-Level Macrotexture Measurement Speed (MPH) 25_50 50_25 50_0_50 50_25_50 1.17 1.18 1.19 1.20 1.21 1.22 1.23 AV G M PD Figure 69. Processed data collected by Device 5 during the variable-speed experiment. for the variable-speed test fell within the device’s coefficient of repeatability, the effect was found to be negligible. 4.5.3 Main Findings The analysis resulted in the following findings: • In general, speed did not have a practically significant effect on the measurement of macro- texture parameters. This finding applied to all devices except for one device that was equipped with a lower frequency laser. • Similarly, neither acceleration nor deceleration produced significant practical differences in the measured macrotexture parameters for any of the devices tested.