Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 Pavement Design, Construction, and Condition Evaluation Selected Papers from the Proceedings of the International Airfield and Highway Pavements Conference 2021 >> June 8–10, 2021 Edited By Hasan Ozer, Ph.D John F Rushing, Ph.D., P.E Zhen Leng, Ph.D Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved AIRFIELD AND HIGHWAY PAVEMENTS 2021 PAVEMENT DESIGN, CONSTRUCTION, AND CONDITION EVALUATION SELECTED PAPERS FROM THE INTERNATIONAL AIRFIELD AND HIGHWAY PAVEMENTS CONFERENCE 2021 June 8–10, 2021 SPONSORED BY The Transportation & Development Institute of the American Society of Civil Engineers EDITED BY Hasan Ozer, Ph.D John F Rushing, Ph.D., P.E Zhen Leng, Ph.D Published by the American Society of Civil Engineers Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Published by American Society of Civil Engineers 1801 Alexander Bell Drive Reston, Virginia, 20191-4382 www.asce.org/publications | ascelibrary.org Any statements expressed in these materials are those of the individual authors and not necessarily represent the views of ASCE, which takes no responsibility for any statement made herein No reference made in this publication to any specific method, product, process, or service constitutes or implies an endorsement, recommendation, or warranty thereof by ASCE The materials are for general information only and not represent a standard of ASCE, nor are they intended as a reference in purchase specifications, contracts, regulations, statutes, or any other legal document ASCE makes no representation or warranty of any kind, whether express or implied, concerning the accuracy, completeness, suitability, or utility of any information, apparatus, product, or process discussed in this publication, and assumes no liability therefor The information contained in these materials should not be used without first securing competent advice with respect to its suitability for any general or specific application Anyone utilizing such information assumes all liability arising from such use, including but not limited to infringement of any patent or patents ASCE and American Society of Civil Engineers—Registered in U.S Patent and Trademark Office Photocopies and permissions Permission to photocopy or reproduce material from ASCE publications can be requested by sending an e-mail to permissions@asce.org or by locating a title in ASCE's Civil Engineering Database (http://cedb.asce.org) or ASCE Library (http://ascelibrary.org) and using the “Permissions” link Errata: Errata, if any, can be found at https://doi.org/10.1061/9780784483503 Copyright © 2021 by the American Society of Civil Engineers All Rights Reserved ISBN 978-0-7844-8350-3 (PDF) Manufactured in the United States of America Airfield and Highway Pavements 2021 iii Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Preface Airfield and highway pavements are critical components of our transportation infrastructure Increasing demand on these assets creates a unique challenge for researchers and practitioners to find sustainable solutions to managing their life-cycle The airfield and highway pavements specialty conference is a unique setting where the world’s foremost experts in pavement design, construction, maintenance, rehabilitation, modeling, management, and preservation meet and present most recent developments in the pavement engineering area Building on the success of our past conferences, the 2021 International Airfield and Highway Pavements Conference of ASCE’s Transportation and Development Institute (T&DI) displayed the adaptive nature of our profession as we held our first completely virtual event from June 8-10, 2021 The 2021 virtual conference was designed to feature plenary sessions and panel discussions on important topics facing government agencies and industry Technical breakout sessions allowed researchers and practitioners to present deeper technical content on breakthrough practices and technologies The virtual poster session allowed “on-demand” access to cutting edge research The proceedings of the 2021 International Airfield and Highway Pavements Conference have been organized into three publications and are described as follows: Vol I: Airfield and Highway Pavements 2021: Pavement Design, Construction, and Condition Evaluation This volume includes papers concerning mechanistic-empirical pavement design methods and advanced modeling techniques for highway pavements, construction specifications and quality monitoring, accelerated pavement testing, rehabilitation and preservation methods, pavement condition evaluation, and network-level management of pavements Vol II: Airfield and Highway Pavements 2021: Pavement Materials and Sustainability This volume includes papers describing laboratory and field characterization of asphalt binders, modifiers and rejuvenators, asphalt mixtures and modification, recycled and waste materials in asphalt mixtures, unbound base/subgrade materials and stabilization, pavement life-cycle management, interactions of pavements and their environment, and recent advances in cementitious materials characterization and concrete pavement technology In this volume, we also included papers introducing cutting edge innovative and sustainable technologies used in pavement applications Vol III: Airfield and Highway Pavements 2021: Airfield Pavement Technology This volume includes papers on recent advances in the area of airfield pavement design, construction, and rehabilitation methods, modeling of airfield pavements, use of © ASCE Airfield and Highway Pavements 2021 accelerated loading systems for airfield pavements, and airfield pavement condition evaluation Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved The papers have undergone a rigorous peer review by at least two to three international highway pavement and airfield technology experts and a quality assurance process before becoming a publication of ASCE – the world’s largest publisher of Civil Engineering content The success of the conference is a tribute to the incredible efforts of the leadership team consisting of Conference Co-Chairs (Hasan Ozer, John Rushing, and Zhen Leng) and Advisory Board (Imad Al-Qadi and Scott Murrell) along with an outstanding Conference Steering Committee (Amit Bhasin, Rick Boudreau, Zeijao Dong, Jeffrey Gagnon, Tom Harman, Andreas Loizos, Geoffrey Rowe, Injun Song, Leif Wathne, and Richard Willis) and terrific support from ASCE T&DI staff The efforts of the Conference Scientific Committee are graciously acknowledged for their role in reviewing papers and providing critical feedback to the authors We thank everyone who attended the virtual conference and hope to see everyone again in 2023! Hasan Ozer, Ph.D., A.M.ASCE, University of Illinois at Urbana-Champaign John Rushing, Ph.D., P.E., M.ASCE, U.S Army Engineer Research and Development Center Zhen Leng, Ph.D., M.ASCE, Hong Kong Polytechnic University © ASCE iv Airfield and Highway Pavements 2021 v Contents Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Influence of Tire Footprint Contact Area and Pressure Distribution on Flexible Pavement Ainalem Nega and Daba Gedafa Monte Carlo Simulation for Flexible Pavement Reliability .13 Anastasios M Ioannides and Jeb S Tingle Calibration of Transverse Cracking and Joint Faulting Prediction Models in Pennsylvania for JPCP in AASHTOWare Pavement ME Design 26 Biplab B Bhattacharya and Michael I Darter Calibration of Fatigue Cracking and Rutting Prediction Models in Pennsylvania Using Laboratory Test Data for Asphalt Concrete Pavement in AASHTOWare Pavement ME Design 37 Biplab B Bhattacharya and Michael I Darter Probabilistic Viscoelastic Continuum Damage Approach for Fatigue Life Prediction of Asphalt Mixtures: Challenges and Opportunities .49 Ayat Al Assi, Husam Sadek, Carol Massarra, and Carol J Friedland Multiscale Modeling of Heterogenous AC and Damage Quantification .61 Zafrul Khan and Rafiqul Tarefder Implementation of Fracture Mechanics-Based Reflection Cracking Models for Asphalt Concrete Overlay of Existing Concrete Pavement in AASHTOWare Pavement ME Design 71 Biplab B Bhattacharya; Leslie Titus-Glover, and Deepak Raghunathan Service Life Prediction of Internally Cured Concrete Pavements Using Transport Properties 82 Fariborz M Tehrani Quality Control of Hot Mix Asphalt Pavement Compaction Using In-Place Density Measurements from a Low-Activity Nuclear Gauge 92 Linus Dep, Robert Troxler, Sumuna Mwimba, Chris Croom, and Wes Langston Application of Fourier Transformation to Identify the Onset of Fatigue Damage of Bitumen 103 M Jayaraman and A Padmarekha Characterization of Curling and Warping Influence on Smoothness of Jointed Plain Concrete Pavements 110 Kexin Tian, Bo Yang, Daniel King; Halil Ceylan, and Sungwhan Kim © ASCE Airfield and Highway Pavements 2021 Utilization of Finite Element Analysis towards the Evaluation of the Structural Capacity of Flexible Pavements 120 Nitish R Bastola, Mena I Souliman, and Samer Dessouky Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Intelligent Pavement Roughness Forecasting Based on a Long Short-Term Memory Model with Attention Mechanism 128 Feng Guo and Yu Qian Performance Comparison of HMA Mixes Used on Different Levels of North Dakota’s Highway Performance Classification System 137 Jun Liu, Robeam Melaku, Daba Gedafa, and Nabil Suleiman Influence of Slow-Moving Nature of Super Heavy Load (SHL) Vehicles on the Service Life of Pavement Structures 146 Ali Morovatdar and Reza S Ashtiani Durability of Concrete Pavements Exposed to Freeze-Thaw Cycles in Different Saline Environments 159 Mohammad Pouramini, Fariborz M Tehrani, Saman Sezavar Keshavarz, and Arjang Sadeghi A Quantitative Investigation of the Durability of Asphalt Pavement Materials Using Experimental Freeze-and-Thaw Weathering Data 169 Rojina Ehsani, Alireza Miri, and Fariborz M Tehrani Enhancing the Resilience of Concrete Pavements Using Service Life Prediction Models 178 Sara Kalantari and Fariborz M Tehrani Performance Characteristics of Asphalt Mixes Containing High Percentage of RAP Material .186 Gargi P Jagad, Ambika Behl, and Sanjay M Dave Field Performance Evaluation of Pavement Sections with High Polymer-Modified Asphalt Concrete Overlays .197 Jhony Habbouche, Ilker Boz, Brian K Diefenderfer, and Sayed Adel Comparison of Unconventional and Conventional HMA Mixes’ Performance in North Dakota 209 Rabindra Pariyar, Nabil Suleiman, Daba Gedafa, and Bishal Karki Adoption of 3D Laser Imaging Systems for Automated Pavement Condition Assessment in the United States: Challenges and Opportunities 219 Ryan Salameh and Yichang (James) Tsai GPR Application for Moisture Content Prediction of Cold In-Place Recycling 231 Lama Abufares, Uthman M Ali, Qingqing Cao, Siqi Wang, Xin Sui, and Imad L Al-Qadi © ASCE vi Airfield and Highway Pavements 2021 Impact of Seasonal and Diurnal Profile Measurements on Surface Roughness of Rigid Pavements—LTPP SMP Study 240 Hamad B Muslim and Syed W Haider Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Pavement Image Data Set for Deep Learning: A Synthetic Approach .253 Haitao Gong and Feng Wang Evaluation of the Effect of Interlayer Shear Strength of a Layered Asphalt Pavement on Observed Distresses in the Field: A Case Study 264 R Ghabchi and M Mihandoust Evaluation of Fiber-Reinforced HMA Mixes’ Performance in North Dakota 272 Zeenat Nahar, Nabil Suleiman, and Daba Gedafa The Effects of Automated Vehicles Deployment on Pavement Rutting Performance 280 Ali Yeganeh, Bram Vandoren, and Ali Pirdavani Field Performance Analysis of Open Graded Friction Course: A Case Study in Shreveport, Louisiana 293 Sajidur Rahman Nafis and Nazimuddin M Wasiuddin Comparison of Inverted Pavements with Different Types of Crack Relief Layers .306 Rajan Singh Baghel, Sridhar Kasu Reddy, and Anush K Chandrappa Evaluation of International Roughness Index Measurement Using Cell Phone App and Compare with Pavement Condition Index 317 Mohammad Hossain and Kerrie Schattler Machine Learning Approach to Identifying Key Environmental Factors for Airfield Asphalt Pavement Performance 328 A Z Ashtiani, S Murrell, and D R Brill © ASCE vii Airfield and Highway Pavements 2021 Influence of Tire Footprint Contact Area and Pressure Distribution on Flexible Pavement Ainalem Nega, Ph.D.1; and Daba Gedafa, Ph.D.2 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Dept of Civil Engineering, Curtin Univ., Perth, WA, Australia Email: Ainalem.Nega@curtin.edu.au Dept of Civil Engineering, Univ of North Dakota, Grand Forks, ND Email: daba.gedafa@UND.edu ABSTRACT Effect of tire-pavement contact pressure distribution on flexible pavement is generally complex and dynamic, and it is affected by tire types In addition, there are several inconsistences in data analysis from several experimental studies in measuring contact pressure distribution between tire and pavement The main purpose of this study is to evaluate the influence of tire footprint contact area and pressure distribution on flexible pavements Tirepavement contact stress and interaction model were simulated using 3D finite element for five layers (asphalt concrete, unbound base and subbase, compacted, and natural subgrade) of flexible pavement at various loads An axisymmetric and tire-pavement 3D finite element model was developed A good correlation agreement between contact area and deflection was observed For thin and thick pavement in the static analysis, contact area reduced 3.5% and 3.8%, respectively, while the static deflection for thin pavement decreased from 43.5 mm when 𝐸 = 0.01 to 30.5 mm when 𝐸 = 100 GPa, reduction of 29.9% Whereas for thick pavement, the deflection between static and rolling analysis was not significant, similar trends of deflection between thin and thick pavement were obtained The tire’s finite element model was validated using measured contact area and deflection The results of analysis were then compared to simplify the results of the modeling considering its effects on flexible pavement This finding may have important implication for design of relatively thin asphalt surface layered than thick pavement Keywords: Tire-pavement contact stress; flexible pavement; tire foot print contact area; axisymmetric; thin pavement, thick pavement; finite element model INTRODUCTION The importance of road transportation and development has been growing in the entire world since the past three decades – not only as the result of the development of the road infrastructure but also as a result of the technical development of trucks (Hernandez and Al-Qadi 2016; Hernandez et al 2015; Huhtala et al 1989) Transportation vehicles have become heavier and heavier, and their load – carrying capacity has also become greater and greater Engines are more powerful, cabs more comfortable, and important developments have been made in axles, tires and suspensions (Hernandez and Al-Qadi 2016; Huhtala et al 1989) Where the deteriorating flexible pavement infrastructure and investigation load conditions are, tire-pavement contact stresses usually receives a special attention (Hernandez and Al-Qadi 2016) not clear Pavement contact stresses are not the only directly affect related to several and/or various types of distresses as discussed by Hernandez’s but are also the only feasible © ASCE Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 manner to compare the effect of various tire types on pavement damage (Al-Qadi and Wang 2011; Nega and Nikraz 2017) such as conventional dual-tire and wide-base tires The increased traffic roads and heavier urban vehicles cause much more distress to road pavements than ever before in history The regulation of weights and dimensions are even more significant in the wake of substantial pressure from the development and transportation industry to allow on the highways pavements (Gungor et al 2016; Nega and Nikraz 2017) The effect of tire-pavement contact pressure distribution on flexible pavement is generally complex and dynamic, and it is affected by axles, tires, geometry configurations, and vehicles types including load – carrying capacity (Dessouky et al 2014; Nega and Nikraz 2017) There are several inconsistences in data analysis from several experimental studies to understand and evaluate the influence of tire-pavement contact area and pressure distribution between the tire and pavements including setting the acceptable standard for thin and thick pavements The main purpose of this study is to evaluate the influence of tire print contact area and pressure distribution on flexible pavements Tire-pavement contact stress interaction model was simulated using 3-D finite element for five layers An axisymmetric and tire-pavement 3D finite element model was developed, and deflection for thin and thick pavements was determined using the statics analysis TIRE-PAVEMENT FE MODEL DESCRIPTION The tire-pavement pressure distribution is generally known to be complex and affected by tire type There are many inconsistencies in the data from various experimental studies measuring the distribution of contact pressure between tire and pavement Simplifying assumptions have been used in literatures, including the use of a circular contact with contact pressure equal to the tire pressure (Nega 2016; Nega and Nikraz 2017) The flexible pavement was composed of five layers: Asphalt concrete (AC), unbound base and subbase, compacted and natural subgrade Each layer’s thickness changes depending on the type of flexible pavement considered In the case of a thick pavement, the thickness of the AC and unbound base were 50 and 100 mm, respectively The thickness of unbound subbase, compacted and natural subgrade were 250, 75 mm and infinite, respectively Regarding the material properties, the AC layer was assumed linear elastic with varying modulus between 1.2 and 1.5 GPa Assuming AC as linear elastic material instead of viscoelastic is not expected to have a negative consequence on the conclusion of this particular study (i.e second step for validation) because the main objective is not the study of flexible pavement behavior, but analyzing the impact of tire-inflation pressure on the flexible pavement Unbound subbase, compacted and natural subgrade were determined by the Mohr Coulomb and Drucker-Prager model (Nega et al 2015) because the illustration of the typical cross section of the five layer linear elastic of the flexible pavement model is considered a viscoelastic on the pavement area, which means the unbound base was considered nonlinear for the thin pavement and linear elastic for a thick pavement (Hernandez and Al-Qadi 2016; Nega 2016) The model was verified using Falling Weight Deflectometer (FWD) data from seven main roads and creep test to assure the integrity of the in site experimental data were used in the verification of the finite element (FE) model in ABAQUS The detail experiments can be found (Mulungye et al 2007; Nega et al 2015; Nega et al 2016; Owende et al 2001) For a thin pavement, the stress level in base layer is significant (Hernandez and Al-Qadi 2016), so the stress-dependency of the resilient modulus becomes significant However, in the case of thick © ASCE Airfield and Highway Pavements 2021 323 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved IRI DATA COLLECTION METHODOLOGY Figures 4(a) and (b) show the cell phones set up in the Toyota Prius sedan car Two students collected the IRI data; one student drove the vehicle, and another student operated the Roadroid app and the cell phones The students collected the data during the off-peak time to reduce interference with local traffic The off-peak time was chosen since most of the roads surveyed for this study were two lanes (i.e., one lane, each direction), and the data collection speed was relatively low compared to the posted speed The average lowest speed recorded by the Roadroid app was 8.0 mph., and the average highest speed was 30.0 mph (a) (b) Figures 4(a) and (b) IRI data collection using Motorola G6 (left) and Samsung J3 (right) cell phones Four sets of data were collected; two cell phones and two data were collected for each direction, i.e., North to South or East to West, and returned directions For example, the car drove Northbound, two cell phones collected one set of data, and while the vehicle returned following Southbound, two cell phones collected another set of data The car again drove to North-bound and collected data, and returned to South-bound and collected data The North-bound and South-bound data are then averaged, and one IRI value was calculated for each road and recorded alongside the PCI data collected by another group of students The IRI data collector group coordinated with the PCI data collector group before visiting each site and confirmed their starting and ending data collection location The cell phone data variability can be seen in Figures 4(a) and (b) According to Figure 4(a), the green horizontal bars on both cellphone screen shows that they are collecting consistent IRI data As shown in Figure 4(b), the green and yellow horizontal bars on the Motorola phone and green horizontal bars on the Samsung phone show that they are collecting slightly different IRI data The difference may be due to sensor sensitiveness in two cell phones, or the vehicle vibration is not equally distributed into the vehicle frame For this reason, averaging four sets of data would normalize the variation into the IRI data © ASCE Airfield and Highway Pavements 2021 324 IRI DATA VISUALIZATION Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved The IRI data needs to upload to the Roadroid website for more visualization Figure shows an example of one of the types of data visualization for N Big Hollow Rd The Figure shows locations where the roads are in good, satisfactory, unsatisfactory, and poor conditions and the location of various ranges of IRI values The location interval can be varied, such as the pavement condition can be visualized every in 100 m or 200 m intervals The IRI data is available in the SI unit on the Roadroid website Figure Various road conditions and IRI data visualizations on the Roadroid website IRI (in./mile), Motorola G6 1000 y = 0.9795x + 52.769 R² = 0.8087 800 600 400 200 0 200 400 600 IRI (in./mile), Samsung J3 800 1000 Figure Comparison between IRI values measures by the Roadroid app using Samsung J3 and Motorola G6 cell phones © ASCE Airfield and Highway Pavements 2021 325 Figure shows the comparison between IRI values measured by the Roadroid app using Samsung J3 and Motorola G6 cell phones The lower IRI value means less distressed pavement According to the graph, the R-square value is 0.8087, which means a strong correlation was observed between the IRI values measured using the two cell phones The blue dotted line shows the linear regression line, and the red line shows the line of equity It could be seen that both cell phones show fair agreements as most of the pavements have low to moderate distressed conditions, such as IRI value ranges from 150 to 400 in./mile Even a few pavements have a high distressed condition, such as an IRI value greater than 400 in./mile; the IRI values are close to the line of equity The comparison indicates that the Roadroid app performed well to measure all levels of distressed pavements for most of the pavements surveyed in this study Figures and show the relationship between PCI and IRI data The IRI data were measured by the Roadroid app using Samsung J3 and Motorola G6 cell phones Higher PCI value and lower IRI value mean less distressed pavement and vice versa According to the graph, the R-square values are 0.0248 and 0.0577, respectively, which means the linear regressions not show a strong correlation between PCI and IRI values Thought other research shows a better correlation between IRI and PCI data (Arhin et al 2015, Hasibuan and Surbakti 2019) The blue dotted line shows the linear regression line, and the red line shows the line of equity The data points close to the line of equity show the best correlation between the PCI and IRI data However, there are lots of data that are scattered and away from the line of equity Though the Motorola G6 cell phone collected IRI data shows a better correlation than Samsung J3, the R-square value is not significantly different 1000 y = -1.4648x + 358.35 R² = 0.0248 IRI (in./mile), Samsung J3 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved RESULTS 800 600 400 200 0 10 20 30 40 50 PCI 60 70 80 90 100 Figure Comparison of PCI vs IRI data; the IRI data were measured by the Roadroid app using the Samsung J3 cell phone © ASCE Airfield and Highway Pavements 2021 326 1000 IRI (in./mile), Motolora G6 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved y = -2.4358x + 469.72 R² = 0.0577 800 600 400 200 0 10 20 30 40 50 PCI 60 70 80 90 100 Figure Comparison of PCI vs IRI data, the IRI data were measured by the Roadroid app using the Motorola G6 cell phone CONCLUSIONS The study is done as a pilot project for Peoria County to collect IRI data using cell phone app Roadroid, alongside PCI data for selected roads in the county IRI is well-established indexing to measure pavement conditions as oppose to PCI PCI data collection is time-consuming and laborintensive, and often time lane closure is needed IDOT collects IRI data using vehicle-mounted or hand-held laser devices that are expensive, sophisticated, and skilled technicians required For this reason, a less expensive but reliable method, such as a cell phone app, was used in this project Cell phones can mount in a car or truck, and the vehicle can drive at the near posted speed, and the data collection is faster Two cell phones were used to check the consistency of the IRI data There are various cell phone apps available in the app store to download However, currently, only Android operating system-based apps are available for download A well-established app Roadroid is used in this project In this project, the Roadroid app shows good agreement while collecting IRI data using Samsung J3 and Motorola G6 phones The R-square value is 0.8087 However, it is seen that the cell phone app measured IRI, and PCI data not show a strong correlation It is essential to check the cell phone data collection's consistency using the Roadroid app for future years It is expected that the county roads will show more distresses in the coming years due to traffic and climate, and IRI value will increase along with that The Roadroid app with cell phones should capture the increase in IRI value in future years ACKNOWLEDGMENTS The author is grateful to Peoria County for providing financial support to conduct the study The author appreciates Roadroid for giving access to the app for collecting the IRI data Special thanks to the BU students Mohammed Abdul Aslam, Rizwan Ahmed, Ahmed Alkhatim Idris, Yohannes Fisseha, and Joshua Rodrigues for their dedicated work on the project © ASCE Airfield and Highway Pavements 2021 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved REFERENCES Aleadelat, W., Ksaibati, K., Wright, C H G., and Saha, P (2018) “Evaluation of Pavement Roughness Using an Android-Based Smartphone.” Journal of Transportation Engineering, Part B: Pavement, 144(3), 1–9 Arhin, S A., Williams, L N., Ribbiso, A., and Anderson, M F (2015) “Predicting Pavement Condition Index Using International Roughness Index in a Dense Urban Area.” Journal of Civil Engineering Research, 5(1), 10–17 Asphalt-Surfaced Roads and Parking Lots: PAVERTM Distress Identification Manual, U.S Army Corps of Engineers, ERDC- CERL, June 2009 ASTM D6433-11 Standard Practice for Roads and Parking Lots Pavement Condition Index Surveys, ASTM International, West Conshohocken, PA, 2011, www.astm.org Belzowski, B., and Ekstrom, A (2015) Evaluating Roadway Surface Rating Technologies MDOT Report No RC-1621 MTRI-2015-19 Lansing, Michigan, USA Concrete Surfaced Roads and Parking Lots: PAVERTM Distress Identification Manual, U.S Army Corps of Engineers, ERDC- CERL, June 2009 Douangphachanh, V., and Oneyama, H (2014) “A Model for the Estimation of Road Roughness Condition from Sensor Data Collected by Android Smartphones.” Journal of Japan Society of Civil Engineers, Ser D3 (Infrastructure Planning and Management), 70(5), 103–11 Hasibuan, R P., and Subakti, M S (2019) “Study of Pavement Condition (PCI) Relationship with International Roughness Index (IRI) on Flexible Pavement.” MATEC Web of Conferences, SCESCM 2018, Vol 258 Hossain, M I., and Tutumluer, E (2019) “Methodology for Evaluation of Seal-Coated, Gravel, and Dirt Roads.” Research Report No FHWA-ICT-19-008, Illinois Center for Transportation, Illinois Islam, S., Buttlar, W., Aldunate, R., and Vavrik, W (2014) “Measurement of Pavement Roughness Using Android-Based Smartphone Application.” Transportation Research Record: Journal of the Transportation Research Board, 2457, 30–38 Park, K., Thomas, N E., and Wayne Lee, K (2007) “Applicability of the International Roughness Index as a Predictor of Asphalt Pavement Condition.” Journal of Transportation Engineering, 133(12), 706–709 Pavement Technology Advisory—Data Collection Vehicles—PTA-T2 (2005) Illinois Department of Transportation Bureau of Materials and Physical Research, Illinois Department of Transportation, Springfield, Illinois Sayers, M W., Gillespie, T D., and Paterson, W D O (1986) “Guidelines for Conducting and Calibrating Road Roughness Measurements.” World Bank Technical Paper Number 46 The World Bank, USA Sayers, M W., and Karamihas, S M (1998) The Little Book of Profiling: Basic Information about Measuring and Interpreting Road Profiles University of Michigan, Ann Arbor, Michigan Schlotjes, M R., Visser, A., and Bennett, C (2014) “Evaluation of a Smartphone Roughness Meter.” In Proceedings of the 33rd Southern African Transport Conference, Pretoria, South Africa, 141–53 Technical Manual: Pavement Maintenance Management U.S Army Corps of Engineers Washington, DC: Headquarters, Department of the Army, 1982 © ASCE 327 Airfield and Highway Pavements 2021 Machine Learning Approach to Identifying Key Environmental Factors for Airfield Asphalt Pavement Performance A Z Ashtiani, Ph.D., P.E.1; S Murrell, P.E., M.ASCE2; and D R Brill, Ph.D., P.E.3 Applied Research Associates, Inc., Elkridge, MD Email: aashtiani@ara.com Applied Research Associates, Inc., Egg Harbor Township, NJ Email: smurrell@ara.com Airport Pavement Technology, Federal Aviation Administration, William J Hughes Technical Center, Atlantic City International Airport, Egg Harbor Township, NJ Email: david.brill@faa.gov Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved ABSTRACT Machine learning (ML) techniques are promising methods for developing predictive models involving multiple interrelated predictors A key step in an ML procedure is feature engineering, which is a method for converting raw data into sets of useful and relevant features that provide the best performance model prediction In this study, the Federal Aviation Administration (FAA) applied feature engineering to the problem of identifying the key environmental variables (climate and weather) that influence airfield asphalt pavement performance The FAA implemented various feature selection and feature construction methods based on supervised and unsupervised learning algorithms Selected environmental variables will become inputs to the machine learning models being developed to predict long-term pavement performance Data from the FAA extended airport pavement life (EAPL) program were used in this study The EAPL database includes various pavement performance measures, such as PCI and derivative indexes, surface friction and profile roughness indices, as well as maintenance work histories, historical runway usage, and historical weather data for runways at large- and medium-hub U.S airports In this study, the effect of certain environmental variables was evaluated with respect to the performance index anti-SCI, a derivative of the PCI containing only those distresses that are not directly caused by aircraft loads INTRODUCTION The goal of pavement performance models based on data-driven approaches is to identify and characterize relationships between a series of interrelated features (also known as predictors) in the data and a target which is a performance measure The predictors in these models include critical parameters influencing the pavement performance such as weather, traffic, material properties and pavement structure Methods based on machine learning techniques can be effectively used in problems involving a contribution of multiple predictors Pavement performance models are typically underspecified, meaning that the predictors not fully represent the underlying factors inducing pavement deterioration One reason is that the complete set of predictors is not often available because acquiring them is difficult and expensive The other reason is the inherent uncertainty in the quality of predictors due to such factors as the variability in construction materials and variabilities in environmental and traffic information The prediction models can also be over-specified when a large set of predictors are considered with the hope that they capture some significant part of variations in pavement performance Underspecified models tend to have low predictive accuracy and robustness In over-specified models, especially when the set of incorporated features is large in comparison with the number © ASCE 328 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 329 of samples, the ML learning process is often more complicated, and the chance of over-fitting is high The presence of a large number of insignificant predictors tends to increase the prediction variance, making the predictive model unreliable In addition, an increased number of features increases the difficulty in exploring the effect of each feature on the prediction FAA has initiated the development of pavement performance models based on data collected under the EAPL study (Brill and Parsons, 2017) The FAA has gathered a variety of airfield pavement data and stored them in a dedicated EAPL database, designated PA40 The PA40 database contains various pavement performance measures such as surface friction, profile roughness and surface distresses The database also contains maintenance work histories, historical pavement condition index (PCI) data as well as historical runway usage and weather data A preliminary study by the FAA (Ashtiani et al 2019) used regression analysis to develop predictive models for various pavement performance indices In these models, pavement age was considered the only predictor, assuming the age is a surrogate for traffic and weather events A followed-on study by the FAA considered expanding the performance models beyond pavement age by taking into account additional predictors such as environment and traffic variables This paper focuses on identifying the key environmental variables in the PA40 database that impact the pavement performance by implementing various feature selection methods based on supervised learning algorithms FAA PA40 DATABASE Runway Pavement Performance Indices The FAA PA40 database contains various performance-related indices such as PCI and its load-related component, referred to as SCI, and complementary non-load related component, referred to as anti-SCI Anti-SCI is scaled from zero to 100, and for flexible pavements is calculated using all distresses defined in ASTM D5340, except for alligator cracking and rutting Since anti-SCI contains distresses that are not directly caused by aircraft loads, material aging and weather events are the most probable causes of its deterioration Data from 10 runways with flexible pavements were used in this study as presented in Table Each runway is typically divided into sections Multiple performance data, collected over the years, are typically available for each pavement section Table 1: Airport Runways Studied Airport Boston Logan Airport (BOS) Columbus International Airport (CMH) Greensboro International Airport (GSO) Kansas City International Airport (MCI) LaGuardia Airport (LGA) Miami International Airport (MIA) San Francisco International Airport (SFO) Tucson International Airport (TUS) Runway 4L-22R 10L-28R & 10R-28L 5L-23R 9-27 4-22 12-30 10R-28L 11L-29R & 3-21 Environmental Data The weather records for each airport were obtained from the National Oceanic and Atmospheric Administration (NOAA) and added to the PA40 database, and a system to summarize weather events experienced by a pavement section between construction and each © ASCE Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 330 inspection was implemented into PA40 Weather-related variables in the PA40 database are typically represented as the summation or average of principal variables: (1) temperature, (2) precipitation, (3) dew point, (4) solar radiation, (5) sky cover, and (6) wind speed Climate- or weather-related variables can be characterized by both cumulative and average values between the last major construction or rehabilitation date and the date of a given inspection Thirteen (13) weather variables that may influence the pavement performance were initially identified for investigation (Table 2) Weather variables exhibited temporal behavior, i.e., their values were not constant over the years the performance data were collected This indicates that in some years the pavements were exposed to more severe weather conditions than other years, which could accelerate pavement deterioration Table 2: Weather variables considered in this study Climate Variables Freezing Degree Days (FDD) Freeze Thaw Cycles (FThC) Days Temp Over 90˚F (Temp90) Days Precipitation (DPrec) Total Precipitation (TPrec) Freeze Precipitation Days (FPD) Hydration Days (HD) Avg Daily Temperature (Avg Temp) Avg Daily Temperature Difference (Temp Diff) RHumidity Avg Avg Wind Speed Thornthwaite Index Sky Cover Units o F cycles days days inches days days o F o F % mph % oktas CORRELATIONS AMONG CLIMATE VARIABLES A common problem in ML model development is predictors that are highly correlated (Tsagris et al, 2018) This phenomenon is called multicollinearity or collinearity The collinear variables are often different manifestations of the same process Collinearity occurs when the whole set of information is described by relative quantities For example, FDD and Temp90 are two extreme representations of temperature, which in some regions could have negative correlations Collinearity may also be caused due to inherent limitations in data, for example when the sample size is small, and predictors are not identically distributed in the input domain Perfectly correlated predictors may lead to a misspecified model In this case one or more of the perfectly correlated predictors can be omitted because no additional information is gained by including them (Guyon and Elisseeff, 2003 and Dormannet al., 2013) High collinearity may negatively affect the prediction performance of the model However, the major problem with high collinearity is that it makes it difficult for the predictive model to assess the relative importance of the predictors with respect to the target To this end, the temporal dependency among weather variables based on yearly fluctuations was examined using climate © ASCE Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 data spanning from 1989 to 2016 from airports identified above Various collinearity tests were performed to explore the correlation between the weather variables One collinearity test is to create Pearson correlation matrices to examine the strength of pairwise linear relationships Correlations exceeding a threshold of 0.7 were considered “high collinearity” Based on this threshold, “Days Precipitation,” “Total Precipitation” and “Thornthwaite Index” are highly positively correlated to one another Also, “Freeze-Thaw Cycles” is highly negatively correlated to “Average Daily Temperature” and positively to “Freezing Degree Days.” The airports in the PA40 database are not distributed randomly across the states Since the locations of asphalt surfaced airports not cover all the geographic and climatic regions, a complete range of potential climate situations may not exist Therefore, some correlations should not be eliminated without using engineering judgment Collinearity patterns are likely to change from one data set to another dataset For example, precipitation and moisture could be highly correlated in some climate regions, but not in other regions Also, the correlation may change over the decades IMPLEMENTATION OF DIMENSIONALITY REDUCTION METHODS ON WEATHER VARIABLES Given that flexible pavement performance data are available currently for only 10 runways at airports, there will be more independent variables than the number of climate locations This may pose challenges in developing reasonable predictive models Therefore, it is desired to select or construct a subset of climate features that are useful to build a good prediction model Feature selection is the process in ML that reduces the dimensionality of data by identifying and removing a subset of irrelevant or redundant variables that can decrease the model accuracy and quality Feature Selection Methods There are two general feature selection approaches: feature ranking and subset selection Feature ranking evaluates relevant variables based on their individual predictive power by assigning weights according to their degrees of relevance Subset selection produces candidate feature subsets that are useful as predictors based on a certain search strategy Ranking algorithms can be inefficient, particularly if some of the features are redundant On the other hand, subset selection methods may exclude some redundant but relevant variables (Guyon and Elisseeff, 2003) Two categories of feature selection methods are: Filter and wrapper Filter methods select a subset of features based on their importance according to their variance and relevance to the response Selecting important features is part of the preprocessing step and there is no learning algorithm involved in filter approaches Wrapper methods try to use a learning algorithm to evaluate the relative benefit of adding or removing a feature from the feature subsets The evaluation is performed based on a selection criteria by directly measuring the variation in model performance stemming from addition or removal of a feature The algorithm is said to have achieved a good performance when its stopping criteria are met (Kotsiantis et al., 2006) The feature selection methods evaluated in this study employ a supervised learning scheme, i.e., a set of predictors is assessed against a target (in this case, anti-SCI) The following Filter methods were used: 1) Pearson Correlation Coefficient (R), which is a fast way for screening the relevance of an individual feature by measuring its correlation with the target Pearson correlation is a univariate feature ranking method that can only detect linear dependencies between the variable and target; 2) Correlation-based Feature Selection (CFS), which is a subset selection multivariate filter algorithm This algorithm evaluates the significance of a subset of features by examining the individual predictive ability of each feature alongside the degree of redundancy or © ASCE 331 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 correlation between them and scores them accordingly (Hall, 1999) Irrelevant features are disregarded due to their low correlation with the target Redundant features are removed due to their high correlation with one or more of the remaining features; and 3) RReliefF, is a modified version of the original Relief algorithm (Kira and Randell, 1992) The original Relief, which was proposed for classification problems, calculates a proxy statistic (weight or score) for each feature to estimate feature relevance to the target (Urbanowicz et al 2018) Relief randomly picks a sample of instances from the data and for each instance it finds the nearest neighbor from the same and opposite class based on the Euclidean distance measure The underlying assumption is that a useful feature should differentiate between data objects from different classes (Bolón-Canedo et al 2012) ReliefF is an extension of Relief for classification and has the ability to deal with multiclass problems ReliefF finds the k nearest neighbor from each other class and averages the weight update based on the prior probability of each class RReliefF for regression is similar to RelifF and tries to penalize the predictors that give different values to neighbors with the same response values, and rewards predictors that give different values to neighbors with different response values The RReliefF algorithm does not assume the conditional independence of the features and can correctly estimate the quality of features in problems with strong dependencies between features (Robnik-Šikonja et al., 2003) Wrapper methods were also implemented using two learners: linear regression and support vector machine (SVM) Cross-fold validation was used to estimate the benefits of adding or removing a feature from the feature subset and to evaluate the model accuracy A forward stepwise selection approach was used in which the algorithm starts with an empty set of features and searches forward by adding features until the performance does not improve further A score or weight are given to each feature based on their relevance to the target or based on their prediction power Each method has its unique algorithm for assigning scores to the features Variables that most improves the model (i.e having higher score) are selected A Gaussian kernel function was used for the SVM learning Feature Selection Implementation on PA40 Data Two approaches were used for the implementation of feature selection on weather variables In the first approach, climate variables were treated as sums, i.e., cumulative values from the date of construction/rehabilitation to the date of inspection For each pavement performance record in the dataset, six cumulative weather variables were considered as predictors and the measured anti-SCI was considered as the target Pavement age at the time of anti-SCI measurement and the previous anti-SCI measurement were also considered as predictors In the second approach, weather variables were treated as averages, i.e., average values of weather variables over a given period of time (between the last rehabilitation and the inspection date, or between two inspections) Eleven weather variables were considered as the predictors, and the rate of anti-SCI deterioration (change in anti-SCI divided by the time between rehabilitation and inspection, or between two inspections) was considered as the target Since the rate of anti-SCI deterioration is not constant, particularly for ages over 10 years, only performance data for pavements less than 10 years old were used in the second approach This was to isolate, at least in part, the influence of the pavement age on the deterioration rate when performing the feature selection analysis Table lists the input and output variables in both approaches Feature Selection Results The “Attribute Selection” module within the WEKA software program was used to perform the feature selection analysis Weka is a free and open-source software program for data mining and machine learning (Witten, 2002) For Approach 1, Table summarizes the rank of each feature, and the associated weights or scores as calculated by each © ASCE 332 Airfield and Highway Pavements 2021 333 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved ranking algorithm According to these ranking algorithm, pavement age and previous anti-SCI are the most informative, and freezing degree days and freeze-thaw cycles the least informative variables Table summarizes the subset of informative features as determined by three subset selection algorithms These algorithms search for dependency between the features and remove redundant or correlated features that have minimal impact on the prediction performance Pavement age, previous anti-SCI, and number of days with a high temperature above 90ºF are all identified as informative variables in all the methods The wrapper methods retained freezing degree days, freeze-thaw cycles and total precipitation as relevant features Table 3: Input and Output variables used in Feature Selection Analysis Approach Approach 1: Cumulative Weather Variables Approach 2: Average Weather Variables Input Variables FThC, FDD, Temp90, DPrec, TPrec, HD, Age, Previous Anti-SCI FThC, FDD, Temp90, DPrec RHumidity, HD, Temp Diff TPrec, Sky Cover, Avg Wind Speed Thornthwaite Index Target anti-SCI Rate of anti-SCI change Table 4: Feature Selection using Ranking Algorithms in Approach Filter Methods Pearson RReliefF Correlation Input Input R Score Variable Variable Previous Previous 0.77 0.040 anti-SCI anti-SCI Age -0.77 Temp90 0.022 DPrec Temp90 TPrec HD FThC FDD -0.63 -0.59 -0.57 -0.57 -0.54 -0.50 Age TPrec HD DPrec FDD FThC 0.016 0.010 0.009 0.009 0.006 0.005 Wrapper Methods SVM - Gaussian Liner Kernel Regression Input Input Score Score Variable Variable Previous Age 12.3 5.7 anti-SCI Previous 10.5 Age 5.6 anti-SCI Temp90 9.3 DPrec 3.5 DPrec 9.1 Temp90 2.9 HD 8.0 TPrec 2.8 TPrec 7.6 HD 2.7 FDD 7.1 FThC 2.4 FThC 6.7 FDD 2.0 Table 5: Feature Selection using Subset Selection Algorithms in Approach Filter Method CFS Age (year), Previous Anti-SCI, Temp90 © ASCE Wrapper Methods SVM - Gaussian Kernel Liner Regression Age, Previous anti-SCI, FThC, FDD, Temp90, TPrec, HD Age Previous Anti-SCI FThC, FDD, Temp90, DPrec Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 334 For Approach 2, Table summarizes the rank and score of each weather variable This table shows that both wrapper methods and the Spearman correlation gave the highest rankings to the average daily temperature difference and precipitation variables, and lower rankings to freezing degree days and freeze-thaw cycles The correlation coefficients (R) are relatively low, and for some of the variables the sign of the correlation contradicts engineering expectations Moreover, the ranking scores are roughly the same for all the climate variables This suggests that the rate of anti-SCI reduction is not significantly correlated to any of the climate variables In other words, no individual climate variable can describe the changes in anti-SCI deterioration In addition, the feature rankings obtained by filter and wrapper methods are inconsistent One reason for this inconsistency is the different criteria used by the algorithm for selection of features The inconsistency may also be due to the insignificant correlation between the features and the target Table summarizes the subset of relevant features as identified by filter and wrapper methods The CFS method identified the Average Daily Temperature Difference as the only relevant feature The wrapper methods identified Freezing Degree Days and number of days of precipitation as relevant features in addition to Average Daily Temperature Difference Table 6: Feature Selection using Ranking Algorithms in Approach Filter Methods Pearson Correlation ReliefF Input Input R Score Variable Variable Temp Diff 0.28 Sky Cover 0.023 DPrec -0.21 Avg Temp 0.019 TPrec -0.20 RHumidity 0.013 Temp90 0.16 Wind Speed 0.013 RHumidity -0.15 FThC 0.011 FDD 0.11 TPrec 0.010 Wind Speed -0.09 Thornthwaite 0.009 HD -0.09 Temp Diff 0.009 Thornthwaite -0.07 DPrec 0.01 FThC 0.07 Temp90 0.01 Avg Temp 0.05 FDD 0.01 Sky Cover -0.05 HD 0.00 Wrapper Methods SVM - Gaussian Kernel Liner Regression Input Input Score Score Variable Variable Temp Diff 3.1 Temp Diff 0.1 TPrec 3.1 TPrec 0.0 DPrec 3.1 DPrec 0.0 Temp90 3.1 Temp90 0.0 RHumidity 3.1 Sky Cover 0.0 HD 3.1 Thornthwaite 0.0 Thornthwaite 3.1 Avg Temp 0.0 Sky Cover 3.1 Wind Speed 0.0 FDD 3.07 RHumidity -0.01 Wind Speed 3.06 HD -0.02 Avg Temp 3.04 FDD -0.02 FThC 3.04 FThC -0.02 Table 7: Feature Selection using Subset Selection Algorithms in Approach Filter Method CFS Temp Diff © ASCE Wrapper Methods SVM - Gaussian Kernel Liner Regression Temp Diff, FDD, DPrec Temp Diff, FDD, DPrec RHumidity Avg, HD, Temp90, TPrec, Thornthwaite 335 Table showed that the Pearson correlation coefficient between each cumulative climate variable and anti-SCI was less than the correlation between age alone and anti-SCI This means that none of the climate variables alone is a better predictor than age Since the cumulative climate variables in approach are dependent on pavement age, any high correlation with anti-SCI will not imply a cause of performance deterioration On the other hand, a low correlation between a climate variable and performance does not imply that the variable is completely useless For example, the cumulative number of freezing degree days had the lowest correlation with anti-SCI, which goes against the common expectation The reason for this low correlation is that FDD is a weather variable representing an extreme weather condition, making it a significant factor in the cold weather region but an insignificant factor in the warmer region Figure illustrates the scatter plot of anti-SCI versus the number of cumulative FDD for each pavement section in the database It is obvious from Figure that there are two distinct clusters of data One cluster includes the data from three airports in no-freeze regions (SFO, TUS and MIA) and the other cluster represents the rest of the airports which are located in the northern and central part of the USA The FDD has a very low correlation with anti-SCI in those airports with nearly zero FDD, and a relatively high correlation with anti-SCI for the rest of the airports with a higher chance of freezing events Evaluation of the influence of weather inputs on pavement performance would be more meaningful if individual distress types were considered, as opposed to an overall surface condition indicator such as anti-SCI The specific types of pavement distresses contributing to anti-SCI deterioration can differ greatly from one climate region to another For example, while thermal cracking due to severe freezing is a common type of distress in a cold region, block cracking and rutting could be primary distresses in an airport with warm weather and high solar radiation Therefore, it is anticipated that a climate variable such as FDD will have a high correlation with particular distress types such as thermal cracking In this case, FDD can be described as the primary factor causing thermal cracking 100 90 80 70 anti-SCI Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 60 50 R² = 0.2069 40 R² = 0.5995 30 Freeze Regions 20 No-Freeze Regions 10 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 9,000 Freeze Degree Days Figure 1: Correlation of anti-SCI with Cumulative Freeze Degree Days © ASCE Airfield and Highway Pavements 2021 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved SUMMARY AND CONCLUSION This paper presents results of feature selection analysis to evaluate the influence of weather parameters on pavement performance Feature selection is known as a key step in developing predictive models based on machine learning techniques Several feature ranking and subset selection algorithms were implemented using data from the FAA PA40 database These algorithms, which are based on supervised learning methods, analyzed relationships between interrelated predictors such as pavement age and weather variables and a target which was antiSCI performance index Results showed that ranking algorithms that score variables individually and independently of each other are not able to determine which combination of variables would give the best model performance Ranking algorithms can give some information on how relevant each variable is to performance, but not enough to determine which should be kept or discarded for the prediction of performance It can be concluded from the results of filter methods in approach that a higher rank given to a weather variable does not suggest the significance of that variable in anti-SCI deterioration The results of implementation of ranking algorithms on the average climate variables in approach showed that the variation in the rate of anti-SCI deterioration cannot be related to any individual climate variable However, it is likely that when all the climate variables are taken together, they can describe some variation in anti-SCI Subset selection methods seemed to be more promising in identifying the most relevant weather variables to the pavement performance ACKNOWLEDGMENTS The work described in this paper was supported by the FAA Airport Technology R&D Branch, Dr Michel J Hovan, Manager The contributions of Mr Timothy Parsons of ARA are gratefully acknowledged by the authors The contents of the paper reflect the views of the authors, who are responsible for the facts and accuracy of the data presented within The contents not necessarily reflect the official views and policies of the FAA The paper does not constitute a standard, specification, or regulation REFERENCES Ashtiani, A Z., Shirazi, H., Murrell, S., Speir, R., and Brill, D R., (2019) “Performance Model Development for Extended Airport Pavement Life” International Airfield and Highway Pavements Conference, Chicago, IL, USA Bolón-Canedo, V., Sánchez-Maroño, N., and Alonso- Betanzos, A (2012) “A review of feature selection methods on synthetic data” Knowledge and Information Systems Brill, D R., and Parsons, T A (2017) “Development of new FAA design procedures for extended airport pavement life.” 10th International Conference on the Bearing Capacity of Roads, Railways and Airfields, CRC Press, Athens, Greece Dormann, C F., et al (2013) “Collinearity: a review of methods to deal with it and a simulation study evaluating their performance” Ecography (36) pp 27-46 FAA (2016) Airport Pavement Design and Evaluation Advisory Circular 150/5320-6F, November 10, 2016 Washington, DC: US Department of Transportation, Federal Aviation Administration, Office of Airport Safety and Standards © ASCE 336 Downloaded from ascelibrary.org by Nanjing University on 06/08/21 Copyright ASCE For personal use only; all rights reserved Airfield and Highway Pavements 2021 Guyon, I., and Elisseeff, A (2003) “An Introduction to Variable and Feature Selection” Journal of Machine Learning Research, 3, pp 1157-1182 Hall, M A (1999) Correlation-based feature selection for machine learning PhD Thesis, University of Waikato, Hamilton Kira, K., and Rendell, L A (1992) “The feature selection problem: traditional methods and anew algorithm”, in: AAAI, 2, pp 129–134 Kotsiantis, S B., Kanellopoulos, D., and Pintelas, P E (2006) “Data Preprocessing for Supervised Leaning” International Journal of Computer Science, Volume 1, Number Robnik-Šikonja, M., and Kononenko, I (2003) “Theoretical and Empirical Analysis of ReliefF and RReliefF” Machine Learning volume 53, pages23–69 Tsagris, M., Lagani, V., and Ioannis, T (2018) “Feature selection for high-dimensional temporal data” BMC Bioinformatics 19, 17 Urbanowicz, R J., Meeker, M., La Cava, W., Olson, R S., and Moore, J H (2018) “Reliefbased feature selection: Introduction and review” Journal of Biomedical Informatics, 85, pp.189-203 Witten, I H., and Frank, E (2002) “Data mining: practical machine learning tools and techniques with Java implementations” Acm Sigmod Record 31 (1), 76-77 © ASCE 337