Evaluating the Use of Spatially Precise Diurnal Population Data in Aviation Noise Studies (2020)

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

3 2 Emerging Practices and Activity-Based Models 2.1 Introduction to Diurnal Population Movement and Spatiotemporal Data In this section we examine the current and emerging methods and practices for using spatiotemporal population data sets and their applicability to aircraft noise modeling. Spatiotemporal data considers both the spatial and temporal aspects of population counts, measuring the number of people present at a given location during a specific time period. Diurnal, in this context, refers to the movement of people over the course of the course of a 24-hour day. This includes movement from home to work, school, shopping, entertainment, and the like. Diurnal movements can be modeled in spatiotemporal population data which reflects the population present by time period. A review of research studies, in a variety of domains, applied diurnal population in their analyses demonstrated the potential insights that can be gained from the spatiotemporal movement of population as well as the limits of the available data sources. The most common applications for spatiotemporal population models are presented, along with a summary of articles and research papers from the current literature, including one study that has applied spatiotemporal population to aircraft noise. Special attention is given to sources and methods for generating diurnal population data sets and how that influences their applicability to aircraft noise studies. It is only recently that daytime population distributions have been actively developed. Daytime population attempts to quantify the number of people who are present in an area during typical business hours, including workers, children in school, people in hospitals or other short-term medical facilities, people temporarily staying in lodging facilities, customers present at retail locations, and persons remaining in residences. Bhaduri (2008) defined daytime population as “the distribution of population at times other than when they are expected to be at their residences at night which extends the duration from business hours to include the evening hours as well.” The U.S. Census Bureau first published daytime population estimates using Census 2000 data and the American Community Survey (ACS). The original estimates from Census 2000 were limited to the city or urban area location of workers during a typical workday. Later, the Census Bureau also provided commuter-adjusted population estimates based upon travel-to-work information collected from the ACS. It is based solely upon workers and does not account for any other type of travel during the day. In 2013, the U.S. Census Bureau released commuter-adjusted daytime population estimates based on the 5-year 2006–2010 ACS estimates (Rapino et al., 2014). These estimates took into account the total population and the number of workers commuting into and out of each area daily. The commuter-adjusted daytime population estimates provided an overall estimate of population fluctuation in a given area based on commuting throughout the day. Since that release, the U.S. Census Bureau conducted research on further delineating the commuter-adjusted daytime population estimates. These new dynamic daytime population estimates include an additional dimension—the delineation of the daytime population estimates into 15-minute increments. The key challenge of developing a diurnal population model is gathering enough information to come up with a population count at a sufficient level of spatial granularity. The minimum spatial granularity required will vary by the intended application, but no method of developing diurnal

4 population data approaches the granularity available from census data. In the United States, census blocks represent the smallest spatial unit available, and may be as small as a few houses or a city block. Aviation noise studies use census block data. So, any aviation noise study using diurnal population data will involve estimates at a coarser resolution. (A detailed discussion of diurnal population data sets is provided in section 3.) Spatiotemporal databases of human population distribution are of significant value in disease burden estimation, epidemic modeling, resource allocation, disaster management, accessibility modeling, transport and city planning, poverty mapping, and environmental impact assessment. The process to generate the spatiotemporal data set might vary by application, but one common method is activity-based modeling, which classifies persons according to the activities they are engaging in as they move about, such as commuting to work or school, shopping, or engaging in recreational activities. These models describe their movements over time and can be used to create snapshots of population density at discrete time intervals. Research applying activity-based population models most commonly falls into one of the following categories: • Transportation Planning • Emergency Response • Marketing Studies These are discussed in the succeeding sections. 2.2 Activity-Based Models and Transportation Planning Transportation planning is one of the most developed areas for activity-based modeling. For example, over many years the Atlanta Regional Commission (ARC) has developed an activity-based model (Vovsha et al., 2016) to describe the movements of persons throughout the Atlanta area and the transportation choices they make. The model is then used to reveal demand patterns for all modes of transportation by location and time. It classifies individual persons by age and activity, such as full-time worker, student, or retired person, among other categories. The same model allows researchers to create synthetic populations, that is, a snapshot of the spatial distribution of the population that varies with time, and further allows segmenting the population according to age, income, and other demographic factors. Segmenting the population by activity allows researchers to model the discretion a person has in being at a certain place and time. A person traveling to work would have little discretion compared to a person engaging in a recreational activity or eating out. They would also potentially have less practical discretion in the mode of transportation to get to their workplace reliably. The data sets to support the activity-based model come from surveys in which respondents answer questions about when and where they travel, by what mode, how long they spend at their destinations, and the activity they engage in at the time. The survey responses are then scaled up to the total population using census data. The model also relies on a land use model which is then used to allocate population by activity. The researchers also create synthetic populations for future scenarios. These scenarios include assumptions about development and population growth and then model the likely movements of the population based on that development.

5 The ARC has developed a visualization tool (http://atlregional.github.io/ABMVIZ/ ) to show the movement of the population throughout the region through time. The Atlanta region is divided into more than 5,800 Travel Area Zones (TAZ), and movement is modeled to and from the different zones. Figure 1 shows the number of people currently traveling (“not at home”) at noon, and Figure 2 shows the population not at home at 11:00 pm. By using the slider at the bottom of the visualization tool, the not-at-home population can be seen rising and falling throughout a 24-hour period in 48 half-hour steps. Figure 1: Atlanta Activity-Based Model – People Not at Home, 12:00 PM Figure 2: Atlanta Activity-Based Model – People Not at Home, 11:00 PM Other research has looked at various ways to build urban activity-based population sets. Greger (2015) uses a novel method for estimating building-level temporal population density estimates in urban areas. The researcher reviewed several existing population data sets and methods for deriving population data sets as part of the study and noted their advantages and drawbacks. A key point was that most spatiotemporal population estimates for urban areas were at a resolution that was not fine enough for some use cases. This study used previously published population estimates

6 by differentiating two activities (“home” and “work”) and by operationalizing the temporal dimension of human activities using a very detailed movement profile data set. The approach utilized multiple data sets—population census data, employment data, address point data, and movement data—to provide fine-grained results of the estimated populations within different usage categories for each building in the study area on a given timescale. The researcher also enumerated the different categories encompassed by building occupancy—homes, business and offices, educational buildings (schools, universities, libraries, etc.), retail and service, leisure and hotels, and public institutions. The approach adopted a timescale that is more detailed than the diurnal separation of daytime and nighttime populations and does not rely on a priori assumptions and expert judgment regarding the time profiles of human activities. The model uses a variety of micro-scale data sets to operationalize the categories of human activities and their spatiotemporal representations. Additional studies of activity-based modeling for transportation planning include McKenzie et al. (2010), who derived commuter-adjusted population estimates from three ACS-based population components. The commuter-adjusted population is largely a function of the resident population, and the difference between the two varies little for most places. The employment/residence ratio is defined as the total number of workers working in a specific geographic area divided by the number of workers living in that area. They go on to estimate the differences between static population and commuter-adjusted population due to commuting. They suggest that the ACS-based commuter-adjusted population estimates can be useful to provide a standardized, widely available, and relatively basic measure to serve as a baseline for more detailed analyses. 2.3 Activity-Based Models and Emergency Response Emergency response is another major area that uses activity-based population models. These models seek to inform first responders with not only the absolute size of a population in a given area, such as a downtown office district in the middle of a workday, but also the description of the population by age, activity the population is likely to be engaged in at the time, or other factors. A summary of research into developing spatiotemporal data to assist with emergency response is presented below: • Sleeter and Wood (2006) developed an accurate representation of population distribution along the Oregon Coast, estimating daytime versus nighttime fluctuations, to be used in the event of a tsunami. By using a dasymetric mapping technique to manipulate U.S. Census enumeration units, population values were distributed to 10-meter (m) grid cells with the input of a tax parcel data set acting as the primary ancillary indicator of inhabited land use/land cover. An area interpolation technique was then applied to disaggregate census population data into homogenous residential or commercial zones reflecting inhabited land use. A dasymetric approach introduced ancillary information to redistribute the standardized data into zones of increased homogeneity taking into consideration actual changing densities within the boundaries of the enumeration district. They used the census-block group as the source polygons of known values and a tax parcel database as the target polygons. The parcel database was divided into four density classes, high, medium, low, and exclusion, to act as a hierarchical measure within the interpolation process that used employment data from the InfoUSA database from 2005. They concluded

7 that the net daytime populations of the study areas decreased significantly when compared to the residential population but did not account for non-residents such as tourists. • Freire (2007) used a data-driven model developed to map the spatial and temporal distribution of the population in two municipalities in Portugal. An approach based on dasymetric mapping was used to combine existing physiographic and statistical data sets to map daytime and nighttime population densities in 2001 for each 25-m grid cell in the study area. • Freire (2010) notes that many methods require assumptions that over-simplify the reality or disaggregate population totals based on heuristic or empirical parameters. They also note that sources such as LandScan have a spatial resolution that is too coarse to adequately support analysis at the local level. In their research they utilize a dasymetric mapping approach to refine population distributions in Portugal. A spatial resolution of 25 m was adopted to approximate the size of a single-family residence (half block). Pre-processing and modeling of geographical data were conducted in ESRI® ArcGIS 9.1, a geographic information system (GIS) application. The workforce population surface was created by georeferencing workplaces and schools and respective workforce and students in the study area. Finally, a map of nighttime population distribution was obtained by using a grid binary dasymetric mapping method to disaggregate residential population from census zones to residential streets on a cell-by-cell basis. Their final results consist of raster surfaces of nighttime (residential) population, daytime residential population, daytime worker and student population, total daytime population, and ambient population. They recommend that future improvements could be made to account for people in transportation networks, hospitals, prisons, leisure locations, and shopping areas, as well as accounting for other variations such as weekend versus weekdays, seasons, and tourism. • Bell’s 2011 Master’s thesis titled “Comparing Methods for Estimation of Daytime Population in Downtown Indianapolis, Indiana” focuses on accurate population estimation for emergency response. Four types of population estimates were evaluated in this study. The two old models were the ESRI 2009 USA daytime population model that includes only the employee count and the 2008 LandScan Global Population Model developed by the ORNL on a grid with a resolution of 30 arc‐seconds. The other newly developed models were the parking lot study method, which was based upon the gross square footage of buildings, and the direct count method that estimated those segments of the population at fixed locations and consisted of employees, day care and school children, prisoners, and the daytime residential population. A simulated gas plume was used to estimate affected daytime populations. The direct count method and ESRI models showed very similar results, while the LandScan and parking lot models had low estimates of population affected. Overall the direct count method is promising for use in emergency response at the local level, as it included more segments of the population than the ESRI model. The disadvantages of the direct count method include the time-consuming process of collecting, evaluating, and geocoding the data, and the proper allocation of the employee population for businesses with multiple buildings. The ESRI Daytime Population data is a readily available source that can be utilized at the Census Block Group level. • Kobayashi et al. (2011) used the U.S. Department of Transportation (DOT) data on the population throughout the day in the Census Transportation Planning Package (CTPP)

8 from 2000. This data set, provided at the Traffic Analysis Zone (TAZ) level, is more detailed than the census tract level. This study relied on a formula provided by the U.S. Census Bureau to calculate present populations based on the number of residents and the number of workers in any region. The Pycnophylactic interpolation method was used for areal interpolation to generate and visualize dynamic urban population surfaces. To create the hourly population surfaces, a TAZ feature class database was populated with data for each of the 24 hours. Next, the feature class was rasterized with the constraint that the cell size must be smaller than the smallest TAZ. Fifty-meter cells were used for this study. This study presents a simple methodology to represent and rapidly visualize diurnal population change for metropolitan areas and offers significant improvement over using static population estimates for emergency management. • McPherson et al. (2003) created a preliminary estimate of the diurnal temporal shift in population due to employment. Information was sourced from the American Business Directory, the Census County‐to‐County Journey to Work data, and the daytime residential population. Separate population grids, at a 250-m resolution, were created for nighttime residential, daytime residential, and daytime workplace population. The residential population location coefficient (pri) was calculated according to the equation pri=1/rbg, where rbg is a number of road grid cells per Census Block Group. Nighttime residential population for the cell (NRPI) is calculated as NRPI=Wbg·pri, where Wbg is the residential population in each block group. Although the methodology used in this technique provided national data sets on daytime and nighttime population, there is a significant source of uncertainty in the source data as well as lack of coverage for educational and retail areas. • Dhondt et al. (2012) used Forecasting Evolutionary Activity-Travel of Households and their Environmental Repercussions (FEATHERS) to derive statistical representative information on the location of populations to assess the impact of air pollution. FEATHERS is an activity-based micro-simulation modeling framework used for transport demand forecasting and implemented to model travel behavior in the regions of Flanders and Brussels (Belgium). At the heart of the activity-based model lies a scheduler, which assumes a sequential decision process consisting of 26 decision trees that simulate the way individuals build activity-travel schedules throughout the day. The scheduler inside FEATHERS is based on the Dutch model ALBATROSS, a multi-agent, rule-based model of activity pattern decisions. FEATHERS is able to predict, for all individuals within a synthetic population, which activities are conducted, when, where, for how long, with whom, and the transport mode involved. To derive the necessary “training data” for the scheduler, activity-based schedule information was gathered from 8800 persons living in the study area, by means of an activity diary survey. Hourly pollutant concentration maps, provided by the air pollution models, were used for calculating both the static and the dynamic exposure approach. For the static exposure approach, the concentration estimates at the receptor points were first transformed into Thiessen polygons using ESRI’s ArcGIS 10. Address locations were then used to weight the different polygons in each zone in order to provide hourly average concentration levels per zone. These weighted concentrations per zone were used as the static approach to assess population exposure. For the dynamic exposure approach, zonal exposure was taken into account. For the in-transport exposure, a different approach was adopted, as FEATHERS was not capable of exactly localizing its individuals when they are in transport. The in-transport exposure

9 was calculated as an hourly average concentration on the road network for the whole study area. The average transport concentration was estimated by selecting the hourly concentrations from the grid points closest to each road segment and averaging those to a concentration value for the whole study area. The authors found a significant difference in the health impact of pollutants by including population mobility compared to the assumption that the people are always at home. • Ahola et al. (2007) studied planning for emergency response in the Finnish capital Helsinki. They used data from the Helsinki Metropolitan Area Council that contained information on building use that could be classified as office, school, etc. They combined this with traffic data, information about school populations, and information about traffic at shopping areas to develop a model of population distribution at different times of day. The research was ultimately to develop a risk assessment model to estimate the number of people exposed to fire risk from a terrorist attack or other emergency. They explored various data structures for their model and various ways to estimate population density and how population distribution changes with time. For this research, the data structures presented did not offer anything new in terms of diurnal population for aircraft noise, compared to existing data sets. • Friere et al. (2015) introduced and tested a new approach for refining spatiotemporal population distribution at high resolution by combining diverse geo-information layers. The model was implemented in GIS using the following data sets as inputs: LandScan population grid, reference densities of worker/student population by land use class (LU weights), Urban Atlas land use map, Global Human Settlement, European Settlement Map grid, official census data, and the ratio (or percent) of residents commuting for work or study. Using this methodology, four raster population distribution surfaces were produced, at 10 m resolution: (1) nighttime (residential) population, (2) daytime residential population, (3) daytime worker and student population, and (4) total daytime population. These grids were aggregated to 50 m cell size for subsequent analysis and visualization. The main contributions of the model were the addition of a temporal dimension to population data and the refinement of the spatial distribution for areas where detailed statistics do not exist or are not available.

10 2.4 Spatiotemporal Data for Use in Marketing Spatiotemporal data can be used to better understand a target market for a product or service, or when and where to reach a target customer. One common way to build such data is with mobile phone location data. Mobile phone companies have historically sold data about customer location to third-party data aggregators. The location data is typically calculated by triangulating the phone’s position when its signal is detected by multiple cell towers. This method is less precise than Global Positioning System (GPS) positioning. Recent concerns over privacy have led mobile phone companies to limit or stop selling customer location data to third-party aggregators.1 More recently, other companies will sell aggregated or raw device location data sold by application developers. Many smartphone and tablet apps collect or use GPS location data which is sent back to the developer. Using apps eliminates the need to go through a mobile phone carrier and can be used to collect GPS location data whether the device is using a Wi-Fi or cellular data connection. Such companies tend to operate in a gray area given sensitivities over privacy, and one of the companies we contacted did not even want their company name quoted in this research. We identified additional studies using mobile phone data to build spatiotemporal models. Two examples are summarized below. • Deville et al. (2014) found that individual mobile phones can be located by identifying the geographic coordinates of the transmitting tower and the associated cell. The mobile phone location-based approach was used to downscale census population data accurately and compare it with the existing method of downscaling through remote sensing (RS) and other geospatial data. Precision and accuracy statistics, including the Pearson product–moment correlation coefficient and root-mean-square error, were calculated to compare the performance of the mobile phone and RS downscaling methods, using the baseline census- derived population densities as a reference. Their approached provided insights into the recreational movement of people on holidays or during weekends and also offered a detailed visualization and quantification of the dynamic popularity of a given place over time. Their research demonstrated that mobile phone data collected every day by the phone network providers can complement traditional census distributions to quantify and visualize temporal movement. • Williams et al. (2015) used mobile phone Call Detail Records (CDR) to estimate spatiotemporal trajectories of individual users. This was done by linking the CDRs associated with that phone with the locations (latitude and longitude) of the cellular towers that handled the calls. Standard measures of mobility derived from CDRs included the number of towers used, distance traveled in a straight line, maximum distance traveled, and the radius of gyration. The researchers analyzed these measures of mobility and found limitations, leading them to develop two additional measures of mobility by gridding subject areas into individual cells. They hypothesize multiple uses for their work that would allow for use of measures of mobility in multiple fields of research. 1 For example, see: https://www.washingtonpost.com/news/the-switch/wp/2018/06/19/verizon-will-suspend-sales-of- customer-location-data-after-a-prison-phone-company-was-caught-misusing-it/ (accessed 10/4/2019)

11 2.5 Spatiotemporal Data and Aircraft Noise Studies Noise impact modeling is typically undertaken with a variety of scientific models that assess the exposure of people on the ground to sources of noise such as vehicles, trains, aircraft, etc. Noise impact modeling typically uses static population distributions obtained from the U.S. Census. However, the census data represents the distribution of residential or nighttime population and does not accurately represent where people are during the day. There are only a few sources of pre-developed population data that provide daytime distributions. These include ORNL’s LandScan data set, the ESRI daytime population data set, and the LEHD LODES data set. Mobile phone data is also a very promising area of spatiotemporal population data, but the format and availability vary widely. Outside of these sources, most researchers seem to develop their own statistical methods for deriving diurnal or spatiotemporal population data. Our literature search yielded only one study where spatiotemporal population data was used specifically to study the effects of aviation noise. Stephen Greaves, Andrew Collins, and Nitin Bhatia in their paper “Disaggregate Assessments of Population Exposure to Aircraft Noise” examined hourly flight movement information and hourly population estimates derived from a household travel survey in Sydney, Australia. The travel survey was based on a set of travel diaries recorded by survey participants that indicated where they went during their day and included trip start and end times. The city was divided into travel zones, and the survey results were used to compute the number of people present in each travel zone polygon by time of day. The data representing a sample of the population was then scaled up to population counts in the 2001 census to produce a total population estimate. The aircraft noise estimates were calculated at grid points in a noise study such as those analyzed in the Integrated Noise Model (INM). The researchers spread the travel zone population evenly among the grid points in each zone. They examined multiple noise exposure scenarios under different levels of flight activity and different runway configurations used at Sydney’s Kingsford Smith International Airport. Researchers found that the dynamic population sets consistently showed a higher number of people exposed to significant aircraft noise compared to the static (residential-based) population estimates. The reason was that more people were present in areas exposed to aircraft noise than in the static data, which the researchers note could be an airport-specific factor. Researchers also noted that noise exposure was highly variable and very sensitive to the runway configuration in use. 2.6 Conclusions Our research to date has not revealed any current or emerging practices that utilize spatiotemporal population data in aviation noise studies. What is evident, however, is that in many planning activities the use of accurate spatiotemporal population data is of increasing importance. There is a growing field of research and application in activity-based modeling that is both a source of and a consumer of spatiotemporal population data to support these planning activities. Most methods and practices begin with a baseline population data set, such as census data, and either derive a new spatiotemporal population data set using additional inputs and methods (e.g., LandScan, ESRI Daytime Population, LODES) or utilize the baseline population data set as an underlying layer in a custom-built population activity model (e.g., ARC model, dasymetric modeling). Depending on the project type and scope, we have found that researchers and planners employ an approach that is tailored to their situation. For example, researchers studying the global distribution of poverty might use a data set such as LandScan whereas a city planner might develop a specialized

12 dasymetric model that utilizes census data. Typically, a localized, activity-based model will require localized knowledge and potentially be dependent on the inclusion of a manual survey of a subset of the population. Aviation noise studies have hitherto relied on static population distributions—namely the census block points or count of residential units. As the metric most commonly used is the DNL, studies conducted using the census data are inherently conservative in their assessment of populations impacted by adverse noise, due to their particular emphasis of nighttime operations over residential areas. Our research shows that there are a variety of data sets and methods that could be used to support diurnal population distributions for aviation noise modeling. Data sets such as LandScan, ESRI Daytime Population, and LODES could be used as-is by applying simple resampling to census block points (LandScan and ESRI) and where necessary applying proportional scaling (e.g., for LODES). Based on the variety of methods and models that have been developed, it is feasible to develop activity-based models for aviation noise modeling that utilize one or more methods from the research, one or more data sets from the research, and localized inputs from the modeler. As the use and distribution of mobile phones and GPS-enabled devices continue to grow, location data obtained from these devices promises to provide the most accurate picture of spatiotemporal population movements. Pioneering use of mobile phone locations triangulated using cellular broadcast towers is being superseded by GPS location data obtained by application developers. At the current moment, coverage from mobile devices is in the 20–25% range, and purchasing such data is expensive. However, indications are that the percentage of the population covered will expand and cost will simultaneously decrease. While these data sets show the number of people present at a given location by time period, they do not indicate the level of sensitivity to aircraft noise. As we note in Section 4.4 Further Research Needs, an activity-based model that combines a model of noise sensitivity with location is necessary to take full advantage of spatiotemporal population data for aviation noise studies.