Traffic Sign Management: Data Integration and Analysis Methods for Mobile LiDAR and Digital Photolog Big Data

Transcription

1 Utah State University All Graduate Theses and Dissertations Graduate Studies 2016 Traffic Sign Management: Data Integration and Analysis Methods for Mobile LiDAR and Digital Photolog Big Data Majid Khalilikhah Utah State University Follow this and additional works at: Part of the Civil Engineering Commons, and the Transportation Engineering Commons Recommended Citation Khalilikhah, Majid, "Traffic Sign Management: Data Integration and Analysis Methods for Mobile LiDAR and Digital Photolog Big Data" (2016). All Graduate Theses and Dissertations. Paper This Thesis is brought to you for free and open access by the Graduate Studies at It has been accepted for inclusion in All Graduate Theses and Dissertations by an authorized administrator of For more information, please contact

2 TRAFFIC SIGN MANAGEMENT: DATA INTEGRATION AND ANALYSIS METHODS FOR MOBILE LIDAR AND DIGITAL PHOTOLOG BIG DATA by Majid Khalilikhah A dissertation submitted in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY in Civil and Environmental Engineering Approved: Dr. Kevin Heaslip Major Professor Dr. Guifang Fu Committee Member Dr. Ziqi Song Committee Member Dr. John Rice Committee Member Dr. Laurie McNeill Committee Member Dr. Mark McLellan Vice President for Research and Dean of the School of Graduate Studies UTAH STATE UNIVERSITY Logan, Utah 2016

3 ii Copyright Majid Khalilikhah 2016 All Rights Reserved

4 iii ABSTRACT Traffic Sign Management: Data Integration and Analysis Methods for Mobile LiDAR and Digital Photolog Big Data by Majid Khalilikhah, Doctor of Philosophy Utah State University, 2016 Major Professor: Dr. Kevin Heaslip Department: Civil and Environmental Engineering This study links traffic sign visibility and legibility to quantify the effects of damage or deterioration on sign retroreflective performance. In addition, this study proposes GIS-based data integration strategies to obtain and extract climate, location, and emission data for in-service traffic signs. The proposed data integration strategy can also be used to assess all transportation infrastructures physical condition. Additionally, nonparametric machine learning methods are applied to analyze the combined GIS, Mobile LiDAR imaging, and digital photolog big data. The results are presented to identify the most important factors affecting sign visual condition, to predict traffic sign vandalism that obstructs critical messages to drivers, and to determine factors contributing to the temporary obstruction of the sign messages. The results of data analysis provide insight to inform transportation agencies in the development of sign management plans, to identify traffic signs with a higher likelihood of failure, and to schedule sign replacement. (129 pages)

5 iv PUBLIC ABSTRACT Traffic Sign Management: Data Integration and Analysis Methods for Mobile LiDAR and Digital Photolog Big Data Majid Khalilikhah Traffic signs often convey critical information to drivers. However, traffic signs are only effective when clearly visible and legible. This study aims to determine the effects of various damage and deterioration forms on sign retroreflectivity, identify the most important factors affecting traffic sign visual condition, predict traffic sign vandalism that obstructs critical messages to drivers, and identify important environmental factors contributing to the temporary obstruction of the sign messages. To do so, two data sets are used. A sample data of over 1,700 signs was manually collected in the field, and the background retroreflectivity of each sign was measured using a handheld retroreflectometer. In addition, sign data of over 97,000 traffic signs was digitally collected by driving an equipped vehicle. Sign visual condition and damage/deterioration data were obtained from inspection of daytime digital images taken of each individual sign. A GIS-based strategy is proposed to extract location and climate data for every individual sign. Various statistical tests and models are also used to accomplish the goals of study.

6 v DEDICATION I dedicate my dissertation work to my parents. My loving parents are very special. They have never left my side. I also dedicate this to my siblings for always being there for me. Majid Khalilikhah

7 vi ACKNOWLEDGMENTS Foremost, I would like to express my deepest thanks to my advisor, Dr. Kevin Heaslip, for his continuous support, motivation, enthusiasm, and for his friendship. Dr. Heaslip is a truly special person. I have been extremely lucky to have a supervisor who cared so much about not only my work, but my life and helped me map out my future. I could not have imagined having a better advisor for my Ph.D. studies. Besides my advisor, I would like to thank the rest of my dissertation committee: Dr. Guifang Fu, Dr. Laurie McNeill, Dr. John Rice, and Dr. Ziqi Song for their insightful comments and questions. Especially, my appreciation goes to Dr. Fu for helping me through my studies. This study has also benefitted from discussion with Dr. David Tarboton and Dr. Kathleen Hancock. Their assistance is gratefully acknowledged. I thank Wesley Boggs, Travis Evans, and Kevin Gardiner, all of whom collected the manual data. I would also like to express thanks to Mandli Communications Inc. for collecting mobile-based data. This research was conducted with the help and support from the Utah Department of Transportation. I would like to convey my gratitude to UDOT for sponsoring the data collection efforts. Additionally, I thank my friends Navid Zolghadri, Hadi Malek, Sadra Sharifi, and Mehrdad Shahabi for their support and friendship. Majid Khalilikhah

8 vii CONTENTS Page ABSTRACT... iii PUBLIC ABSTRACT... iv DEDICATION... v ACKNOWLEDGMENTS... vi LIST OF TABLES... x LIST OF FIGURES... xii CHAPTER INTRODUCTION Introduction Sign Assessment and Management Methods Research Questions and General Approach Data Collection Manual Method Mobile-based Method Anticipated Contributions Research Outline CHAPTER THE EFFECTS OF DETERIORATION ON TRAFFIC SIGN VISIBILITY Introduction Data Analysis Sign Deterioration Deterioration Categories Type I Signs Type III and III HIP Signs Type IX and XI Signs In-service Sign Age Linear Regression Polynomial Regression Discussion Conclusions CHAPTER

9 IMPORTANT FACTORS AFFECTING TRAFFIC SIGN VISUAL CONDITION Introduction Traffic Sign Attributes Sign Length and Width Sign Mount Height Sign Facing Direction Sign Background Color Climate and Location Data Integration Ground Elevation Temperature and Precipitation Land Cover Solar Statistical Methodology Decision Tree Random Forests Variable Importance Measure (VIM) Results and Discussion Conclusions CHAPTER PREDICTION OF TRAFFIC SIGN VANDALISM THAT OBSTRUCTS CRITICAL MESSAGES TO DRIVERS Introduction Vandalism by Demographics Data Population Density Ethnicity Age Income Education Gender Regression Model Vandalism by Sign Attributes Sign Background Color Sign Length and Width Sign Mount Height Exposure Road Type Discussion viii

10 4.4 Conclusions CHAPTER ENVIRONMENTAL FACTORS CONTRIBUTING TO THE TEMPORARY OBSTRUCTION OF THE SIGN MESSAGES Introduction Explanatory Variables Sign Mount Height Sign Facing Direction Ground Elevation Exposure Precipitation Wind Statistical Model Results Conclusions CHAPTER CONCLUSION AND FUTURE WORK REFERENCES ix

11 x LIST OF TABLES Table Page 1-1 Minimum retroreflectivity levels (cd/lx/m 2 ) Summary of surveyed sign attributes using manual and mobile methods Retroreflectivity specifications for Type I signs Retroreflectivity specifications for Type III signs (cd/lx/m 2 ) Performance of Type III aging vs. non-aging signs Retroreflectivity specifications for Type III HIP signs Retroreflectivity specifications for Type IX signs Retroreflectivity specifications for Type XI signs Retroreflectivity specifications for green signs Regression models of signs retroreflective measurements Odds ratio for failure rates Sign condition by length Sign condition by width Sign condition by mount height Sign condition by facing direction Sign condition by color Sign condition by elevation : Sign condition by average annual temperature : Sign condition by average annual precipitation Sign condition by land cover...56

12 xi 4-1 Damaged signs by county Demographic characteristics by county in Utah Vandalism rates by population density Vandalism rates by ethnic majority Vandalism rates by median age Vandalism rates by income Vandalism rates by education level Vandalism rates by gender Regression models of sign vandalism rate Summary of vandalized signs by MUTCD type Traffic sign vandalism by color Traffic sign vandalism by width Traffic sign vandalism by length Traffic sign vandalism by mount height Traffic sign vandalism by sign exposure Traffic sign vandalism by road type Dirty traffic signs by sign mount height Dirty traffic signs by sign direction Dirty traffic signs by ground elevation Dirty traffic signs by sign exposure Dirty traffic signs by precipitation Dirty traffic signs by wind...101

13 xii LIST OF FIGURES Figure Page 1-1 Locations of traffic signs collected manually Snapshot of the process of manual sign data collection LiDAR equipped vehicle used for UDOT sign data collection LiDAR imaging and associated photolog Taking image from traffic signs Locations of mobile-based traffic signs Summary of surveyed signs by type and color Deterioration categories Summary of surveyed signs with known age Linear Regression Models for Green Signs Sign retroreflectivity failure by deterioration categories Process of the study Examples of traffic signs rated as (a) Good; (b) Fair; (c) Poor Locations of traffic signs by elevation (m) Locations of traffic signs by average annual temperature ( C) Locations of traffic signs by average annual precipitation (mm) Locations of traffic signs by land cover Locations of traffic signs by solar resource (kwh/m 2 /Day) Variable importance plot for traffic sign s visual condition ratings...64

14 xiii 4-1 Samples of traffic sign vandalism Locations of vandalized traffic signs in Utah Population density by county Ethnicity by county Median age by county Household income by county Education by county Number of vandalized traffic signs by legend type Municipalities in Utah Vandalized signs by mount height vs. exposure and road type Examples of dirty signs Locations of dirty traffic signs in Utah : Locations of traffic signs by wind power classes Variable importance ranking for traffic sign dirt...104

15 CHAPTER 1 INTRODUCTION 1.1 Introduction A variety of factors associated with drivers, vehicles, and the roadway contribute to the likelihood of crashes (Pour-Rouholamin, et al., 2015; Baratian-Ghorghi, et al., 2015). Transportation agencies continually make efforts to design safety improvements that will reduce such outcomes, with particular concern for fatal and serious injury crashes (Sadauskas, 2003; Ivan, et al., 2012; Yannis, et al., 2014; Jalayer, et al., 2015; Baratian- Ghorghi & Zhou, 2016; Pour-Rouholamin & Zhou, 2016). Traffic signs, specifically warning and regulatory signs, are an element of transportation infrastructure which can provide critical safety related information to drivers and other users. However, these signs are only effective when clearly visible. Generally, sign replacement is a low-cost safety treatment helping drivers navigate roads in a safer and more efficient manner through removal of ineffective signs (McGee, 2010). To better understand traffic signs, one must take factors such as visibility and retroreflectivity into account (Boggs, et al., 2013). By incorporating sheeting made of retroreflective material, even signs that are not illuminated by external lights are still visible at night. Retroreflection works by redirecting light from the sign face back to the source (McGee, 2010). The U.S. Congress first introduced standards for minimum levels of sign retroreflectivity to the Secretary of Transportation (United States Department of

16 2 Transportation, 1992): The goal of the new minimum retroreflectivity requirements was to improve safety on our nation s streets and highways, and was meant to ensure that drivers, especially the elderly, would be able to detect, comprehend, and react to traffic signs accordingly and help to facilitate safe, uniform, and efficient travel. (Re & Carlson, 2012). To fulfill that mandate, in 2009, the Manual on Uniform Traffic Control Devices (MUTCD) established minimum retroreflectivity standards for traffic signs, including an obligation for agencies to replace signs that were not in compliance with these levels. The coefficient of retroreflectivity, RA, commonly referred to as retroreflectivity, is the ratio of a sign s luminance to its illuminance. The MUTCD outlined five methods that would guide agencies in achieving and maintaining minimum RA levels. These methods included assessment methods (visual nighttime inspection and sign retroreflectivity measurement) and management methods (expected sign life, blanket replacement, and control signs). In 2012, final revisions were adopted to MUTCD that changed the three original target compliance dates for minimum retroreflectivity levels (MUTCD, 2012). While retroreflectivity efficiency ensures visibility of the traffic sign, it does not properly describe the legibility of the sign. The existence of damage and deterioration on a traffic sign s face causes a decline in the overall legibility of the sign during both day and nighttime conditions (Boggs, et al., 2013). The effects of sign damage on the legibility of traffic signs vary considerably with respect to the form of damage (Ng & Chan, 2008). In addition, sign damage and deterioration on the faces may lead to the decline in sign retroreflective performance.

17 3 One of the challenges in fulfilling MUTCD mandate is collection of sign data due to the sheer size of sign inventories. Accurate data are important as they serve as the basis for cost efficient and compliance effective strategies. To address the data challenge, the Utah Department of Transportation (UDOT) has sponsored field investigations by a team of researchers at Utah State University to investigate the effectiveness of data collection techniques. Over the course of the past few years, the team has developed data taxonomies, field collection methods, and post-collection analysis methods. Early research included the development of standard practices for measuring retroreflectivity with handheld retroreflectometers. The lessons learned in those early efforts were incorporated into a data collection system that included the use of mobile computing devices to capture key parameters. Data from this effort include were collected on over 1700 signs located across the state s major climatological regions in both rural and urban environments. While the prior work produced useful findings to further the body of knowledge for measurement techniques and data requirements, it did not address the issue of scale. To address this, a third phase of research was conducted leveraging a mobile-based sign data collection effort that examined over 97,000 signs. The hypothesis was that a mobilebased method could be used to cost effectively collect data on a large number of signs across a large area. Employing manual and mobile-based in-service traffic sign data that were collected in the field, this study discusses the effects of various forms of damage and deterioration on traffic sign visibility, determines the most important contributing factors to sign visual condition, identifies which traffic signs are more vulnerable to

18 4 vandalism, and predicts the contributing factors to traffic sign dirtiness. To do so, Geographic Information System based (GIS-based) data integration and non-parametric machine learning methods are utilized. 1.2 Sign Assessment and Management Methods The minimum retroreflectivity levels established by the MUTCD require transportation agencies to develop and adopt assessment or management methods to maintain retroreflective compliance with minimum levels. The coefficient of retroreflectivity, RA, commonly referred to as retroreflectivity, is the ratio of a sign s luminance to its illuminance and is measured with the unit of (cd/lx/m 2 ). Each individual assessment and management method outlined in the MUTCD has its own strengths or weaknesses. In general, assessment methods evaluate each individual sign of the inventory on a periodic basis to determine whether or not it meets minimum retroreflectivity levels. In contrast, management methods categorize signs based on their attributes, such as color, sheeting type, age, and geographic conditions, in order to predict their retroreflective degradation over time without involving evaluation of each in-service sign. The following is a summary of the methods outlined in the MUTCD: Visual Nighttime Inspection (assessment method) is not dependent on installation dates and an existing sign inventory database. Also, the invested resources are absolutely minimal. However, the consistency of the method may be questioned, since the accuracy of inspections dramatically depends on the amount of training received by inspectors. A study showed that

19 method accuracy varied significantly between agencies with respect to the amount of training provided (Rasdorf et al., 2006). 5 Retroreflectivity Measurement (assessment method) requires using a retroreflectometer to directly measure the sign retroreflectivity value. The retroreflectivity measurements must be compared to the standard MUTCD levels. Table 1-1 shows the minimum retroreflectivity levels required by the MUTCD. In this table, the retroreflectivity sheeting types are defined with respect to the standard specification in American Society for Testing and Materials (ASTM) D Although the assessment rates can vary depending on spatial distribution of the signs, this method definitely requires spending a significant number of person-hours. Also, the location of the retroreflectivity measurement on traffic signs is an important factor affecting the accuracy of the method. In addition, crew safety, weather conditions, and measuring retroreflectivity of overhead signs are major issues with this method. Expected Life (management method) is heavily dependent on installation date information. Using the expected sign life method, signs are replaced before their retroreflectivity degrades below minimum levels. Such sign replacement may be established based on different factors, such as manufacturers warranties and retroreflective deterioration forecasting. However, conducting the expected sign life management method might be questioned in terms of its cost-effectiveness, since the deterioration of

20 retroreflective sheeting in different environments is not the same, and actual sign service life might be longer than its warranty. 6 Sign Color White on Green Black on Yellow or Black on Orange White on Red Black on White Table 1-1 Minimum retroreflectivity levels (cd/lx/m 2 ) Sheeting Type (ASTM D ) Prismatic Beaded Sheeting Sheeting Additional Criteria III, IV, VI, VII, I II III IX, X W W W W 250; G 15 Overhead G 7 G 15 G 25 W W 120; G 15 Ground-mounted G 7 Y; Text and fine symbol signs Y 50; O 50 O measuring at least 1200 mm Y; Text and fine symbol signs Y 75; O 75 O measuring less than 1200 mm W 35; R 7 Sign constant ratio 3:1 W Blanket Replacement (management method) is similar to the expected life method, except it is based on geographical area, corridor, or sheeting type and color. Thus, this method is not dependent on sign installation date and is simple for an agency to implement. However, using this method, inefficiency arises, due to high variance in expected sign deterioration levels. Control Signs Method (management method) uses a sample set of in-field signs to determine the life of signs. To assess the performance of a sample

21 7 set, retroreflectometers are used. For each individual sheeting type and color, traffic signs are replaced before the retroreflectivity measurement degrades below the minimum MUTCD levels. When implementing this method, the frequency of retroreflectivity measurements, the number of sites, and the number of signs in each sample set are the key factors that should be taken into consideration by agencies. 1.3 Research Questions and General Approach The current research makes an effort to find a solution for this question: Why do traffic signs get damaged and deteriorated? In other words, this study is conducted to determine the most significant contributing factors to signs visual condition. To address this issue, it is necessary to identify the effects of sign damage and deterioration on sign performance. In addition, it is required to categorize sign damage and deterioration forms into appropriate groups based on their source. Then, integrating traffic sign data with GIS-based climate and location data should yield the desired results. This study is categorized into four parts to find the answer of the mentioned questions which can be written as follows: What are the effects of various damage and deterioration forms on traffic sign retroreflectivity? To answer this question, based on the source of damage and deterioration, manual data of over 1700 traffic signs are organized into appropriate categories. Then, t-test, odds ratio, and regression models are employed to analyze the data. Various damage

22 forms are compared in terms of their impact on traffic sign visibility. A sign replacement plan is also suggested based on the results of data analysis. 8 What are the most important factors affecting traffic sign visual condition? This section examines various factors that may affect a sign s visual condition including sign attributes, localized condition, and climate data. To do so, mobile-based data of more than 97,000 traffic sign are utilized. Various official online sources are used to obtain weather and localized conditions of traffic signs. ArcGIS software is employed to extract climate and location data for each individual sign. Due to large size of data and statistical issues associated with traditional methods, non-parametric tree-based models are used to analyze data. Which traffic signs are more likely to get vandalized? Traffic sign vandalism is exclusively the result of humans. This section intends to answer two subsequent questions: (a) Is there an association between the demographics of local populations and traffic sign vandalism rates? (b) Does traffic sign vandalism correspond to specific types of signs? To answer these questions, traffic sign attributes and localized condition are integrated with demographic data. Chi-square and trend tests and regression model are used to analyze data. What are the environmental factors contributing to the temporary obstruction of the sign messages? Among damaged signs, dirty traffic signs are unique since their damage is not permanent and they just can be cleaned instead of replaced. To identify the most important contributing factors to traffic sign dirt, traffic signs were combined with

23 9 location and climate data using ArcGIS. The Chi-square test was employed to identify contributing factors to traffic sign dirt. Finally, random forests statistical model was utilized to analyze the data and rank all of the factors based on their importance to the sign dirt. 1.4 Data Collection Transportation agencies must maintain the minimum retroreflectivity levels provided by MUTCD. To do so, agencies are required to efficiently evaluate and manage sign data inventory. In addition, the Moving Ahead for Progress in the 21st Century Act (MAP-21) performance measures for asset management drives the need for datainformed decisions for asset management. To successfully implement such programs, the collection of reliable and accurate data is a key factor. Two sign data collection methodologies, manual and mobile, were used in the state of Utah in the period from 2011 to 2012 to fulfill those mandates Manual Method In 2011, a preliminary study was conducted to collect a sample set of traffic signs under UDOT s jurisdiction. In order to provide an overall perspective of compliance with minimum retroreflectivity levels across the state, different regions were considered, and the sample dataset collection routes were selected. Figure 1-1 displays the locations of the recorded signs. A snapshot of the process of data collection is shown in Figure 1-2. The

24 10 overall effort was accomplished by a three-person team, with specific tasks assigned to each member. Throughout the data collection, a handheld retroreflectometer was used, as well as a global positioning system (GPS) unit that included a customized data dictionary to record specific attributes of signs, including: Location (GPS coordinates) Roadway type (rural, urban, canyon, mountain) Background color (green, red, white, yellow) Sheeting type (Type I, III, III HIP, IX, XI) Retroreflectivity measurements (cd/lx/m 2 ) MUTCD type and code (warning, regulatory, guide) Face direction (north, east, south, west) Mount height (ft) Offset from the edge of roadway (in) Installation date (month/day/year, if known) Form and severity of sign deterioration

25 Figure 1-1 Locations of traffic signs collected manually 11

26 12 Figure 1-2 Snapshot of the process of manual sign data collection The team also took a photo for each individual sign. At its completion, the sample data population consisted of more than 1700 traffic signs located across the state s major climatological regions in both rural and urban environments. Of more than 97,000 traffic signs maintained by UDOT, almost 1.5% were recorded by the team, providing an appropriate sample size to ensure a 95% confidence level. Using a Delta RetroSign Model 4500 retroreflectometer, the retroreflectivity values were also measured. The Model 4500 illuminates the sign at a -4 entrance angle, with an angle of observation at 0.2. Holding the retroreflectometer vertical and stable against the sheeting was a detail that the team took into consideration during the collection process. Following ASTM E1709 (2009) standards, four retroreflectivity measurements were taken for each individual sign. Then, the four measurements were averaged in order to determine the overall retroreflectivity of each individual sign.

27 Mobile-based Method Asset management provides a basis for examining the performance of transportation facilities in a cost-effective manner (Asset Management, 2007). To obtain comprehensive information about its multi-billion dollar road assets, UDOT embarked on an effort to collect data along the nearly 6,000 centerline miles (approximately 18,000 lane miles) of state routes and interstates in 2012 (Ellsworth, 2013). Since mobile-based methods are being increasingly considered by transportation agencies across the country to address the size of necessary data collection (U.S. Department of Transportation, 2013), they were investigated and chosen to conduct the study. In this way, data can be collected with high speed particularly in areas with restricted access and safety concerns (Williams, et al., 2013). Recently, mobile-based methods have been used to collect data in various fields, including bridge management (Demann, 2010; Zolghadri, et al., 2014; Zolghadri, et al., 2016a), environmental monitoring (Weibring, et al., 2003), petroleum (Puente, et al., 2013), and water resources (Tang, et al., 2014; Hassan-Esfahani, 2015; Hassan-Esfahani, et al., 2015). Previous studies discussed mobile-based data collection methods. For example, the prototype mobile luminance (Miles & Ge, 2014), mobile laser scanning (Gong, et al., 2012; Jalayer, et al., 2014), video-based data collection systems (Balali, et al., 2013; Balali & Golparvar-Fard, 2014), and advanced mobile asset collection systems (Carlson, 2011) were evaluated. Another effort compared mobile and manual data collection methods (Findley, et al., 2011; Khalilikhah, et al., 2015b). UDOT s comprehensive approach was deployed by an instrumented vehicle driving at freeway speeds and

28 14 collecting data on the following types of assets: signs, pavements, markings, guardrails, lights, and reflectors. Figure 1-3 illustrates the vehicle used for data collection. The sensors on the UDOT data collection vehicle included a LiDAR sensor, a laser road imaging system, a laser rut measurement system, a laser crack measurement system, a road surface profiler, and a position orientation system. In addition, imaging technologies were integrated to automatically collect high-resolution detailed images from the assets. Figure 1-4 shows an example of LiDAR imaging and the associated digital photo. At the completion of the effort, many sign attributes, including location, size (length and width), mount height and facing direction, were measured. Figure 1-3 LiDAR equipped vehicle used for UDOT sign data collection (Source:

29 15 Figure 1-4 LiDAR imaging and associated photolog (Source: The first phase of the project, data gathering, included the collection of data on a wide variety of roadway assets along state roads while driving only one vehicle. The second phase of the project, post-processing, was conducted by survey. Many sign attributes, such as location, size (length and width), mount height, collection date, and facing direction were measured. In addition, during the post-processing analysis, the captured daytime digital images were examined by trained operators, to rate each sign s visual condition as good, fair, or poor (GFP) with respect to their overall condition. Moreover, traffic signs with any form of damage on the face were acquired throughout the entire data set. Figure 1-5 provides examples of photos taken from traffic signs.

30 16 Figure 1-5 Taking image from traffic signs (Source: Categorizing traffic signs into these groups (GFP) was based on personal judgment with respect to the signs general condition, overall legibility, and daytime appearance. Generally, a sign in good condition did not exhibit any notable form of damage or deterioration and showed an adequate visibility. Signs that showed damage or deterioration on the face or exhibited ineffective visibility were categorized as poor or fair signs, based on the damage severity and its effects on the legibility and visibility of the sign. At the end, over 97,000 signs under the jurisdiction of UDOT were measured. Figure 1-6 shows the locations of mobile-based signs, created using ArcGIS.

31 17 Figure 1-6 Locations of mobile-based traffic signs Table 1-2 compares the captured attributes of traffic signs provided by each method (manual vs. mobile-based). Sheeting type, installation date, and retroreflectivity measurements constituted important information provided only by the manual method.

32 18 Table 1-2 Summary of surveyed sign attributes using manual and mobile methods Sign Attribute Method Manual Mobile LiDAR MUTCD Code * * Direction * * Sheeting Type * Sheeting Backing * Mount Height * * Offset from Roadway * Deterioration Form * * Survey Date * * Installation Date * Location * * Condition * Retroreflectivity Measurements * 1.5 Anticipated Contributions In the past, many researchers have studied traffic signs from the perspective of the road user. Though such studies of road user characteristics have resulted in some progress, the attributes of traffic signs themselves are just as important, since the process of conveying messages also includes an interaction between signs and drivers. However, research focused on measuring the effectiveness of traffic signs is not well-developed. The chief contribution of the study is the size of traffic sign data. Analysis of over 97,000 in-service traffic signs is unique in comparison with similar studies that mostly were conducted using small data sets. The other anticipated contribution of this dissertation is to address how the existence of sign damage and deterioration may impact traffic sign retroreflectivity. Previous research discussed that retroreflectivity efficiency ensures visibility of the traffic sign, while damage and deterioration affect the legibility of the

33 19 sign. In other words, previously, damage/deterioration and retroreflectivity were studied separately. In addition, the proposed data integration strategy to obtain and extract GIS data can be used to assess all transportation infrastructures physical condition. Moreover, although significant efforts have been made to safeguard human health and the environment from air pollutants, little has been done to study the effects of air pollutants on the deterioration of sign material. To assess the effects of air pollutants on traffic sign deterioration is another contribution of the dissertation. In addition, few studies have focused on traffic sign vandalism due to a lack of detailed information about traffic sign vandals. This study contributes by combining demographics of local population, traffic sign data, and GIS data to address this issue. Every year, transportation agencies across the country spend tens of thousands, or even hundreds of thousands, of dollars to repair or replace inefficient traffic signs. Therefore, traffic sign degradation results in increased cost to taxpayers. This study provides a good basis for agencies regarding choosing the appropriate sheeting type before installing signs in the field, while more resistant types are needed for signs more vulnerable to failure. In addition, by identifying traffic signs with a higher likelihood of failure, agencies can schedule more frequent inspections for those signs. The replacement of ineffective traffic signs with new signs could help drivers navigate roads in a safer and more efficient manner.

34 Research Outline The remainder of this dissertation is organized as follows: Chapter 2 discusses the effects of various damage and deterioration forms on traffic sign retroreflectivity. Chapter 3 identifies the most important factors affecting traffic sign visual condition. Chapter 4 determines which in-service traffic signs are more vulnerable to vandalism. Chapter 5 identifies the most important contributing factors to traffic sign dirt. The key conclusions of the study and recommendations for future studies are provided in Chapter 6.

35 21 CHAPTER 2 THE EFFECTS OF DETERIORATION ON TRAFFIC SIGN VISIBILITY 2.1 Introduction After adopting final revisions to the Manual on Uniform Traffic Control Devices (MUTCD) in 2012, the three original target compliance dates for minimum retroreflectivity levels were changed. The minimum retroreflectivity levels established by the MUTCD require transportation agencies to implement an assessment or management method to maintain sign retroreflectivity minimum levels. Over the course of the past few years, transportation agencies have aggressively developed methodologies to meet the MUTCD mandate. Of all the assessment and management methods, the expected sign life has been selected most often as a primary or secondary method (Re & Carlson, 2012). Using the expected sign life method, signs are replaced before their retroreflectivity degrades below the minimum levels. Similar to the other management methods, expected sign life depends on establishing a baseline data set and conducting periodic follow-up surveys. However, the expected life of a sign has been shown to exhibit discrepancies, depending on the manufacturer, sheeting type, color, and geographic location (Evans, et al., 2012). This chapter of the study was conducted using sample data collected across the state of Utah. In 2011, the Utah Department of Transportation (UDOT) sponsored field investigations by a team to investigate the compliance of traffic signs with the guidelines set forth by the 2009 MUTCD. Over 1700 in-service traffic signs were recorded during

36 22 the effort. Figure 2-1 illustrates a summary of the captured signs based on the sheeting type and color. As shown in the figure, recorded retroreflectivity sheeting included Type I, III, III HIP, IX, and XI (FHWA, 2011). Approximately 60% of the surveyed signs were either white or yellow, while sheeting Type III was used by the majority of the dataset (58% of the signs). Figure 2-1 Summary of surveyed signs by type and color If the captured retroreflectivity was below the minimum level, the sign was recorded as failing in the database. At the conclusion of the study, the rate of failure for the entire sample population was nearly 8%. If extrapolated to the entire population, it could be inferred that 9 out of 10 of signs throughout the state pass the MUTCD

37 23 requirements. More specifically, a 95% confidence interval for the compliance rate is (90.7%, 93.3%). In other words, there is strong evidence that between 90.7 % and 93.3% of signs across Utah comply with MUTCD s requirements. By considering the measured retroreflectivity of each individual sign, the objective of this chapter is to compare the effects of various deterioration forms on traffic sign retroreflective performance. To do so, a range of measured sign retroreflectivity is provided from the sample data, with special notice given to sign sheeting type and color. By focusing on signs with a known installation date, the second goal is to determine the association between sign age and retroreflectivity by developing regression models. 2.2 Data Analysis During data collection effort, it was found that only 17% of the 1,683 surveyed signs included an installation date. UDOT has not required recorded installation dates since 2004 (Evans, et al., 2012), which explains why this percentage is so low. In addition, it was observed that a number of signs exhibited a variety of age-related deterioration forms on their faces. In these cases, they found that it is more likely that deterioration led to the decline in sign retroreflective performance. In this chapter, deterioration forms were categorized into three groups: aging, environmental, and vandalism. Data analysis of this chapter is presented in two sections. First, after considering the entire sample data, the relationship between different forms of deterioration and the retroreflective performance of the traffic signs is examined to compare the effects of

38 24 aging deterioration to the other forms of deterioration. To do this, a range of retroreflectivity measurements for each sign deterioration form was obtained, based on the sign sheeting type and color. Then, by focusing on the signs with a recorded installation date, the association between sign age and performance was investigated. Almost 64% of signs with an installation date were green guide signs. Finally, regression models were developed for these guide signs Sign Deterioration This section of the chapter is focused on the effects of various deterioration forms on sign retroreflective performance. First, the deteriorated signs are organized into categories. This is followed by a summary of statistical measurements, including the highest and lowest values of the measured retroreflectivity for each group of deteriorations, as well as the mean and standard deviation for each group. Also, the necessary number of recorded signs with background retroreflectivity above or below the minimum levels is counted in order for each category to yield the desired results Deterioration Categories Various deterioration forms were observed throughout the research team s data collection effort. In addition, photos were taken of every surveyed sign in order to further classify each form. Major and minor deteriorations were recorded and ranked, based on severity. For this study, only major deterioration forms were taken into consideration, since minor deterioration had negligible effect on sign performance. Ultimately, the

39 25 deteriorated signs were categorized into three groups: aging, environmental, and vandalism, as shown in Figure 2-2. The following categories are an extension of those defined by (McGee, 2010; Evans, et al., 2012; Boggs, et al., 2013): Figure 2-2 Deterioration categories Aging: All of the signs surveyed with cracking, peeling on their face, or faded colors were classified as aging signs. Cracking deterioration consisted of the retroreflective

40 background cracking and degrading over time, while peeling damage occurred on the legend of a sign. This form of deterioration was mostly widespread on Type I sheeting. 26 Environmental: This group of signs included signs bent by wind or snow, signs bent or knocked down by vehicles, and signs damaged by contact with trees or tree sap. Bending damage described signs with significant portions of the sheeting bent, causing light to be reflected away from its origin. Signs with multiple cuts on the sheeting as a result of transportation and installation of the sign were also categorized into this group, since the damage was not deliberate. Vandalism: The most diverse category of sign deterioration included any deliberate damage to the sign face caused by humans, such as paintball damage, ballistic damage from firearms, glass bottle impacts, eggs, stickers, dents, graffiti, and over painting Type I Signs Table 2-2 provides a summary of the retroreflectivity specifications for Type I signs. As seen in the table, with the exception of yellow signs, the average retroreflectivity of aging signs was lower than the other signs. White and yellow signs showed the greatest range of measurement. Yellow Type I signs showed the highest rate of failure, with 91% out of compliance with minimum levels. The majority of average values were also below the minimum levels. Of 139 Type I sheeting signs, the failure rate was 71%. Also, 8 out of every 10 (84%) aging signs were out of compliance. It is

41 27 necessary to mention that regardless of what obtained from this study, UDOT is replacing Type I signs, due to such poor performance. By comparing the current data, with no Type I red sheeting signs, to the data collected in 1999, in which there were a significant number of Type I red sheeting signs, UDOT s process of replacement of Type I signs is evident, where 45% reduction in the number of Type I red signs was observed. Color Green Red White Yellow Table 2-1 Retroreflectivity specifications for Type I signs Deterioration # of Retroreflectivity (cd/lx/m 2 ) # of Signs # of Signs Above Below Signs High Mean Low S. D. Minimum Minimum Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None Type III and III HIP Signs As shown in Table 2-2, with the exception of green signs, for all colors of Type III signs, the average retroreflectivity of aging signs was lower than the other signs. Only

42 28 32 failures out of 980 signs demonstrated a good performance for Type III signs. However, the retroreflective performance of Type III signs with aging deterioration was much worse than the others, with a 10% rate of failure. The compliance rate of red and yellow aging signs is especially low. Yellow signs had the highest rate of failure, as well as the highest range of variability. The performance of white signs was good, with respect to minimum level compliance. Color Green Red White Yellow Table 2-2 Retroreflectivity specifications for Type III signs (cd/lx/m 2 ) Deterioration # of Retroreflectivity (cd/lx/m 2 ) # of Signs # of Signs Above Below Signs High Mean Low S. D. Minimum Minimum Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None Since Type III signs exhibited all of the deterioration forms, a statistical test was conducted to determine whether or not there is a statistically significant difference between the retroreflective performance of aging deteriorated signs and

43 29 environmental/vandalism signs. The null hypothesis was that there is no difference between average retroreflectivity measurements for aging and non-aging deteriorated signs, while the alternative hypothesis stated that there is difference between them. To do so, a two tailed t-test was performed, and the results are shown in Table 2-3. As seen in the table, the p-values were statistically significant at the level of 0.05 for green, red, and yellow signs. As a result, in the context of the collected data, the average measurements of sign retroreflectivity for age-related deterioration forms vs. other forms were different for green, red, and yellow signs. Signs that exhibited aging deterioration also had a lower mean value. In contrast, depending on deterioration form, difference in the retroreflective performance of white Type III signs was not confident. Color Green Red White Yellow Table 2-3 Performance of Type III aging vs. non-aging signs Deterioration # of Retroreflectivity (cd/lx/m 2 ) Signs Mean Difference S. D. t-statistic P-Value Aging Non-Aging Aging Non-Aging Aging Non-Aging Aging Non-Aging <0.001 As seen in Table 2-4, Type III HIP (high intensity prismatic) signs showed a very high performance level, though their sample population was small. All Type III HIP signs recorded across the state of Utah were in compliance with the minimum required

44 30 standards. Only one Type III sheeting sign was observed with age-related deterioration on the face, so no conclusion might be drawn with respect to this form of deterioration. In total, only 13 out of 190 (almost 7%) Type III HIP signs were recorded with any form of deterioration. By considering the standard deviation values, the range of variability of Type III HIP signs was also large in comparison with Type III signs. Color Green Red White Yellow Table 2-4 Retroreflectivity specifications for Type III HIP signs Deterioration # of Retroreflectivity (cd/lx/m 2 ) # of Signs # of Signs Above Below Signs High Mean Low S. D. Minimum Minimum Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None Type IX and XI Signs In UDOT s sign inventory, Types IX and XI are the most recently adopted sheeting types. Table 2-5 and Table 2-6 summarize the retroreflective performance of

45 31 Type IX and Type XI, respectively. With the exception of a few signs, almost all of these signs performed well. Only three green Type IX sheeting did not maintain the minimum levels. In addition, fewer aging signs were observed in the Type IX and XI samples, an evidence of good performance of these types. For those observed in the color yellow, the average measurement of retroreflectivity was less than that of other signs. Color Green Red White Yellow Table 2-5 Retroreflectivity specifications for Type IX signs Deterioration # of Retroreflectivity (cd/lx/m 2 ) # of Signs # of Signs Above Below Signs High Mean Low S. D. Minimum Minimum Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None

46 32 Color Green Red White Yello w Table 2-6 Retroreflectivity specifications for Type XI signs Retroreflectivity (cd/lx/m 2 ) # of Deterioration Signs High Mean Low S. D. # of Signs Above Min # of Signs Below Min Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None Aging Environmental Vandalism None In-service Sign Age Since 2004, UDOT has required agencies to have an installation sticker on both the front and back of all signs placed in the field. Typically, the sticker on the front of the sign has a transparent background with a black legend for the year it was installed, whereas the back contains the month and year of installation, as well as the company that constructed the sign. However, this policy was not consistently fulfilled by the stations and contractors installing signs for UDOT (Evans, et al., 2012). As a result, at the completion of the data collection, only 17% of the surveyed signs were placed with an installation date. A summary of these signs, based on the sign background sheeting type and color, is shown in Figure 2-3. Approximately 6 in 10 (64%) of these groups of signs

47 33 were green guide signs. Classifying signs by their color, sheeting type, and age also led to low numbers of red, yellow, and white signs in each group. Based on the data, it was not possible to draw strong conclusions for any one category. Therefore, this section focuses on green signs including D1 (destination), D2 (distance), D10 (milepost), E1 (interchange), and E5 (exit) MUTCD type signs. The range of installation dates of the surveyed signs was from 2003 to Of the 186 green signs, only 5% were measurably deteriorated. The majority of signs were also Type III sheeting. No Type I sheeting was surveyed, since UDOT has started replacing Type I signs. Also, the majority of green signs were placed in the years ranging from 2003 to It is interesting to note that all of the in-service green signs with known installation dates maintained the minimum retroreflectivity levels. However, these measured retroreflective values were variable. After considering all of the possible contributing factors, determining the effects of sign age on the observed degradation is the main goal of this section. Table 2-7 presents a summary of descriptive statistics. To make better sense of the association between sign age and retroreflective performance, the sign age data were classified into categorical groups in 2-year intervals. Generally, with the increase in sign age comes a steady decrease in the average values of sign retroreflectivity, with few exceptions though. In this study, 104 out of 110 (94%) Type III signs were between 6 and 8 years old; thus, more details on sign age are needed to provide more comprehensive conclusions. However, regression models might answer this question. In Type III HIP green signs, the average retroreflective value of signs 4 to 6 years old is lower than that of signs 6 to 8 years old. This same result should be obtained for Type IX signs, if signs in

48 # of surveyed signs # of surveyed signs groups 2-4 are removed due to their low numbers. The majority of Type XI signs are zero 34 to two years old, although surprisingly a larger standard deviation is observed in RA values. 70 Type III Type III HIP Type IX Type XI year sheeting type White Green Yellow Red year color Figure 2-3 Summary of surveyed signs with known age

49 35 Table 2-7 Retroreflectivity specifications for green signs Sheeting # of Retroreflectivity Age Type Signs High Mean Low S. D III III HIP IX XI To examine the effects of sign age on retroreflective performance, a modeling study was conducted based on background color. Developing regression models with respect to sign age and retroreflective measurements should yield the desired results. Since the relationship between age and retroreflective measurements is unknown, a linear regression model was developed, as well as quadratic and cubic polynomial models for each sheeting type, to examine the association. Table 2-8 provides the final fitted models with respect to sheeting types. Since cubic polynomial regressions were ill-fitted to the task, their results are not shown in this section.

50 36 Table 2-8 Regression models of signs retroreflective measurements Sheeting Age Model Constant Type (years) Age 2 R 2 III Linear 71.39* -2.74* Polynomial * 2.3* 0.36 III HIP Linear 80.51* Polynomial * IX Linear 84.15* Polynomial 90.19* XI Linear 91.22* Polynomial 87.09* *P-Value < 0.05, statistically significant at level of Linear Regression Figure 2-4 plots the linear regression models obtained for each sheeting type created using SAS software. As expected, the slope of all models, with the exception of Type III HIP, was negative. These results showed that by increasing sign age, sign retroreflectivity will decrease. The positive slope of the Type III HIP plot may, in fact, reflect the range of the signs ages. The majority of surveyed Type III HIP signs ranged in age around 6 years old, so the positive slope is likely caused by the skewed sample. After focusing on plots with degrading slope, Type III showed the sharpest slope, which means the decrease in the value of sign retroreflectivity for this type of sign is higher than that of the others. It is important to notice that the coefficient of the age variable was statistically significant at level of 0.05 only for the Type III model Polynomial Regression When the relationship between response and predictor variables is unknown, polynomial regression may be used (Weisberg, 2005). In other words, polynomial

51 37 regression approximates the association between variables, while a function is smooth but not straight. After developing second-order polynomial regression models, it may be concluded, again, that the coefficient of age and quadratic age variables was statistically significant at the 0.05 level for only the Type III model. In addition, the obtained R 2 values for Type III sheeting were significantly better than those for Types IX and XI. In comparison with linear regression models, the R 2 values obtained from polynomial models were improvements, but still somewhat ill-fitted to the task. (a) Type III (b) Type III HIP (c) Type IX (d) Type XI Figure 2-4 Linear Regression Models for Green Signs

52 38 Generally, the developed models for in-service green signs were not statistically significant enough to predict the expected service life of the signs. In addition, poor values of R 2 reflected the point that there are numerous other contributing factors affecting sign performance. Other studies focused on developing models for retroreflectivity with respect to sign age have found the same results, in that the fitted models were not significant (some of them with a lack of degradation) and yielded a poor value of R 2. In other words, the claim of predicting sign service life with respect to MUTCD minimum levels (Black, et al., 1992; Bischoff & Bullock, 2002; Re, et al., 2011; Pike & Carlson, 2014) is seriously questioned. To address this issue, we take into consideration the effects of deterioration on the sign s retroreflectivity. 2.3 Discussion Figure 2-5 represents the rate of retroreflective failure in signs surveyed with aging, environmental, vandalism, or no deterioration. As can be seen, aging has, by far, the highest rate of failure, with 55 out of 165 (33%) signs exhibiting this form of deterioration to the extent of non-compliance with MUTCD retroreflectivity requirements. For all other deteriorated signs, only 17 in 211 (8%) failed. With regards to vandalized signs, 12% did not maintain, and the failure rate for environmental deteriorated signs was unremarkable. For signs without any form of deterioration on their face, only 62 out of 1307 (4.7%) signs did not maintain compliance.

53 rate of failure 39 35% 33% 30% 25% 20% 15% 10% 5% 2% 12% 5% 0% Aging Environmental Vandalism None Figure 2-5 Sign retroreflectivity failure by deterioration categories A contingency table according to sign deterioration form (aging, environmental/vandalism) and retroreflective performance (fail, pass) was created for this study to compare sign failure rates. Certain potential measures were taken into account: Difference of Proportion, Relative Risk, and the Odds Ratio (Agresti, 2007). Of these three possible measurements, we used the odds ratio for our study. Odds Ratio is the most widely used measurement in practice by far since it is invariant regarding whether a study is prospective or retrospective and it is best-suited when the outcome is relatively rare (Corcoran, 2013). If P0i is the probability that signs with no damage form i fail to comply with the minimum MUTCD standards and P1i is the probability that signs with damage form i on the face fail to maintain the minimum levels, formally speaking, we were interested in a test of: H 0 : P 0i = P 1i or H 0 : log[ Odds(P 1i) Odds(P 0i ) ] = 0 H 1 : P 0i P 1i or H 0 : log [ Odds(P 1i) Odds(P 0i ) ] 0 (2-1)

54 In addition, a 95% confidence interval for the odds ratios are shown in the table, given by (α=0.05, thus z α 1 = z ): 2 40 exp (log[ Odds(P 1i) Odds(P 0i ) ] ± z S. E. (log[ Odds(P 1i) ]) (2-2) Odds(P 0i ) Table 2-9 shows the odds ratio calculated to compare aging signs with environmental/vandalism signs in terms of their retroreflective performance. At the conclusion of this study, it was found that the odds of sign failure for signs with aging deterioration were 9.11 times higher than those for signs with environmental or vandalism, or no deterioration, with a confidence interval equaling (6.14, 13.52). After focusing on deteriorated signs, we are 95% confident that the odds of sign failure for aging signs, including cracking, peeling, and fading deterioration forms are between 3.16 and times those for signs with vandalism or environmental deteriorations. Table 2-9 Odds ratio for failure rates Failure Rate Odds 95% CI Aging Env/Van Ratio Lower Upper Aging vs. all others Aging vs. deteriorated

55 Conclusions Initial analysis indicated that approximately 8% of signs were not in compliance with minimum MUTCD retroreflectivity standards. A t-test confirmed the result that there was a statistically significant difference between signs retroreflective performance based on sign deterioration. Thus, the strong association between deterioration on the face of sign and retroreflective performance of traffic signs was evident. Age-related deterioration had, by far, the highest rate of retroreflective failure, with an odds ratio 9.11 times higher than signs with environmental deterioration, vandalism deterioration, or no deterioration. Based on the results, traffic signs surveyed with cracking, peeling on their face, or faded colors, are more likely to fail to convey their message to the road users. This failure is even more critical for regulatory and warning signs. Since the deterioration affects retroreflective performance of traffic signs which leads to safety issues for drivers, transportation agencies should plan on the replacement of key deteriorated signs. To do this, agencies could schedule more frequent sign inspections to identify deteriorated signs, or people might report problems with traffic signs to agencies. For example, in New York City traffic sign problems (deteriorated, blocked, or missed signs) can be reported to the department of transportation by phone call or online (NYC, 2015). Then, taking the findings of this study into consideration, agencies may prioritize the replacement of deteriorated signs to provide safer environment for the users.

56 42 CHAPTER 3 IMPORTANT FACTORS AFFECTING TRAFFIC SIGN VISUAL CONDITION 3.1 Introduction In the past, many researchers have studied traffic signs from the perspective of the road user. For example, the understandability and comprehensibility of traffic signs, as well as drivers recognition of signs have been studied (Al-Madani & Al-Janahi, 2002; Ng & Chan, 2007; Kirmizioglu & Tuydes-Yaman, 2012; Liu, et al., 2014). Though these studies resulted in some progress, the attributes of traffic signs themselves are equally important. The process of conveying messages involves an interaction between signs and drivers. However, research focusing on the effectiveness of traffic signs is not welldeveloped. Some studies analyzed sign performance based solely on sign age (Kirk, et al., 2001; Wolshon, et al., 2002; Huang, et al., 2013). Other studies reported that sign age was not the only significant factor and the expected life of a sign may exhibit discrepancies, depending on the manufacturer, sheeting type, color, and geographic location (Black, et al., 1992; Bischoff & Bullock, 2002; Hildebrand, 2003; Re, et al., 2011; Pike & Carlson, 2014; Khalilikhah & Heaslip, 2016c). In addition, many in-service signs may not include an installation date, so their age is unknown (Evans, et al., 2012). Therefore, it is important to determine the contributing factors that lead to a decrease in sign performance.

57 43 Multiple studies were performed, focusing on assessment and management of traffic signs retroreflectivity (Carlson & Lupes, 2007; Kipp & Fitch, 2009; Harris, et al., 2009; Balali & Golparvar-Fard, 2015). A simulation of the sign inspection process to optimize sign management was conducted (Rasdorf, et al., 2006). Other studies analyzed the effects of emissions on sign retroreflective performance and deterioration rate (Khalilikhah, et al., 2015a, Khalilikhah & Heaslip, 2016b). In addition, a risk-based approach for agencies to follow when checking for retroreflective sign compliance was developed (Liang, et al., 2012). There have also been studies focused on long-term deterioration of traffic signs, with special attention given to color and retroreflectivity, in order to provide information related to select types of signs (Brimley & Carlson, 2013). The objective of this chapter is to determine which explanatory variables are most significantly associated with traffic sign degradation. Figure 3-1 depicts the process of study.

58 44 Figure 3-1 Process of the study 3.2. Traffic Sign Attributes This study is conducted using mobile-based data of over 97,000 traffic signs across the state of Utah. The traffic sign data included a variety of sign attributes, including size of the sign (length and width), mount height, facing direction, and background color. In addition, with regard to their visual condition, traffic signs were rated at three levels (good, fair, and poor). These levels were obtained by inspection of daytime digital images taken of each sign. Categorization of the traffic signs into these

59 45 groups (GFP) was based on personal judgment with respect to the sign s general condition, overall legibility, and daytime appearance. Generally, a sign in good condition did not exhibit any forms of damage or deterioration. Signs that showed damage or deterioration on the face were categorized as poor and fair signs, based on the damage severity and its effects on the legibility of the sign. Google Street View images were used to ensure the accuracy of the sign data visual condition ratings (Figure 3-2). At the end, over 97,000 signs under the jurisdiction of UDOT were recorded. Among all sign ratings, almost 7% of the surveyed signs were in fair or poor condition. (a) examples of traffic signs rated as good (b) examples of traffic signs rated as fair (c) examples of traffic signs rated as poor Figure 3-2 Examples of traffic signs rated as (a) Good; (b) Fair; (c) Poor (Source: Google Street view images)

60 Sign Length and Width In order to provide adequate message comprehensibility, the appropriate size of each sign (length and width) must be determined with respect to certain criteria, including the task of the sign and the prevailing traffic speed on the road (Manual on Uniform Traffic Control Devices, 2012). Tables 3-1 and 3-2 display the corresponding sign ratings of each category of length and width. It can be noted that the percentage of signs in good condition changes very little among the different categories of length or width. Sign Length (in) # of Traffic Signs Table 3-1 Sign condition by length Sign Condition Good Fair Poor Count Percent Count Percent Count Percent <20 24,100 23, ,056 20, ,318 33, , ,904 7, ,830 3, >100 2,109 2, Sign Width (in) # of Traffic Signs Table 3-2 Sign condition by width Sign Condition Good Fair Poor Count Percent Count Percent Count Percent <20 19,333 18, ,744 31, , ,712 20, , ,446 7, ,528 3, ,580 4, >100 4,964 4,

61 Sign Mount Height In addition to sign size, the height of each sign above the road was also obtained for each individual traffic sign. According to the summary of sign condition by mount height in Table 3-3, it sounds signs placed higher performed better. For signs placed 10 feet or more above the road, the percentage of signs in good condition was 98%, whereas the percentage for the signs with a mount height of less than three feet was 91.2%. Sign Height above Road (ft) Table 3-3 Sign condition by mount height # of Traffic Signs Sign Condition Good Fair Poor Count Percent Count Percent Count Percent <3 2,384 2, ,648 28, , ,225 46, , ,547 6, >13 6,845 6, Sign Facing Direction A study concluded that sun exposure has an impact on sign condition (Rasdorf, et al., 2006), thus the directions that signs face was also measured. Table 3-4 summarizes the sign ratings based on the direction the sign faced. The percentage of good performing signs changed very little with regards to the direction that the sign faced.

62 48 Sign Facing Direction Table 3-4 Sign condition by facing direction Sign Condition Good Fair Poor Count Percent Count Percent Count Percent # of Traffic Signs North 16,129 15, Northeast 9,044 8, Northwest 8,662 8, East 14,672 13, West 14,441 13, Southeast 8,458 7, Southwest 9,076 8, South 16,860 15, Sign Background Color The background colors of observed signs were white, yellow, green, black, blue, brown, orange, and red. Respectively, 29%, 23%, and 22% of the surveyed signs were white, yellow, and green, while the other five colors collectively made up only 26% of the total number of signs. Table 3-5 depicts a summary of the sign conditions based on the sign background color. Yellow signs tend to have twice the failure rate of other colors. Sign Color # of Traffic Signs Table 3-5 Sign condition by color Sign Condition Good Fair Poor Count Percent Count Percent Count Percent White 28,610 27, Yellow 22,844 19, Green 21,112 20, Others 24,748 23,

63 Climate and Location Data Integration To examine the effects of all potential contributing factors to traffic sign condition, the collection of climate and location information was also necessary. Several different online sources were used to obtain elevation, temperature, precipitation, and land cover measurements. Afterwards, ArcGIS software was used to combine this online data information with sign location (latitude and longitude). The resulting values of climate and location data for each individual traffic sign were extracted from the raster data. There have been multiple studies that used geographic information system to obtain the desired data, such as landscape attributes (Shelley, et al., 2000), sight distance (Castro, et al., 2011), and volcanic ash cloud (Scaini, et al., 2014). In the following sections, SI units are used in the created maps to simplify the conversion from/to US customary systems to/from SI units Ground Elevation During the data collection process, the equipped vehicle recorded the geographic coordinates of each traffic sign. To ensure the accuracy of each recorded elevation, we used the National Elevation Dataset (NED30) digital elevation model from the United States Geological Survey (USGS, 2013) in this study. Ground elevation measurements for each individual traffic sign were extracted from the raster data using ArcGIS (Figure 3-3). The results of this extraction are summarized in Table 3-6. The majority of the recorded traffic signs are in areas with an elevation between 3200 and 6500 feet. As expected, the percentage of signs in good condition decreased as the elevation increased.

64 These results probably reflect the effects of UV radiation and snow frequency, both of which increase with higher elevation (Boggs, et al., 2013) Temperature and Precipitation The Parameter-elevation Regressions on Independent Slope Model (PRISM, 2010) climate mapping system was used to obtain the thirty-year average ( ) annual temperature and precipitation data across the state of Utah. The PRISM group reveals spatial climate data obtained from a wide range of observations. The mean temperature and precipitation measurements for each individual sign were extracted from the PRISM raster data using ArcGIS (Figures 3-4 and 3-5). Tables 3-7 and 3-8 summarize these results. With some minor exceptions, the rate of traffic sign degradation decreases with an increase in the mean temperature, and increases with an increase in mean precipitation.

65 51 Figure 3-3 Locations of traffic signs by elevation (m) Elevation (ft) Table 3-6 Sign condition by elevation Sign Condition Good Fair Poor Count Percent Count Percent Count Percent # of Traffic Signs <3,200 1,870 1, ,200-5,000 49,165 46, , ,000-6,500 34,198 31, , ,500-8,200 9,728 8, >8,200 2,353 1,

66 52 Figure 3-4 Locations of traffic signs by average annual temperature ( C) Mean Temperature ( F) Table 3-7: Sign condition by average annual temperature # of Traffic Signs Sign Condition Good Fair Poor Count Percent Count Percent Count Percent <41 4,339 3, ,326 9, ,382 19, , ,167 34, , ,112 19, >55 4,988 4,

67 53 Figure 3-5 Locations of traffic signs by average annual precipitation (mm) Mean Precipitation (in) Table 3-8: Sign condition by average annual precipitation # of Traffic Signs Sign Condition Good Fair Poor Count Percent Count Percent Count Percent <8 5,447 5, ,395 17, , ,163 19, , ,486 26, , ,357 10, >23 11,466 10,

68 Land Cover I employed 16-category land cover classification obtained from the National Land Cover Database 2011 to examine the effects of the environment surrounding the sign (NLCD, 2011). NLCD 2011 applied their land cover classification system consistently across the country at a spatial resolution of 100 feet. The classification system categorized land cover into the following groups: Water (Open Water, Perennial Ice/Snow) Developed (Open Space, Low Intensity, Medium Intensity, High Intensity) Barren Land (Rock/Sand/Clay) Forest (Deciduous, Evergreen, Mixed) Shrub Land (Dwarf Scrub, Shrub/Scrub) Herbaceous (Grassland/Herbaceous, Sedge/Herbaceous, Lichens, Moss) Planted/Cultivated (Pasture/Hay, Cultivated Crops) Wetlands (Woody Wetlands, Emergent Herbaceous Wetlands) Table 3-9 demonstrates the summary of sign condition on each category of land cover. Land cover data for each traffic sign was extracted from raster data created by ArcGIS (Figure 3-6). Almost 84% of the captured signs were located in developed areas. By focusing on developed areas, a trend could be observed: open space areas showed the highest percentage of signs in fair or poor conditions, while the lowest rate of failure was exhibited by high or medium intensity areas (Table 3-9). This finding coincided with

69 previous studies, which reported that the sign damage rate for rural signs was greater than that of urban signs (Boggs, et al., 2013). 55 Figure 3-6 Locations of traffic signs by land cover

70 Table 3-9 Sign condition by land cover # of Sign Condition Land Cover Traffic Good Fair Poor Signs Count Percent Count Percent Count Percent Developed-Open Space 26,956 24, , Developed-Low Intensity 24,560 22, , Developed-Medium Intensity 19,777 18, Developed-High Intensity 10,843 10, Bare Rock/Sand/Clay Forest 1,585 1, Shrub/Scrub 9,980 9, Grasslands/Herbaceous Planted/Cultivated 2,183 2, Wetlands Solar The National Renewable Energy Laboratory (NREL, 2013) database provides monthly and annual average daily total solar insolation resource data. Using annual average direct normal solar resource data, it was found that 85% of UDOT s traffic signs were placed in areas with the same solar condition (Figure 3-7). Thus, solar was considered insignificant variable for this chapter.

71 57 Figure 3-7 Locations of traffic signs by solar resource (kwh/m 2 /Day) 3.4 Statistical Methodology In order to provide robust conclusions regarding which predictors are important for traffic sign degradation, employing powerful statistical models is necessary. However, some serious challenges existed in this particular case, including: The large sample size

72 58 The extremely biased response variable (93% of measured signs were in good condition) The varied data structure of multiple explanatory variables (nominal or ordinal, continuous or categorical, quantitative or qualitative) The complex relations among the explanatory variables, some of which may confound the others, or maybe there are variable interactions The unknown, but likely, nonlinear relation between response and predictors The difficulties of this data made traditional statistical approaches, such as logistic multiple regression, log linear models, and analysis of variance (ANOVA), less powerful. These methods make assumptions (for example linearity) that directly neglect problems that exist and persist in real life. Instead, machine learning methods can address these issues. Among machine learning methods, the Random Forests (RF) model is able to simultaneously handle all of these challenges (Cutler & Breiman, 2004; Breiman & Cutler, 2007; Moisen, 2008). A RF model is purposely designed to handle challenges of nonlinearity, complexity, varied data structure, interactions, multicollinearity, and high dimensionality without making any misleading assumptions (Moisen, 2008). Random forests procedure has been widely used in different fields of science and engineering, including public health (Ye, et al., 2005; Goldstein, et al., 2010), urban planning (Lu, et al., 2014), vehicles angle crashes (Abdel-Aty & Haleem, 2011), crashes at traffic zones (Siddiqui, et al., 2012), and air traffic delays (Rebollo & Balakrishnan, 2014). This

73 section of the paper begins with a discussion of decision trees. Then, we describe random forests model and variable importance measure Decision Tree Decision trees can work for both regression and classification problems. For the sign data, the response (sign condition rating) is categorical, and hence, belongs to a classification issue. A decision tree segments the predictor space into a number of smaller distinct and non-overlapping regions, with each region produced as a function of explanatory variables. Thus, a decision tree has a hierarchical structure consisting of a number of nodes, whereas splitting starts at the top of the tree with root node, continues at internal nodes, and ends at the bottom with leaves or terminal nodes. The splitting process is achieved by purifying the leaf node, i.e. trying to put the observations belonging to the same class into the same node. To do so, Gini index is used to give a measurement for the purity level of each leaf node. Gini index is defined by: G = p mk(1 p mk) k (3-1) Here p mk is the proportion of observations in the m th leaf node and k th class. In summary, the process of building a decision tree includes two steps: (1) dividing the predictor space, and (2) making the same prediction for all observations in the same region (James, et al., 2013).

74 Random Forests Although a decision tree has various advantages when compared with other statistical methods, over fitting and less prediction accuracy may be its shortcoming. After a minor change in the data, the results obtained from a decision tree may change significantly. Random Forests (RF) improve prediction accuracy of a decision tree by involving a very large number of decision trees (James, et al., 2013). Generally, Random Forests tend to have lower variance by taking repeated samples from a single data set and combining them together. When developing a RF model, two randomization schemes are processed. Firstly, a subset of m predictors is randomly selected from the full set of p predictors in each split step. By considering the random subset instead of the full set, each predictor is given equal chance of consideration. Thus, it avoids the generation of similar trees because the most significant predictors are selected in each tree. Secondly, random bootstrap samples are generated to construct the trees. The bootstrap samples can be used to compute predictor error easily and minimize variance. The remaining nonselected sample is called an Out-of-Bag sample (James, et al., 2013; Palczewska, et al., 2013) Variable Importance Measure (VIM) Although random forests have higher accuracy than the single decision tree, the interpretation of random forests is cumbersome, due to the involvement of hundreds of trees. To be more specific, the obvious structure easily observed in one tree is impossible to be concluded from multiple trees because each tree comes up with different structure.

75 61 In other words, random forests improve prediction accuracy at the expense of interpretability (James, et al., 2013). To address this challenge, the ranking based on variable importance measure (VIM) of each predictor is provided. A larger importance value indicates that the predictor has a more important role on the response. The variable importance measure is related to the total amount of decrease in the Gini index caused by each predictor when splitting. To obtain the VIM, all of the differences in prediction accuracy before and after permuting the variable x i should be averaged. In other words, VIM of x i can be defined as (Louppe, et al., 2013): N VIM(x i ) = 1 N p(n) G(s n, n) n=1 n N:y(s n )=x i (3-2) Where N equals number of trees, p(n) G(s n, n) is the weighted impurity decreases (for classification trees, Gini index is used to measure impurity) for all nodes n that x i is employed. Also, p(n) is the proportion of samples touching n, and y(s n ) is the variable used to split s n. Ultimately, all the variables are ranked from most important to least important through applying the VIM. After examining the data more closely, it was observed that predictors were measured with very different units. In order to avoid possible bias caused by a variety of scales, a standardization of predictors was conducted with each variable subtracting its mean and dividing by its standard error. After standardizing, measurements of all predictors ranged from -1 to +1. Then, a random forests package was created in R

76 62 software (R, 2014). The subset of variables considered in each splitting is suggested to be m = p (p = 10, m = 4) in previous studies (James, et al., 2013). As for the number of trees, no particular rule or optimal number was suggested in the literature. A larger number of trees does not lead to consistently better performance (Oshiro, et al., 2012). I finally defined the number of trees as 1500, based on similar studies. To make interpretation easier, importance values were also normalized, such that the most important predictor had an importance of 100 (Rebollo & Balakrishnan, 2014). 3.5 Results and Discussion Figure 3-8 displays the variable importance measure for each predictor obtained from the random forests model. We observed that the height of sign above the road is the most important predictor for a sign s visual condition. This result may reflect the impact of environmental and vandalism damages based on sign mount height. For example, only traffic signs close to the ground may be bent or knocked down by vehicles or damaged by tree rubbings and tree sap. Moreover, vandalism damage on the face of traffic signs, such as paintball damage, ballistic damage from firearms, glass bottle impacts, eggs, stickers, dents, graffiti, and paint are more frequent on ground-mounted signs (Boggs, et al., 2013). Ground elevation is ranked as the second most important factor in Figure 3-8. As mentioned earlier, by increasing ground elevation, both UV radiation and snow frequency increased, which explains why elevation plays a critical role in sign visual condition. A number of sign damages occur because of snowplows throwing snow and roadway debris

77 63 against the face of signs. A study proved that an increase in elevation typically is associated with an increase in precipitation in Utah (Boggs, et al., 2013). Temperature and precipitation are the next most important factors in Figure 3-8. Besides their critical importance, the closeness of these two rankings may also indicate an inner correlation or interaction between these two predictors. It is known that decreasing temperatures increases the possibility of snowfall. The majority of the signs in Utah experience seasonal temperature swings spanning an immense 50 to 64 degrees of Fahrenheit (Boggs, et al., 2013; Zolghadri, et al., 2016b). Generally, when a sign becomes cut, cracked, or punctured, water accumulates within the layers of sign sheeting. Thus, after weathering for several seasons, sign legibility deteriorates, ultimately leading to a decline in performance and sign failure. These findings were similar to those of Black et al. (1992) and Hildebrand (2003), who reported that climate and location had significant effects on sign performance. As seen in Figure 3-8, the facing direction, size, land cover, and color of a sign are not as particularly important as the aforementioned factors to traffic sign failure.

78 64 Elevation Temperature Precipitation Coarse Particulate Matter Facing Direction Sign Width Sign Length Land Cover Sign Color Figure 3-8 Variable importance plot for traffic sign s visual condition ratings 3.6 Conclusions The goal of the study was to examine the varying importance of multiple explanatory variables to traffic sign s visual condition. At the completion of our project, it was found that the height of sign above the road and the ground elevation were the most important variables. In terms of effects on sign degradation, the results also showed that climate and localized conditions were more important than sign attributes. Remarkably, it was revealed that air pollutants were among the most important contributing factors to traffic sign deterioration. Based on the findings, traffic signs that more frequent inspection may be warranted for them include ground-mounted signs that are installed in areas with high ground elevation, and are exposed to cold weather, and experience frequent snowfall. The results of the study can also provide reliable

79 65 theoretical guidance for any further manual experiments. When transportation agencies perform this experiment manually, the findings from this study may provide some of the necessary preliminary information about which factors are important and worthy of the time and cost to measure and study.

80 66 CHAPTER 4 PREDICTION OF TRAFFIC SIGN VANDALISM THAT OBSTRUCTS CRITICAL MESSAGES TO DRIVERS 4.1 Introduction In order for road users to comprehend traffic signs, high legibility and visibility are critical (Ye, et al., 2014; Balali, et al., 2015). The overall legibility of a sign declines when the face of a sign is damaged (Boggs, et al., 2013). Depending on the form of damage, the effects on legibility vary considerably (Khalilikhah, et al., 2016; Khalilikhah & Heaslip, 2016d). The overall day and nighttime legibility of the sign can be affected by vandalism (Evans et al., 2012). In addition, tremendous amounts of money are spent to repair or replace vandalized signs (Harris, 1992). Types of human vandalism include shooting paintballs, shooting bullets, throwing beer bottles, putting stickers on signs, and painting graffiti on signs (Figure 4-1). Although traffic sign vandalism has become a serious problem for traffic agencies (Chadda & Carter, 1983; Harris, 1992), few studies are conducted that focus on this issue. Previous studies estimated the costs of sign vandalism (Smith & Simodynes, 2000), developed methods for sign vandalism detection (Mueller, 1995), and examined the effects of releasing information in the media to reduce sign vandalism (Ellison, 1996). The countermeasures against vandalism have been discussed with regard to the form of sign vandalism, including utilizing more resistant materials to construct signs, mounting signs higher, applying penalty notices to signs, and using public information campaigns (Picha, 1997; Perkins & Barton, 1997). However, it

81 is necessary to identify traffic signs that are more vulnerable to vandalism before installing them. 67 Figure 4-1 Samples of traffic sign vandalism (Source: Photos taken by Utah State University research team) After focusing on the sign data collected across the state of Utah, it was observed that almost 7% of 97,314 measured signs were damaged, of which at least 22% of the damages were intentionally caused by humans. Due to a lack of detailed information about traffic sign vandals, this chapter aims at answering the following questions: How does the vandalism rate change with respect to the demographics of the local population where the sign is placed? What traffic signs are more likely to be damaged by vandals? To accomplish this goal, creating contingency tables is necessary to yield the desired results. U.S. Census Bureau data were used to obtain the demographic data, including population, ethnicity, age, income, education, and gender composition data. To identify more significant demographics associated with sign vandalism, the chi-square and trend statistic tests with respect to the categorical variables were applied. Based on

82 68 these findings, a linear regression model was finally developed to predict the rate of sign vandalism with respect to these demographics. In addition, mobile-based data provide sign attributes data, including background color, sign size, and mount height. Location data obtained from online sources combined with the traffic sign data was imported into ArcGIS to acquire localized conditions for each individual sign. Figure 4-2 shows the locations of vandalized traffic signs. Figure 4-2 Locations of vandalized traffic signs in Utah

83 Vandalism by Demographics Data Table 4-1 summarizes the number of damaged signs for each of the 29 counties across the state of Utah. Since the locations of measured traffic signs (latitude and longitude) were recorded by the equipped vehicle, traffic sign data were imported into ArcGIS software to extract the counties in which the signs were placed. As shown in Table 4-1, the rate of damage and vandalism differed by county. After examining the results more closely, diverse vandalism rates were illustrated by counties, in which Piute County had, by far, the highest rate of sign vandalism, and Davis County the lowest. In order to assess the association between the vandalism rate and the socioeconomic characteristics of people living in the counties where traffic signs are placed, U.S. Census Bureau data could be useful (Table 4-2). Online sources were used to obtain various desired demographic data from U.S. Census data (Utah Median household income, 2010; Matthews, 2010; Age Distribution in Utah, 2011; Bureau of Economic and Business Research University of Utah, 2011; Howden & Meyer, 2010; United States Census Bureau, 2014). Utah is composed of 29 counties, with an overall population of 2,763,885 people, based on the 2010 United States Census Bureau data. County populations in the state of Utah range from about 1,000 to 1,000,000 people. Salt Lake County is the most populated county, wherein approximately 37% of Utah s inhabitants reside. The population of four counties, including Daggett, Piute, Rich, and Wayne Counties, each one is less than 5,000. Moreover, Davis County is the smallest county, with an area of 774 km 2, whereas San Juan County is the largest, with 20,253 km 2. In this study, area and population data were obtained to examine whether a county s

84 70 population density and the number of people per unit of area influenced its vandalism rate. In general, Utah is not a high-density state. With an average population density of person per square kilometer, Utah is ranked 41st out of the 50 US states. To provide some perspective, the national average population density is (WorldAtlas, 2014). Table 4-1 Damaged signs by county County Damage # of % % Aging/ Signs None Vandalism Unknown Damage Vandalism Environmental Beaver 1,585 1, Box Elder 5,152 4, Cache 2,821 2, Carbon 2,130 1, Daggett 1,180 1, Davis 5,279 5, Duchesne 2,038 1, Emery 3,071 2, Garfield 3,216 3, Grand 2,514 2, Iron 3,105 2, Juab 1,642 1, Kane 1,731 1, Millard 3,698 3, Morgan 1, Piute Rich Salt Lake 15,790 15, San Juan 3,910 3, Sanpete 2,198 2, Sevier 4,198 3, Summit 3,715 3, Tooele 3,118 2, Uintah 1,838 1, Utah 8,711 8, Wasatch 1,665 1, Washington 3,481 3, Wayne 1,592 1, Weber 4,950 4, Total 97,314 90,

85 County Table 4-2 Demographic characteristics by county in Utah Area (sq km) Population % of Ethnic Majority Median Age (yr) Median Household Income, (USD) % of Adults with Associate Degree, or Higher Gender Ratio (M:F) Beaver 6,708 6, , Box Elder 14,880 49, , Cache 3, , , Carbon 3,829 21, , Daggett 1,805 1, , Davis , , Duchesne 8,394 18, , Emery 11,557 10, , Garfield 13,403 5, , Grand 9,509 9, , Iron 8,538 46, , Juab 8,786 10, , Kane 10,334 7, , Millard 17,022 12, , Morgan 1,578 9, , Piute 1,963 1, , Rich 2,664 2, , Salt Lake 1,922 1,029, , San Juan 20,253 14, , Sanpete 4,118 27, , Sevier 4,948 20, , Summit 4,848 36, , Tooele 17,977 58, , Uintah 11,602 32, , Utah 5, , , Wasatch 3,044 23, , Washington 6, , , Wayne 6,373 2, , Weber 1, , , (Source: U.S. Census Bureau) 71

86 72 By considering people of all backgrounds, Hispanics, Blacks, Indians, and Asians make up 20% of the total population of Utah, while Whites make up 80% of the population (Bureau of Economic and Business Research University of Utah, 2011). There is a higher percentage of Whites in Utah than in the United States in general (63%). San Juan County is the only county in Utah where Whites are not the majority group (43.9%). However, 74% of Salt Lake City s residents are White. Morgan County has the highest percentage of Whites (96.1%). Of all the United States, Utah is also the youngest state, with a median age of 29.2 years old, while the national median age is 37.6 (Age Distribution in Utah, 2011). The state s youngest population is centered in Utah County, with a median age of 24.6, followed closely by Cache County (25.5). Kane, Daggett, and Piute Counties have the highest median ages in Utah. With respect to the 5-year estimates of median household income obtained by the U.S. Census Bureau, the average income value for Utah ($56,330) is higher than the national value of $51,914 (Utah Median Household Income, 2010). As defined by the American Community Survey (ACS): Income of household includes the income of the householder and all other individuals 15 years old and over in the household, whether they are related to the householder or not (Utah Median Household Income, 2010). Summit is the richest county in Utah, with a household median income of $79,461. This is followed by Morgan County ($70,152). The values for Daggett, Piute, and San Juan Counties are less than half of that of people living in Summit. The percentage of Americans between the ages of 25 and 64 with at least a two-year college degree is 38.7 percent, whereas 40.3% of the working-age adults in the state of Utah hold at least an

87 73 associate degree (Matthews, 2010). In addition, residents of Beaver and Summit Counties have respectively the least and highest rates of educated people. By considering gender composition data, the gender ratio (male to female) for the state of Utah is 1.009, higher than the U.S. gender ratio of (United States Census Bureau, 2014). The following subsections discuss the rates of the surveyed vandalism, with special attention given to the demographics of the state of Utah. The chi-square test was used to discuss the association between the two categorical variables. In addition, a trend test was applied to increase the power of analysis. According to Agresti (2007), having a 2 x J (or I x 2) table with categorical column (row) variable Y that fairly shows a trend (increase or decrease) in the studied feature, the power of analysis can be increased by using a test statistic with a more specific alternative, such as trend tests. All of the following maps were created using ArcGIS to visualize the association between counties demographics and vandalism rates. Counties are shaded based on vandalism rate while the size of the symbols reflect the variation in each demographic characteristic Population Density By using the population density of counties, the association between vandalism rate and population can be determined. To do this, population and area data were obtained from the 2010 United States Census Bureau data. Population density (people per km 2 ) by county across the state is displayed in Figure 4-3, and the effects of counties population density on the vandalism rate are summarized in Table 4-3. As seen in Table 4-3, the denser counties were less likely to experience sign vandalism. According to the

88 74 results of the chi-square test, there is strong evidence of an association between vandalism rate and population density. The trend test illustrated that this association could be linear, whereas in denser counties, the rate of vandalism will decrease. This finding was similar to those of Black, et al., (1992) and Boggs, et al., (2013), who reported that vandalism was more frequent in rural canyon areas Ethnicity The question of interest for this section is whether or not traffic sign vandalism in Utah corresponds to a specific community living in the state. Previously, a number of studies found that race plays a role in the road users behavior, probably due to its impact on drinking (Harper, et al., 2000; Romano, et al., 2006). In Figure 4-4, the percentage of White people in comparison with other races across the state is demonstrated. The other groups include Hispanics, Blacks, Indians, Asians, and various other races (Bureau of Economic and Business Research University of Utah, 2011). Almost 20% of the people living in Utah are minorities. Table 4-4 shows whether or not vandalism rate is affected by the ethnicity of the counties in Utah. Although the chi-square value is statistically significant, its value is not large. Also, there is no evidence of linear association between vandalism and ethnicity after considering the trend test results because of the high p- value. As a result, race does not play a significant role in traffic sign vandalism in Utah.

89 75 Table 4-3 Vandalism rates by population density Density Vandalism # of (people per km 2 Signs Yes No ) % Vandalized <1 24, , , , , , , , >100 26, , Chi-Square Test Statistic= (P-Value < ) Trend Test Statistic = (P-Value < ) Figure 4-3 Population density by county

90 76 Table 4-4 Vandalism rates by ethnic majority % Major # of Vandalism Group Signs Yes No % Vandalized <80 24, , , , , , >90 19, , Chi-Square Test Statistic= (P-Value < ) Trend Test Statistic = (P-Value = 0.20) Figure 4-4 Ethnicity by county

91 Age Previously, a study addressed the effects of road users age on the understanding of traffic signs (Ng & Chan, 2008). However, research that assesses the association between age and vandalism rates is not well-developed. To fill this gap, this study derived the median age data for each county from the 2010 U.S. Census (Age Distribution in Utah, 2011). Utah has a median age equal to 29.2 years old, less than the national median age of 37.6 years. Figure 4-5 demonstrates the median age of the State of Utah by county. Table 4-5 summarizes the effects of each county s median age on their vandalism rate. Similar to ethnicity, the chi-square value is statistically significant, but it is not large; the same is true for the value of trend test. It is perhaps a reflection of the narrow range of variability of median age in Utah s counties Income There have been studies that evaluated the correlation between income and driving safety in the past (Traynor, 2008). In order to examine the association between income and vandalism, this study used the Utah median household income levels collected in the American Community Survey (ACS). Generally, 10.8% of people living in Utah are considered to be living in poverty. In the U.S., poverty status is determined by comparing annual income to a set of dollar values called poverty thresholds that vary by family size, number of children and age of householder (Utah Poverty Rate by County, 2010). For this study, data consisted of the values between the years of 2006 and Recorded income levels include the income of the householder and all other

92 78 individuals 15 years old and over in the household, whether they are related to the householder or not (Utah Median household income, 2010). Figure 4-6 displays the median incomes by county. A summary of the vandalism rates based on income is presented in Table 4-6. The result of the chi-square test shows evidence of a very strong association between the vandalism rate and median income, whereas their association is linear, based on the trend test. In other words, there is statistically evidence that higherincome counties were less likely to experience traffic sign vandalism Education Education level is another factor that may affect the rate of vandalism. According to the 2011 Census data, by focusing on Utah s working-age adults (25-64 years old), 40.3% of people across the state hold at least an associate degree (Matthews, 2010). Percentages of educated adults by county are shown in Figure 4-7. The effects of education level on the vandalism rate are summarized in Table 4-7, wherein counties with a higher percentage of educated people were less likely to experience sign vandalism. According to the results of the chi-square test, there is evidence of a strong association between the vandalism rate and education level. Trend tests illustrated that there is also statistical evidence that their association is linear. Based on the trend test results, by increasing the percentage of people with at least an associate degree, the rate of vandalism will decrease. This finding was also reported by Lewis, who stated that education plays a vital role in effectively addressing the sign vandalism issue (Lewis, 1998).

93 79 Table 4-5 Vandalism rates by median age Median Age # of Vandalism % (years) Signs Yes No Vandalized <29 16, , , , , , >34 17, , Chi-Square Test Statistic= (P-Value < ) Trend Test Statistic = 3.90 (P-Value < ) Figure 4-5 Median age by county

94 80 Median Household Income, (USD) Table 4-6 Vandalism rates by income Vandalism # of Signs Yes No % Vandalized <42,000 12, , ,000-45,000 13, , ,000-50,000 11, , ,000-55,000 13, , ,000-60,000 31, , >60,000 14, , Chi-Square Test Statistic= (P-Value < ) Trend Test Statistic = (P-Value < ) Figure 4-6 Household income by county

95 81 Table 4-7 Vandalism rates by education level Vandalism # of Signs Yes No % Adults (25-64) with at least an Associate Degree % Vandalized <30 17, , , , , , , , >42 22, , Chi-Square Test Statistic= (P-Value < ) Trend Test Statistic = (P-Value < ) Figure 4-7 Education by county

96 Gender In order to investigate whether or not a county s gender ratio influences the traffic sign vandalism rate, gender composition data from the 2010 Census data were analyzed. The ratio of males to females for Utah is 1.009, higher than the national gender ratio (0.967) (Howden & Meyer, 2010). By focusing on Table 4-8, there is evidence of an association between the gender ratio and the vandalism rate in Utah s counties. The value of the trend test is not large, although it is statistically significant. However, drawing conclusions based on Utah s gender composition is not reliable, due to the uniformity of counties sex ratio. Gender Ratio (Males to Females) Table 4-8 Vandalism rates by gender Vandalism # of Signs Yes No % Vandalized <1 13, , , , , , > , , Chi-Square Test Statistic= (P-Value < ) Trend Test Statistic = 2.73 (P-Value = ) Regression Model To provide more robust evidence for the association between vandalism rate and demographics of local population, a regression model was developed. However, after considering the scatter plot of predictors, a majority of variables had a high correlation with other predictors. Creating a correlation matrix, a high correlation between predictors

97 83 was also evident. Thus, it was necessary to eliminate confounding from the fitted model by removing highly correlated variables. The model development began by considering all predictors. Table 4-9 shows the fitted model, wherein no predictor was statistically significant. By taking into consideration backward elimination procedure, insignificant predictors were dropped from the model one after another. In other words, the variables without enough statistically significant value were eliminated from the model (based on the t-test, final variables were significant at a 5% level). As seen in Table 4-9, income was the only included predictor in the final model. Upon completing the linear regression analysis, by increasing the percentage of people who have families with a higher income in the county, the rate of vandalism will decrease. With just one predictor, the final model yielded a good value of R 2, In addition, the F-test for lack of fit was equal to 10.3 for a p-value less than 0.001, which showed the developed model provides a good fit. Table 4-9 Regression models of sign vandalism rate All Variables Significant Variables Estimate Pr(> t ) Estimate Pr(> t ) Intercept * <0.001 Population Density Ethnicity Median Age Income * <0.001 Associate Degree Gender F-Statistic * <0.001 R *P-Value < 0.05, statistically significant at level of 0.05

98 Vandalism by Sign Attributes Table 4-10 provides a summary of the vandalized traffic signs based on the Manual on Uniform Traffic Control Devices (MUTCD) types. The Manual of Traffic Signs website was used to list traffic signs based on their type and sub-type (Manual of Traffic Signs, 2014). Respectively, regulatory, warning, marker, and guide signs made up approximately 18%, 20%, 20%, and 23% of over 97,000 measured traffic signs. As a whole, warning signs, by far, exhibited the highest vandalism rate. Of vandalized signs, 53% were warning signs. With regard to the percentage of sign vandalism, turn and curve warning signs showed the highest rates, comprising approximately 30% of vandalized signs. Almost 8% of the vandalized signs were advance warning/crossing, 7% were speed regulation signs and 9% were object markers. Interestingly, nearly 11% of vandalized signs were signs that deal with speed limits (speed regulation and advisory speed signs). Based on the sign legend type, vandalized signs were categorized into four groups: text, symbol, arrow, and text/symbol/arrow. To be categorized as a text sign, these signs need words or digits to accomplish their tasks. Examples of text signs included speed limit signs (R2-1), mileposts (D10-1), and supplemental distance signs (W16-2 & 3). Symbol signs consisted of those that use symbols, rather than words or digits to interact with road users. Examples included school signs (S1-1), no pedestrian signs (R9-3a), and cattle or deer crossing signs (W11-3 & 4). The arrow category included any traffic signs employing arrows to regulate, warn, or guide drivers. For example, straight optional lane signs (R3-6L or R), reverse turn signs (W1-3L or R), and arrow auxiliary signs (M6-1L or R). The data also identified a final group of signs that

99 85 included a combination of text, symbols, and arrows. These signs were included in the text/symbol/arrow grouping, such as do not enter signs (R5-1), stop ahead signs (W3-1), and destination with distance signs (D1-1, 2, & 3). Ultimately, it was found that arrow signs had the highest rate of vandalism, followed by text signs (Figure 4-9). The question of interest was if traffic sign vandalism corresponds to specific types of signs. In other words, questions to be answered include: do vandals select traffic signs based on sign color, size, or mount height? Or do localized conditions, such as exposure (urban or rural) and road type (major or ramp) make them more vulnerable to vandalism? The next sections find answers for these questions Text Symbol Arrow Text/Symbol/Arrow Figure 4-8 Number of vandalized traffic signs by legend type

100 86 Regulatory Signs Warning Signs Marker Signs Guide & Information Signs Table 4-10 Summary of vandalized signs by MUTCD type Traffic Sign Name MUTCD Code # of Signs Text Symbol Arrow Text/ Symbol/ Arrow Stop & Yield R1 Series % Speed Regulation R2 Series % Turn & Land Use R3 Series % Movement Regulation R4 Series % Selective Exclusion R5 Series % One Way R6 Series % Pedestrian & Bicycle R9 Series % Traffic Signal R10 Series % Road Closed R11 Series % Turn & Curve W1 Series % Intersection W2 Series % Advance Traffic Control W3 Series % Merge & Lane Transition W4 Series % Divided Highway W6 Series % Hill W7 Series % Pavement Condition W8 Series % Railroad & Light rail W10 Series % Advance Warning / Crossing W11 Series % Low Clearance W12 Series % Advisory Speed W13 Series % Dead End / No Outlet / No W14 Passing Series % Supplemental Plaques W16 Series % Route Markers M1 Series % Junction Signs M2 Series % Cardinal Direction Auxiliaries M3 Series % Advance Turn Auxiliaries M5 Series % Directional Arrow Auxiliaries M6 Series % Object Markers OM Series % Destination D1 Series % Distance D2 Series % Recreational D7 Series % General Services D9 Series % Mileposts D10 Series % Crossover / Freeway Entrance D13 Series % Interchange Advance E1 Series % Exit Gore E5 Series % Destination E6 Series % Destination E10 Series % General Information I Series % S1, S3, School Signs and S % Series Other Signs % %

101 Sign Background Color Table 4-11 depicts a summary of the sign vandalism rates based on the sign background color. Warning signs are typically yellow (Manual of Uniform Traffic Control Devices, 2012). Thus, yellow signs tend to have a relatively higher vandalism rate. According to the results of the chi-square test, there is strong evidence of an association between sign vandalism rate and sign background color. Table 4-11 Traffic sign vandalism by color Color # of Vandalism % Signs Yes No Vandalized Green 21, , Red 1, , White 28, , Yellow 22, , Others 22, , Chi-Square Test Statistic = P-Value < Sign Length and Width Tables 4-12 and 4-13 display the corresponding sign vandalism ratings for each category of length and width. Generally speaking, the percentage of vandalized signs changes little among different categories of length or width, with the exception of signs with width from 24 to 36 inches and length of between 30 and 40 inches, which is mostly the size of warning signs. The chi-square value is statistically significant. Thus, there is evidence of an association between sign size and vandalism rate.

102 88 Table 4-12 Traffic sign vandalism by width Sign # of Vandalism % Width (in) Signs Yes No Vandalized <18 19, , , , , , , , >54 13, , Chi-Square Test Statistic = P-Value < Table 4-13 Traffic sign vandalism by length Sign # of Vandalism % Length (in) Signs Yes No Vandalized <20 24, , , , , , , , >60 5, , Chi-Square Test Statistic = P-Value < Sign Mount Height According to the summary of sign vandalism by mount height (Table 4-14), signs placed higher were less likely to get vandalized. For signs placed 10 feet or more above the road, the vandalism rate was only 0.12%. Based on the results of the chi-square test, there is evidence of an association between sign vandalism rate and mount height Exposure A variable was defined called sign exposure (urban or rural) with respect to the area that the traffic sign was installed. To obtain sign exposure data, the Geographic

103 89 Information Database of Utah s Automated Geographic Reference Center (AGRC, 2008) website was used. Then, rural and urban signs were identified using ArcGIS. A summary of the traffic signs vandalism by exposure is provided in Table As seen in the table, the number of vandalized signs for rural exposure is indeed higher than for urban areas. The chi-square value was also statistically significant. Therefore, the association between sign exposure and number of vandalized signs was evident. Figure 4-10 shows the municipalities in the state of Utah. Table 4-14 Traffic sign vandalism by mount height Sign Height Vandalism # of % above Road Signs Yes No Vandalized (ft) <5 17, , , , , , , , >10 12, , Chi-Square Test Statistic = P-Value < Table 4-15 Traffic sign vandalism by sign exposure Exposure # of Vandalism % Signs Yes No Vandalized Urban 46, , Rural 50,703 1,090 49, Chi-Square Test Statistic = P-Value <0.0001

104 90 Figure 4-9 Municipalities in Utah Road Type The (AGRC, 2008) data set was used to extract the type of road that traffic sign was installed on. To do so, ArcGIS was employed and ultimately traffic signs were categorized into two groups based on where they were placed. Category one was major road signs (87% of the measured signs), and category two was signs placed in ramps, rest

105 91 areas, or turnarounds (13% of the measured signs) (Table 4-16). After running the chisquare test, the association between road type and sign vandalism rate was evident. The rate of sign vandalism for major road signs was higher than the other signs. Table 4-16 Traffic sign vandalism by road type # of Vandalism % Road Type Signs Yes No Vandalized Major 84,423 1,412 83, Ramp (on/off) 12, , Chi-Square Test Statistic = P-Value < Discussion Although all sign attributes were important to vandalism rates, the height of sign above the road was, by far, the most important variable. The importance of sign mount height reflects the fact that regardless of sign color, size and localized conditions, vandalism damage on the face of traffic signs is more frequent on ground mount signs. Despite this, the strong association between mount height and vandalism rate did not seem to be linear because the rate of vandalism for signs installed within 5 feet of the ground was less than those between 5 and 7 feet. While the average mount height for all measured signs was about 8 feet, the average height of vandalized signs was 6.5 feet above the road. The closeness of rankings of sign size, color, land cover, and road type may indicate an inner correlation or interaction between those predictors. To enable more in-depth analysis, the possible relationship between the most important variable, sign

106 92 mount height, and other factors should be studied. As shown in Figure 4-11, nearly all of the vandalized signs located in rural areas or installed on ramps had a mount height less than 10 feet. Thus, mounting these signs higher can be a good countermeasure against sign vandalism. In addition, 53% of vandalized signs were warning signs. Turn and curve warning signs that comprised approximately 30% of vandalized signs, have an average mount height of 6.6 feet with 1.5 standard deviation. As a result, warning signs can also be mounted higher. However, most of warning signs are located on minor or local roads. Taking into consideration the signs located on ramps or in rural areas, increasing sign heights may dramatically affect sign visibility (especially during hours of darkness) since headlights will not reach higher sign heights. To address this issue, traffic signs can be equipped with internal or external lightening systems. Fig Vandalized signs by mount height vs. exposure and road type