CURRENT LOCATION-SHARING PRACTICES AND STANDARDS

Archaeologists are responsible to a range of governments and institutions for the ethical collection, maintenance, access, and archiving of the data they produce. For example, in the United States, federally funded grant applications require data management plans. At the same time, the push for open access (OA) to scholarship is also securing a foothold in the field (Costa et al. Reference Costa, Beck, Bevan and Ogden2013; Huggett Reference Huggett, Mills, Pidd and Ward2014; Strupler and Wilkinson Reference Strupler and Wilkinson2017). True OA refers to information that is available without restriction on the internet. Evidence of OA in archaeology comes from the rise of journals such as the Journal of Open Archaeology Data (https://openarchaeologydata.metajnl.com/); repositories such as the Digital Index of North American Archaeology (DINAA; http://ux.opencontext.org/archaeology-site-data/) and the Digital Archaeological Record (https://www.tdar.org/about/), which provide data with locational restrictions; and interest groups such as the Society for American Archaeology's Open Science in Archaeology (https://osf.io/2dfhz/).

A balance must be struck in meeting the security and legal requirements of sensitive archaeological information while effectively communicating spatial patterns. If site security is a concern, this balance is usually created by archaeologists in two ways: mapping sites at low resolution or aggregating sites to grids or administrative boundaries. Often, maps of site locations are not accompanied by an explicit description of security concerns. This article introduces the methods available for obscuring archaeological site locations, the efficacy of those methods, and a recommendation for archaeologists to be mindful and explicit when selecting visualization methods.

Sensitive archaeological sites are defined by the U.S. Department of the Interior as those at risk for damage if their locations are disclosed (NPS 1997). Archaeologists must then determine whether a sensitive site has enough cultural significance or potential for future scientific discovery that its location should be obscured to protect it. It is important to note that current evidence points to local knowledge networks, rather than academic data, as being the source for terrestrial site locations for looters (Proulx Reference Proulx2013; Smith Reference Smith2005). Although there is no evidence of a relationship between publication or OA repositories and terrestrial looting, archaeologists are ethically bound, and often also legally bound, to respect the cultural importance of a site by obscuring or not disclosing its coordinates.

The balance between ensuring the security of sensitive sites and sharing site location information is often implicit. However, Sarah Parcak's GlobalXplorer° web platform is one project that provides a high-profile example of archaeologists explicitly addressing visualizations available to the public and site security. The FAQ section of https://www.globalxplorer.org/ states that site maps will not be made publicly available unless deemed appropriate by the “governments and protection agencies involved.” In addition, the images of the earth's surface that are presented to citizen scientists are not “linked to coordinates or other data that would expose the location of a site” (GlobalXplorer° 2019). The random presentation of these tiled images also prevents systematic use by potential looters.

Another example is DINAA, a repository for North American archaeological data. DINAA does not share actual site locations with the public but rather aggregates site counts. Its website indicates that the data are “publicly viewable as KML data with two forms of representation, at the level of US county and in a ~20km grid” (Anderson et al. Reference Anderson, Kansa, Kansa, Yerka and Wells2011). The grid size was chosen as previous work provided precedent (DINAA 2013). DINAA's explicit discussion of data resolution for the purposes of obscuring data is relatively rare and emphasizes protection of the data, with the site going so far as to not store precise locational data at all. New regional trends are being discovered through grid aggregation (Anderson et al. Reference Anderson, Bissett, Yerka, Wells, Kansa, Kansa, Myers, Carl DeMuth and White2017), although more localized spatial patterning is obscured. McCoy (Reference McCoy2017) also explicitly incorporates the aggregation method when discussing his work with Geospatial Big Data. This article explores how other methods help communicate localized spatial patterning while continuing to protect locational data.

There are existing laws and recommendations in regard to the sharing of archaeological site locations; but in many contexts there is a great deal of autonomy in how to visually represent sites and no available archaeological literature on available techniques for obscuring sites while still communicating archaeological findings. Section 304 of the 2016 National Historic Preservation Act states that information should be withheld from the public if “disclosure could result in a significant invasion of privacy, damage to the historic property, or impede the use of a traditional religious site by practitioners.” This includes

The Advisory Council on Historic Preservation also indicates that “if information is already in the public realm but with very limited accessibility, it does not mean that it can no longer be protected from further disclosure” (2016). This restriction applies only to properties that are listed in the National Register of Historic Places as determined by the keeper of the National Register.

The Cultural Resource Geographic Information System Facility Heritage Documentation Programs of the National Park Service present the complexity of maintaining geospatial data at a national scale in the Draft Set of Standards for Cultural Resource Spatial Data (NPS 2019). The NPS maintains geospatial data for “3000 cultural landscapes, 27,000 historic buildings and structures, 1,200 Ethnographic resources, 63,000 archeological sites, and over 500 American battlefields.” Of concern is that the draft declares that “there are no standards for cultural resource spatial data” (NPS 2019). Attempting to collate or share information across administrative and agency boundaries without standards is modestly described as difficult. To begin to alleviate this issue, the NPS established the Cultural Resource Spatial Data Transfer Standards: Guidelines for Use and Implementation (NPS 2014). The standards include increased granularity in regard to restricting the distribution of cultural data, while noting that all archaeological data are restricted from being released as they fall under the 1979 Archaeological Resources Protection Act (ARPA; NPS 2014:19).

Section 9 of ARPA also prohibits public disclosure of information concerning the nature and location of archaeological resources on federal or Indian land that require a permit or other permission under the act for their excavation or removal. Disclosure is allowed as long as there is no risk of “harm to the archaeological resource” (ARPA, Uniform Regulations, Sec. 7.18). Beyond the designation of sensitive data via Section 304 of the National Historic Preservation Act or Section 9 of ARPA, state historic preservation offices vary in their policies for distributing archaeological site data. For example, the Nevada State Historic Preservation Office maintains the Nevada Cultural Resources Information System (https://shpo.nv.gov/services/nvcris), which provides restricted and unrestricted access levels to archaeological data. Researchers meeting the Secretary of the Interior's Standards for Archeological Documentation may request a subscription to restricted data via e-mail.

In 2018 the Geospatial Data Act was signed into federal law. The act formalized many of the groups and processes that govern spatial data at the national level, including the governance of the National Spatial Data Infrastructure (NSDI). The NSDI includes “the technology, policies, criteria, standards, and employees necessary to promote geospatial data sharing throughout the Federal Government, State, tribal, and local governments, and the private sector (including nonprofit organizations and institutions of higher education)” (Geospatial Data Act of 2018, Sec. 752), established in 1994 by Executive Order 12906. The Congressional Research Service report on the act writes that the first goal of the NSDI is “to ensure that geospatial data from multiple sources . . . are available and easily integrated to enhance the understanding of the physical and cultural world” while guaranteeing protection of private and secure data (2018:6). Practical and appropriate workflows for communicating sensitive archaeological site locations are not yet available.

While this review focuses on laws and policies in the United States, there is a wide range of standards employed internationally. All archaeological sites are covered by the Society for American Archaeology's Principle 6, which encourages “taking into account” preservation and protecting sites when disclosing site locations. This article discusses methods for mapping archaeological sites that take into account these concerns and provides descriptions and assessments of geomasking methods to allow researchers to select appropriate geomasking techniques and parameters regardless of where they conduct research. This article calls on findings from the field of public health to explore the degree to which methods obscure actual site locations.

VULNERABILITY OF LOCATION-SHARING PRACTICES

Geographic information systems (GIS) are powerful technologies for collecting, managing, analyzing, and visualizing spatial information. Geospatial data include any data with a locational component. The precision and accuracy of geospatial products depend on researchers’ decision making and the equipment used. As introduced above, archaeologists must balance the security of an archaeological site according to its importance and vulnerability, while meeting expectations for responsible, open science. This article does not provide techniques to assess a site's sensitivity but explores the available options for obscuring site locations and how those options balance security and scholarship.

A number of public health publications expose security flaws in common location-obscuring practices in visualizing spatial data. Brownstein, Cassa, and Mandl (Reference Brownstein, Cassa and Mandl2006) published a correspondence piece in the New England Journal of Medicine in which they called for guidelines for representing patients’ homes to preserve anonymity. They suggested that the common practice of using lower-resolution maps was less effective than aggregating patients to administrative units (e.g., census tracts) or their preferred method of randomly changing a patient's location within a fixed distance. Over the next decade public health specialists conducted reverse geocoding experiments to test geomasking methods (Allshouse et al. Reference Allshouse, Fitch, Hampton, Gesink, Doherty, Leone, Serre and Miller2010; Boulos et al. Reference Boulos, Kamel, Curtis and Malik2009; Brownstein, Cassa et al. Reference Brownstein, Cassa, Kohane and Mandl2006; Hampton et al. Reference Hampton, Fitch, Allshouse, Doherty, Gesink, Leone, Serre and Miller2010; Seidl et al. Reference Seidl, Jankowski and Clarke2017; Zandbergen Reference Zandbergen2014). The process of finding an address from a coordinate pair is referred to as reverse geocoding. Geomasking is a process of obscuring the location of real-world coordinates. After geomasking a set of coordinates, researchers found the probability that a patient's true location could be identified from available patient data, the distance between the observed and offset coordinates, and the population density of the area. Their work reveals a number of geomasking techniques that balance security and open science (sharing unobscured location data) in different ways.

Instead of patient or participant anonymity, archaeologists concerned about site security should consider how difficult it would be to identify a site from a geomasked map and how rigorous an attempt should be made to obscure sites. Only two geomasking techniques are routinely employed by archaeologists, low-resolution mapping and aggregation, with only one or two others having occasional representations in the literature. The following section describes geomasking methods that can be applied to archaeological sites. Each balances security and open data differently.

METHODS FOR VISUALIZING SENSITIVE LOCATION DATA

Aggregation shows the number of sites within a given boundary, with the most common boundaries being administrative units (e.g., counties; Figure 1) and grids (Figure 2). When purposefully and explicitly obscuring archaeological sites, this is the most common method employed. This method can show distributions over a large area but does not allow for spatial patterning at local levels.

FIGURE 1. Sites aggregated to county boundaries. Darker color represents more sites. (State boundary data from Esri and TomTom North America 2019.)

FIGURE 2. Archaeological sites aggregated by 700 km² grid cells. Dark blue cells represent the presence of two sites, and light blue represents one site. (State boundary data from Esri and TomTom North America 2019.)

Low-resolution maps indicate the actual observed location, with the assumption that the low resolution (e.g., a scale of 1:5,000,000) will make it difficult to find the real-world location (Figure 3).

FIGURE 3. Archaeological sites mapped at the low resolution of 1:5,000,000. (Site data from University of Wyoming Department of Geography et al. 2017; state boundary data from Esri and TomTom North America 2019; topographic data from National Geographic Society and i-cubed 2019.)

Heat maps are representations of the density of observations; for example, the density of archaeological sites (Figure 4). Densities are calculated using a neighborhood, or kernel, the size of which is defined by the user.

FIGURE 4. Archaeological sites depicted by heat map produced with a kernel size of 50 km. (State boundary data from Esri and TomTom North America 2019; topographic data from National Geographic Society and i-cubed 2019.)

Bounding boxes are hollow squares surrounding all the observations of a given area but do not depict the site locations (Figure 5).

FIGURE 5. Bounding boxes drawn around individual and clusters of sites. (State boundary data from Esri and TomTom North America 2019; topographic data from National Geographic Society and i-cubed 2019.)

Coordinate patterns without base maps are used to accurately convey the spatial relationships of observations but without the context of topographic or administrative features (Figure 6).

FIGURE 6. Locations of archaeological sites with the topographic base map removed. (Site data from University of Wyoming Department of Geography et al. 2017.)

Random direction with a fixed radius allows the user to define a particular distance from which the geomasked point is shown away from the observed coordinate (Figure 7). The direction in which the geomasked point is placed is randomly selected via algorithm.

FIGURE 7. Random direction fixed radius. The geomasked point (blue) is placed in a random direction 1 km away from the observed point (black). The black ring represents possible locations for a geomasked point.

Random perturbation within a fixed radius allows a maximum distance from which a geomasked point may be offset from an observed point (Figure 8). The geomasked point will be placed in a random direction away from the observed point and at a random distance within the maximum distance specified by the user.

FIGURE 8. Random perturbation. The geomasked point (blue) is placed in a random direction at a random distance within 1 km of the observed point (black). Brown represents the area in which a geomasked point could be placed.

Random perturbation donuts are similar to random perturbation within a fixed radius but also include a minimum distance from which the geomasked points must be placed from the observed point (Figure 9). Overall, the donut method is generally the most effective at masking locations in public health studies while still providing a visual representation of spatial patterning.

FIGURE 9. Random perturbation donut. The geomasked point (blue) is placed in a random direction in a random distance within 1 km of the observed point (black) at a minimum distance of 250 m. Brown represents the area in which a geomasked point could be placed.

Gaussian displacement resembles random perturbation within a fixed radius, with the exception that geomasked points are distributed in a normal distribution away from the observed point, with the maximum distance being set by a contextual variable (Figure 10). For example, in public health, the greater the population density (the contextual variable), the smaller the maximum distance from an observed point.

FIGURE 10. Gaussian displacement. A geomasked point (blue) is offset from the observed point (black) at a distance determined by at least one contextual variable, such as density of sites. The orange field presents the area in which a geomasked point may be placed, while the intensity of the orange indicates the likelihood that a geomasked point will be placed in that specific location.

Gaussian donuts, also referred to as bimodal Gaussian displacement, are similar to Gaussian displacement but also set a minimum threshold at which the geomasked points can be placed from the observed points (Figure 11). Gaussian donuts are also effective in masking true locations while showing spatial patterning: the added dimension of adjusting the maximum displacement range allows for smaller offsets in densely populated areas.

FIGURE 11. Gaussian donuts. A geomasked point (blue) is offset at a distance determined by at least one contextual variable, such as density of sites, at a minimum distance of 250 m. The orange field presents the area in which a geomasked location may be placed, while the intensity of the orange indicates the likelihood that a geomasked point will be placed in that specific location.

Programming models, such as linear programming and digit switching, involve using code to systematically obscure real-world coordinates. Linear programming considers each observation and applies models that include the probability of successful reverse geocoding, constraints, and objectives. Digit switching changes two or more digits of coordinates in the Military Grid Reference System. Another code is used as a key to replace the original digits. By specifying the digits to be switched, the user has control over the maximum distances to which coordinates will be geomasked. Illustrations of linear programming can be found in Wieland and colleagues (Reference Wieland, Cassa, Mandl and Berger2008). Illustrations of digit switching can be found in Clarke (Reference Clarke2016). Another example of applying models to randomly perturb coordinate locations is described in Fronterrè and colleagues (Reference Fronterrè, Giorgi and Diggle2018).

The Voronoi or Thiessen method identifies the center of polygonal areas, usually land parcels in public health studies, and creates Voronoi, also known as Thiessen, polygons. Voronoi polygons are the result of lines drawn through the midpoints between observed points. Geomasked points for each observation are placed on the closest part of the closest Voronoi edge. For parcel data, this method generally avoids placing a geomasked point near a polygon centroid. Illustrations of this method can be found in Croft and colleagues (Reference Croft, Shi, Sack and Corriveau2016).

Location swapping identifies neighborhoods with similar geographic characteristics to a point being geomasked. The original location of that point is then swapped with a location from one of those similar neighborhoods. Variations of this method, such as location swapping with donut, are explored in Zhang and colleagues (Reference Zhang, Freundschuh, Lenzer and Zandbergen2017).

Scholarship and security are balanced differently in each of these approaches. First, skewed heavily in the direction of security are aggregation, heat maps, and bounding boxes. These techniques make it improbable that the actual sites will be identified, at the cost of entirely obscuring spatial patterning at the local level, for other archaeologists and the public. Also skewed heavily in favor of security are observed site locations published without a base map. Without topographical clues, identifying the actual sites would be very difficult. However, there is the risk that once one site is identified, the other site locations could be identified using the published spatial relationships.

Versions of donut masking strike a more calculated balance between scholarship and security. By placing a geomasked point at some distance from an observed point, researchers decrease the size of the area that would need to be searched to locate an archaeological site. Studies in public health show that the risk of reverse geocoding an observed location is much smaller with donut methods than with low-resolution mapping—a commonly deployed technique in archaeology and, prior to the last decade, the most commonly deployed in public health. Donut masking, some programming models, and Voronoi methods have also been explored for how well they maintain spatial relationships.

Comparison of geomasking techniques in archaeological contexts is necessary because the data types and the objectives of archaeologists differ from those of public health researchers. The concern in public health is that personal data could be connected back to an individual person. Making this sort of connection would usually involve using parcel data from a city or county to match identifying characteristics such as age, ethnicity, and sex to health data. The archaeological concern is to protect site locations and adhere to related laws and policies. In this regard, the archaeologist is interested in balancing the obscurity of the site (the size of the area that would require survey to identify a site), open science (providing access to real-world coordinates to repeat analysis and interpretation), and facilitating visualization in published works (geomasking coordinates to mimic observed spatial or geographic patterns).

METHODS FOR COMPARING GEOMASKING TECHNIQUES FOR ARCHAEOLOGICAL SITES

A comparison of geomasking techniques was conducted using 50 archaeological sites in Wyoming whose coordinates are publicly available (University of Wyoming Department of Geography et al. 2017). Techniques that do not result in a masked point pattern—for example, those resulting in a masked area (e.g., heat maps and aggregation)—were not included in the comparison. In addition, Voronoi methods and programming methods were not included. Voronoi methods require an underlying polygon layer from which to create centroids, which is more conducive to studying modern residential associations (e.g., parcel data) than archaeological sites. Programming methods may be tailored to archaeology, but no work has been completed in this area. Existing location-switching programs are also more conducive to modern residential data where placing a geomasked point on an existing residence outside the study area is good practice, whereas placing a geomasked point on a known archaeological site outside the study area would not be good practice.

Low-resolution maps, random direction buffers with fixed radius, random perturbation donuts, and Gaussian donuts were applied to the 50 archaeological sites with varying parameters to produce 50 geomasked points for each type. Random perturbation and Gaussian methods without minimum offset “donuts” were not included, because the methods allow geomasked points to fall in close proximity or on top of observed sites. Low-resolution maps were tested by creating static maps of observed site locations at the scales of 1:100,000, 1:5,000,000, and 1:20,000,000 in Google Maps. These maps were brought into ArcGIS Pro, a commercial GIS platform that is in wide use in cultural resource management. Once in ArcGIS Pro, the maps were georeferenced, and then new points were placed on the observed points.

To create geomasked points using random direction fixed radius and random perturbation donuts, buffers were created at a maximum distance of 1,000 m from observed sites and random points were assigned either to the circumference of the buffer for fixed radius types or between the circumference and a minimal offset distance of 250 m. This was done in ArcGIS Pro but could easily be done in an open-source software such as QGIS. Random perturbation without a minimum distance was excluded as the geomasked points would occur in close proximity to the observed site.

Gaussian displacement donuts were created using the v.perturb GRASS command in QGIS (GRASS GIS 2020). This command allows users to specify the mean and standard deviation of a normal distribution of random points within a given area. In this case a minimum distance of 250 m was selected to prevent points from occurring in close proximity to an observed site. Standard deviations of 500 m and 750 m were tested, as these distributions approximate the 1,000 m maximum distance used in the fixed distance and random perturbation methods.

Once each geomasked location set was created using the above methods, three measures were recorded: the distance between each observed and geomasked point, the minimum resulting area that would need to be searched to find the observed site based on the geomask method applied, and the Global Divergence Index (Gdi). The Gdi is one measure of how far displaced geomasked points are from the observed site locations (Kounadi and Leitner Reference Kounadi and Leitner2014; Seidl et al. Reference Seidl, Jankowski and Clarke2017). The smaller the Gdi, the smaller the displacement. The index is calculated by first finding the mean centers and the standard deviational ellipses of observed and geomasked datasets, as proposed by Kounadi and Leitner (Reference Kounadi and Leitner2014). The mean center is found by calculating the average x- and y-coordinates for a set of points. A standard deviational ellipse is created by using the standard deviations of the x- and y-coordinates from the mean center as the major and minor axes of the ellipse. Figure 12 depicts the parts of a standard deviational ellipse used in the calculation. Next, the divergences between the mean centers, orientation of the ellipses, and major axes of the ellipses are calculated. The three divergence values are then averaged to produce the Gdi:

$$\overline {{\rm GDi}} = \lpar {{\rm Mdi}\comma \;{\rm \;}\,{\rm Odi}\comma \;\,{\rm \;MAdi}} \rpar$$

FIGURE 12. Example of a standard deviational ellipse and mean center, which are calculated using the x- and y-coordinates of a set of locations.

Mdi is the divergence of the mean centers. In addition to using the distance between the original and geomasked mean centers, the distance between the observed mean center and the farthest point in the study area is also used. The study area for this project is the state of Wyoming, with the northeast corner of the state occurring farthest from the mean center of the observed archaeological sites:

$$\eqalign{{\rm Mdi} = {\rm \;}\displaystyle{{{\rm distance\;}\,{\rm of\;}\,{\rm observed\;}\,{\rm mean}\,{\rm \;to}\,{\rm \;geomasked\;}\,{\rm mean}} \over \matrix{{\rm distance\;}\,{\rm of}\,{\rm \;observed\;}\,{\rm mean\;}\,{\rm to\;}\,{\rm farthest\;}\,\cr{\rm point\;}\,{\rm in\;}\,{\rm the}\,{\rm \;study}\,{\rm \;area}}}\,{\rm \;} \times \,100}$$

Odi is the divergence between the orientations of the ellipses. The orientation is the degree of the angle between the major axis and north:

$${\rm Odi} = {\rm \;}\displaystyle{\matrix{{\rm orientation}\,{\rm \;of}\,{\rm \;observed}\,{\rm \;ellipse}-\cr {\rm orientation}\,{\rm \;of}\,{\rm \;geomasked\;}\,{\rm ellipse\;}} \over {180}}{\rm \;} \times 100$$

MAdi is the divergence between the lengths of the major axes of the ellipses. In addition to the lengths of the observed and geomasked ellipses, the length of the longest major ellipse axis (Maximum.MA) that could be drawn through the entire study area is required. The length of the major axis for the geomasked ellipse is compared with either the difference between the maximum major axis length and the observed length or the difference between the minimum major axis length and the observed length. Whichever difference is greater is used for the comparison. The minimum possible major axis length for the study area is so close to zero that it does not significantly impact the resulting index value and is therefore not included in the denominator of the second calculation in Figure 13.

FIGURE 13. The Major Axis Divergence Index (MAdi) compares the difference in orientation of major axes of one-standard deviational ellipses formed by observed and geomasked points.