Georeferencing Best Practices

For a location that is a transect, record both the start and end points of the line. This allows the orientation and direction of the transect to be preserved. If the events associated with the transect occur within a given maximum distance from the transect, it is better to represent the location as a polygon (see §2.3.3). If the events associated with the transect can be reasonably separated into their individual locations, it is better to do so, as these will be more specific than the transect as a whole. If that is done, however, ensure that you document that each individual location is part of a transect.

If the locality is recorded as the center of the transect and half the length of the transect is then used to describe uncertainty, information about the orientation of the transect is lost, and the description essentially becomes equivalent to a circle.

Not all linear-based locations are transects or straight lines. We use the term path to highlight this broader concept. Illustrative examples are: ad-hoc observations while walking along a trail, an inventory or count of species while travelling along a river, tracking an individual animal’s movements. Marine transects, tracks, tows, and trawls, are further examples. Paths should be described using shapes (see discussion under §3.3.4) as connected line segments (a polygonal chain), with the coordinates of the starting point followed by the coordinates of each segment beginning and finishing with the end point. One simple way to store and share these is through Well-Known Text (WKT) (ISO 2016, De Pooter et al. 2017, OBIS n.d., W.Appeltans, personal communication 15 Apr 2019).

To determine the uncertainty of a described path using the point-radius georeferencing method, one needs to determine the corrected center – i.e. the point on the path that describes the smallest enclosing circle that includes the totality of the path ("c" on Figure 3). This is very seldom the same place as the center of a line joining the two ends of the path ("y" on Figure 3), nor the center of the extremes of latitude and longitude (the geographic center) of the path ("x" on Figure 3).

When collecting or recording data from an area, for example, bird counts on a lake, a set of nesting or roosting sites on an offshore coral cay, or a buffered transect – the location is best recorded as a polygon. Polygons can be stored using the Darwin Core (Wieczorek et al. 2012b) field called dwc:footprintWKT, in which a geometry can be stored in the Well-Known Text format (ISO 2016). For the point-radius georeferencing method, if the polygon has a concave shape (for example a crescent), the center may not actually fall within the polygon (Figure 4). In that case, the corrected center on the boundary of the polygon is used for the coordinates of the location and the geographic radial is measured from that point to the furthest extremity of the polygon. Note that the circle based on the corrected center (red circle in Figure 4) will always be greater than the circle based on the geographic center (black circle in Figure 4).

Complex polygons, such as donuts, self-intersecting polygons and multipolygons create even more problems, in both documentation and storage.

Grids may be based on the lines of latitude and longitude, or they may be cells in a Cartesian coordinate system based on distances from a reference point. Usually grids are aligned North-South, and if not, their magnetic declination is essential to record. If the extent of a location is a grid cell, then the ideal way to record it would be the polygon consisting of the corners of the grid (i.e. a bounding box). The point-radius method can be used to capture the coordinates of the grid cell center and the distance from there to one of the furthest corners, but given that the geometries for grid cells are so simple, it is best to also capture them as polygons. Often grid cells (e.g. geographic grids) are described using the coordinates of the southwest corner of the grid. Using the southwest corner as the coordinates for a point-radius georeference is wasteful, since the geographic radial would be from there to the farthest corner, which would be twice as far as it would be if the center of the grid cell was used instead. In any case, the characteristics of the grid should be recorded with the locality information.

It is important when converting gridded data to geographic coordinates to also check the locality description. Locality information may allow you to refine the location as in Figure 5 where just having the grids without the locality information (i.e. "on Northey Island") would lead to the circle (c) with its center (a) at the center of the grid. Knowing that the record is on Northey Island, however, allows you to refine the location to the smaller circle (d) with its center at (b). Note that other criteria (such as a change of datum, map scale, etc.) may add to the uncertainty.

Most terrestrial locations are recorded with reference to the terrestrial surface as geographic coordinates, sometimes with elevation. Some types of marine events such as dives and trawls, benefit from explicit description in three dimensions.

Diving events are commonly recorded using the geographic coordinates of the point on the surface where the diver entered the water, called entry point or point of entry. The underwater location should be recorded as a horizontal distance and direction along with water depth from that surface location (see Figure 7). Below the surface the diver may then begin a collection/observation exercise in three dimensions from that point including a horizontal component and a minimum and maximum water depth. These should all be recorded. The reference point should be the corrected center of the 3D-shape that includes the extent of the location. The geographic radial would be the distance from the corrected center of the 3D shape (the three dimensions projected perpendicularly onto the surface) to the furthest extremity of the projection of the 3D-shape in the horizontal plane (i.e. on the geographic boundary).

There are many different types of trawls and tows, including bottom and mid-water trawls. The 3D nature should be captured as above. The geographic reference points would be line segments tracing the route of the trawl, and would be more akin to paths and captured as a shape as described in §2.3.2.

Geographic coordinates are a convenient way to define a location in a way that is not only more specific than is otherwise possible with a locality description, but also readily allows calculations to be made in a GIS. Geographic coordinates can be expressed in a number of different coordinate formats (decimal degrees, degrees minutes seconds, degrees decimal minutes), with decimal degrees being the most commonly used. Geographic coordinates in decimal degrees are convenient for georeferencing because this succinct format has global applicability and relies on just three attributes, one for latitude, one for longitude, and one for the geodetic datum or ellipsoid, which, together with the coordinate format, make up the coordinate reference system. By keeping the number of recorded attributes to a minimum, the chances for transcription errors are minimized (Wieczorek et al. 2004).

When capturing geographic coordinates, always include as many decimals of precision as given by the coordinate source. Coordinates in decimal degrees given to five decimal places are more precise than a measurement in degrees-minutes-seconds to the nearest second, and more precise than a measurement in degrees and decimal minutes given to three decimal places (see Table 3). Some new GPS/GNSS receivers now display data in decimal seconds to two decimal places, which corresponds to less than a meter everywhere on Earth. This doesn’t mean that the GPS reading is accurate at that scale, only that the coordinates as given do not contribute additional uncertainty.

Universal Transverse Mercator (UTM) is a system for assigning distance-based coordinates using a Mercator projection from an idealized ellipsoid of the surface of the Earth onto a plane. In most applications of the UTM system, the Earth is divided into a series of six-degree wide longitudinal zones extending between 80°S and 84°N and numbered from 1-60 beginning with the zone at the Antimeridian (Snyder 1987). Because of the latitudinal limitation in extent, UTM coordinates are not usable in the extreme polar regions of the Earth. A map of UTM zones can be found at UTM Grid Zones of the World (Morton 2006).

UTM coordinates consist of a zone number, a hemisphere indicator (N or S), and easting and northing coordinate pairs separated by a space with 6 and 7 digits respectively, and all in the order given here. For example, for Big Ben in London (latitude 51.500721, longitude −0.124430), the UTM reference would be: 30N 699582 5709431.

Latitude bands are not officially part of UTM, but are used in the Military Grid Reference System (MGRS). They are used in many applications, including in Google Earth. Each zone is subdivided into 20 latitudinal bands, with letters used from South to North starting with "C" at 80°S to "X" (stretched by an extra 4 degrees) at 72°N (to 84°N) and omitting "O". All letters below "N" are in the southern hemisphere, "N" and above are in the northern hemisphere. When using latitudinal bands, "north" and "south" need to be spelled out to avoid confusion with the latitudinal bands of "N" and "S" respectively. Using the latitudinal band method, the coordinates for Big Ben would be: 30T 699582m east 5709431m north.

National and local grid systems derived from UTM, but which may be based on different ellipsoids and datums, are basically used in the same way as UTMs. For example, the Map Grid of Australia (MGA2020) uses UTM with the GRS80 ellipsoid and Geocentric Datum of Australia (GDA2020) (Geoscience Australia 2019b). An example of a location in MGA2020 is "MGA Zone 56, x: 301545 y: 7011991"

When recording a location, or databasing using UTM or equivalent coordinates, a zone should ALWAYS be included; otherwise the data are of little or no value when used outside that zone, and certainly of little use when combined with data from other zones. Zones are often not reported where a region (e.g. Tasmania) falls completely within one UTM zone. This is OK while the database remains regional, but is not suitable for exchange outside of the zone. When exporting data from databases like these, the region’s zone should be added prior to export or transfer. Better still, modify the database so that the zone remains with the coordinates.

Note that Darwin Core (Wieczorek et al. 2012b) supports UTM coordinates only in the verbatimCoordinates field. There are several tools to convert UTM coordinates to geographic coordinates, including Geographic/UTM Coordinate Converter (Taylor 2003)–see Georeferencing Tools. For details on georeferencing, see Coordinates – Universal Transverse Mercator (UTM) in Georeferencing Quick Reference Guide (Zermoglio et al. 2020).

One of the most important issues for consideration when choosing a GPS or GNSS receiver is the antenna. An antenna behaves both as a spatial and frequency filter, therefore, selecting the right antenna is critical for optimizing performance (Novatel 2015). One of the drawbacks with smartphones, for example, is the limited size of the GNSS antenna.

For information on issues to consider when selecting an appropriate GNSS antenna and/or GPS receiver, we refer you to Chapter 2 in Novatel 2015 and Chapter 10 in NLWRA 2008.

Most GPS devices are able to report a theoretical horizontal accuracy based on local conditions at the time of reading (atmospheric conditions, reflectance, forest cover, etc.). For highly specific locations, it may be possible for the potential error in the GPS reading to be on the same order of magnitude as the extent of the location. In these cases, the GPS accuracy can make a non-trivial contribution to the overall uncertainty of a georeference.

The latest US Government commitment (US Deptartment of Defense and GPS Navstar 2008) is to broadcast the GPS signal in space "with a global average user range error (URE) of ≤7.8 m (25.6 ft.), with 95% probability". In reality, actual performance exceeds this, and in May 2016, the global average URE was ≤ 0.715 m (2.3 ft), 95% of the time (GPS.gov 2017). Though it does not mean that all receivers can obtain that accuracy, the accuracy of GPS receivers has improved and today most manufacturers of handheld GPS units promise errors of less than 5 meters in open areas when using four or more satellites. The need for four or more satellites to achieve these accuracies is because of the inaccuracies in the clocks of the GPS receivers as opposed to the much more accurate satellite clocks (Novatel 2015). The accuracy can be improved by averaging the results of multiple observations at a single location (McElroy et al. 2007), and some modern GPS receivers that include averaging algorithms can bring the accuracy to around three meters or less. According to GISGeography 2019a, “A well-designed GPS receiver can achieve a horizontal accuracy of 3 meters or better and vertical accuracy of 5 meters or better 95% of the time. Augmented GPS systems can provide sub-meter accuracy”. Another method to improve accuracy is to average over more than one GPS unit. Note that some GPS/GNSS receivers can record up to 20 decimal places of precision, but that doesn’t mean that is the accuracy of the unit.

The use of Differential GNSS (DGNSS) (incorporating Differential GPS (DGPS)) can improve accuracy considerably. DGNSS references a GNSS Base Station (usually a survey control point) at a known position to calibrate the receiving GNSS signal. The Base Station and handheld GNSS receiver reference the satellites’ positions at the same time and thus reduces error due to atmospheric conditions, as well as (to a lesser extent) satellite ephemeris (orbital location) and clock error (Novatel 2015). The handheld GNSS instrument applies the appropriate corrections to the determined position. Depending on the quality of the receivers used, one can expect an accuracy of <1 meter (USGS 2017). This accuracy decreases as the distance of the receiver from the Base Station increases. It is important to note that differential technology is not available in all areas – for example, in remote locations and remote islands, and the resulting accuracy may be less than expected. Again, averaging can further improve on these values (McElroy et al. 2007). It is important to note, however, that most DGNSS is post-processed. Records are stored in the GPS/GNSS unit and then post-processing software is run to improve the measurements once connected to a computer. Post processing is not as commonly used since the introduction of real-time DGNSS, such as the Satellite Based Augmentation System, see the next subsection below), and is now used mostly in surveying applications where high accuracy is required.

Marine horizontal position accuracy requirements are 2-5 meters (at a 95 percent confidence level) for safety of navigation in inland waters, 8-20 meters (95%) in harbor entrances and approaches, and horizontal position accuracies of 1-100 meters (95%) for resource exploration in coastal regions (Skone et al. 2004, Skone & Yousuf 2007). While DGNSS horizontal error bounds are specified as 10 meters (95%) studies have shown that under normal operating conditions accuracies fall well within this bound.

DGNSS accuracies are susceptible to severe degradation due to enhanced ionospheric effects associated with geomagnetic storms. Degradation can be in the order of 2-30 times in some areas and depending on the severity of the storm.

Satellite Based Augmentation System (SBAS) is a collection of geosynchronous satellites originally developed for precision guidance of aircraft (Federal Aviation Administration 2020) and more recently to provide services for improving the accuracy, integrity and availability of basic GNSS signals (Novatel 2015). SBAS receivers are inexpensive examples of real-time differential correction. SBAS uses a network of ground-based reference stations to measure small variations in the GNSS satellite signals. Measurements from the reference stations are routed to master stations, which queue the received Deviation Correction (DC) and send the correction messages to geostationary satellites. Those satellites broadcast the correction messages back to Earth, where SBAS-enabled GPS/GNSS receivers use the corrections while computing their positions to improve accuracy. Separate corrections are calculated for ionospheric delay, satellite timing, and satellite orbits (ephemerides), which allows error corrections to be processed separately, if appropriate, by the user application.

Ground Based Augmentation Systems (GBAS), also known as Local Area Augmentation Systems (LAAS), provide differential corrections and satellite integrity monitoring in conjunction with VHF radio, to link to GNSS receivers. A GBAS consists of several GNSS antennas placed at known locations with a central control system and a VHF radio transmitter. GBAS is limited in its coverage and is used mainly for specific applications that require high levels of accuracy, availability and integrity, and is the system largely used for airport navigation systems.

Precise Point Positioning (PPP) depends on GNSS satellite clock and orbit corrections, generated from a network of global reference stations to remove GNSS system error and provide a high level (decimeter) of positional accuracy. Once the corrections are calculated, they are delivered to the end user via satellite or over the Internet.

Although similar to Satellite Based Augmentation System (SBAS) systems (see above), they generally provide a greater accuracy and have the advantage of providing a single, global reference stream as opposed to the regional nature of an SBAS system. Whereas SBAS is free, the use of PPP usually incurs a charge to access the corrections, so it is unlikely that the increased accuracy of PPP when compared to that of SBAS, will be a consideration for most biological applications.

Static GPS uses high precision instruments and specialist techniques and is generally employed only by surveyors. Surveys conducted in Australia using these techniques reported accuracies in the centimeter range. These techniques are unlikely to be extensively used with biological record collection due to the cost and general lack of requirement for such precision.

High-end dual and multi-frequency GPS/GNSS devices can bring accuracy to the centimeter level, and even mm level over the long-term (GPS.gov 2017). One of the ways this is done is by removing one of the largest contributors to overall satellite error － error due to the ionosphere (known as ionosphere error) (Novatel 2015).

GPS-enabled smartphones are typically accurate to within 4.9 m (16 ft.) under open sky, however, their accuracy worsens near buildings, bridges, and trees (GPS.gov 2017). A study by Tomaštik et al. 2017 found that the accuracy of smartphones in open areas was around 2-4 m. This decreased to 3-11 m in deciduous forest without leaves, and 3-20 m in deciduous forest with leaves. There are reports that the accuracy in some GPS-enabled smartphones will soon be improved to <1 meter (Moore 2017) and that accuracy in areas with restricted satellite view within cities will be improved drastically with inbuilt 3D smartphone apps and probabilistic shadow matching (Iland et al. 2018). In general, the GNSS chipsets in smartphones are quite good, and any loss of accuracy is usually due to the quality of the antenna, whose chief failing is due to their poor multipath suppression (Pirazzi et al. 2017). In some smartphones where good satellite coverage is unavailable (e.g. in cities and forests), the phone may introduce errors from bias in its internal clock (Pirazzi et al. 2017), leading to occasional large inaccuracies (Arturo Ariño Oct 2019, pers. comm.). Already the technology for better than 1 meter smartphone accuracy exists, but it is not available to the public due to the difficulty and cost of incorporating the technology into small smartphones (Braun 2019). The accuracies reported in most publications refer to studies in the USA, Europe, coastal Australia, India or Japan where good differential stations are plentiful. More studies are needed to test smartphone accuracies in remote locations and where differential stations are not available.

Smartphone GPS technology is changing rapidly and there is likely to be new and updated information even before this document is published.

We are not aware of the characteristics of the accuracy of GPS-enabled cameras, but we expect the accuracy to be similar to that of smartphones. One study, using three different cameras, showed variation between the three and the true location to be less than 3 m from the reported location (Doty 2017). Note that GPS-enabled cameras that are used for snorkeling and diving activities, will only give new GPS readings each time the camera is brought to the surface.

Over the years, a number of methods for tracking a diver underwater with a GPS have been tried with limited success. These included using a floating GPS receiver over the diver’s bubbles, and a GPS receiver on a raft towed by the diver that recorded intermittent readings to provide a dive transect (Schories & Niedzwiedz 2011). The most successful to date has been the use of a GPS antenna on a floating buoy that is attached by a cable to a diver-held GPS. These diver-towed underwater GPS/GNSS handheld receivers have been used for underwater monitoring studies for several years. Most dives using this method are at <20 meters as the signal deteriorates with cable length giving a maximum practical depth of 50 meters (Niedzwiedz & Schories 2013). One problem is cable drag, and it is almost impossible to determine the buoys offset exactly although Niedzwiedz & Schories 2013 provide formulae for attempting to do so. A study by the same authors (Schories & Niedzwiedz 2011) showed displacement of 2.3 m at a depth of 5 m, 3.2 m at 10-m depth, 4.6 m at 20-m depth, 5.5 m at 30-m depth, and 6.8 m at 40-m depth. These are in addition to GPS accuracy discussed under §2.6.2.

A barometric altimeter uses changes in air pressure as a proxy for changes in elevation, and can be a reliable source of elevation if properly calibrated. Calibration requires that the elevation of the altimeter be set to a known starting elevation, which could be determined from a map, for example. Thereafter, as the altimeter goes higher or lower in elevation, it estimates the new elevation directly from the air pressure it experiences. Since weather conditions can change the air pressure independently of changes in elevation, it is important to re-calibrate the altimeter frequently, either by recording the elevation when you stop moving and resetting to that same elevation before starting out again, and/or by recalibrating to known elevations whenever you encounter them.

In theory it would be possible to use a barometric altimeter to determine elevations when in a subterranean location (cave, mine, etc.), but these situations are particularly prone to changes in air pressure independent from elevation changes (especially in caves with narrow openings), so recalibration would have to be particularly careful.

Elevation can be determined using the contours and spot height information from a suitable scale map of the area. In general, the uncertainty in the elevation when read from a map is half the contour interval.

For information on determining accuracy from a map, see §3.4.2.1.

Elevation accuracy as reported from a GPS has improved markedly in recent years, but elevation accuracy is not usually reported by GPS/GNSS receivers. As a general rule, for most non-Satellite Based Augmentation System (SBAS) or Wide Area Augmentation System (WAAS) enabled GPS/GNSS receivers, elevation error is approximately 2-3 times the horizontal error (USGS 2017). It is hard to find definitive information for smartphones, but it would appear that this same multiplier is a good rule for those as well. With WAAS-enabled GPS, the FAA reports that 95 per cent of the time vertical error is less than 4 meters (FAA 2019). However, the elevation reported on the GPS receiver or smartphone is not necessarily referring to mean sea level (MSL) as reported, but to the zero elevation of the ellipsoid of the datum – see discussion below.

Note that GPS elevation readings can represent one of at least two different values, depending on the method used by the GPS. Elevation reported can be the geometric height. This is the only value that GPS devices can actually measure, and is the height based on the ellipsoid of the datum. The elevation reported can also be the elevation above MSL, or orthometric height. These values are not directly measured by the GPS, but are calculated as the difference between the geometric height (measured) and the geoid height. The geoid height depends on the geoid and the datum you are trying to compare it against. Thus, to understand the potential difference between elevations based on MSL and those based on the geometric model, the geometric model (datum) must be known. To calculate the potential error using WGS84 datum at a given geographic location, use the Geoid Height Calculator (UNAVCO 2020). For further discussion about these methods, consult Eos Positioning Systems 2018. For a good explanation of the differences between the geoid and mean sea level, we refer you to GISGeography 2019b.

In 2022, the USA will release a new geometric reference frame and geopotential vertical datum that will replace existing USA geometric vertical datums. Similarly, over the next five years, Australia will move to a new generation height reference frame – the Australian Gravimetric Quasigeoid 2017 (AGQG 2017) (McCubbine et al. 2019). The new reference frames will rely primarily on Global Navigation Satellite Systems (GNSS), as well as on an updated gravimetric geoid model (National Geodetic Survey 2018). The new method of calculating vertical datums will improve vertical accuracies to around 1-2 cm, will provide more accurate GPS-determined elevations (Ellingson 2017), and will allow for dynamic updating. Other jurisdictions are likely to move to new methods of calculating vertical datums over time, meaning that within five years most users will be able to position themselves vertically using mobile Global Navigation Satellite Systems (GNSS) technology with sub-decimeter accuracy (Brown et al. 2019).

Digital Elevation Models (DEM) are based on elevations above mean sea level (or more recently, the geoid). The models are calculated using sophisticated interpolations and do not necessarily correspond to the actual surface elevation. DEM vertical accuracy is influenced by several factors such as grid size, slope, land cover, and geolocation (horizontal) error, as well as other biases due to the original DEM data collection (e.g. satellite imaging geometry) and/or production method (Mukherjee et al. 2013, Mouratidis & Ampatzidis 2019). Global DEMs such as the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global DEM V2 (Meyer 2011) and the Shuttle Radar Topography Mission (SRTM) are based on 1 arc-second grids (about 30 m x 30 m) (Farr et al. 2007) and have an accuracy of better than 17 m and 10 m respectively (except for in steep terrain such as mountains, and areas with very smooth sandy surfaces with low signal to noise ratio, such as the Sahara Desert (Farr et al. 2007)). Local and regional DEMs may have a smaller grid size. For example, a 5 m grid in Australia, which has a vertical accuracy better than one meter, and even to 0.3 meter in some areas (Geoscience Australia 2018) or the European Digital Elevation Model, which has an accuracy of better than three meters (Mouratidis & Ampatzidis 2019). Note also that satellite image-based DEMs, being radar based, vary greatly over different land surfaces, forests, shrub or herbaceous vegetation, agricultural areas, bare areas, rocky surfaces, wetlands, and artificial surfaces such as cities. Also the radar can penetrate into areas of snow, ice, and sand (as in deserts) (Mouratidis & Ampatzidis 2019).

Some smartphones, whether they incorporate GPS capabilities or not, use apps that provide elevation values based on a DEM. With smartphone GPS apps, be aware that some devices and apps incorrectly record the method used. The uncertainty in elevation due to an unknown elevation source can be up to 100 meters. For example, the difference with datum WGS84 between the ellipsoid and geoid or mean sea level methods of reporting elevation is shown in Figure 8. Note also that these uncertainties are in addition to the uncertainties associated with the measurements themselves. The only true way of determining what your GPS receiver or smartphone is recording is to test it against a known elevation. Some preliminary studies by the authors show elevation accuracy from smartphones varies greatly in different areas of the world. In areas in the USA, Europe, Australia, Japan, etc. (where most published results are from) errors are generally within 10 meters or so, but in more remote areas (such as on a remote island in Fiji), errors in the order of ±60 meters are not uncommon. Using two different mobile applications at sea level at one location resulted in reported elevations from −24 m to +58.9 m. These studies are preliminary and more research is needed in different areas of the world.

GPS-enabled digital cameras are like smartphones with respect to positional accuracy as they have similar sized in-board antennas. To conserve battery life, most GPS-enabled digital cameras have options to set positional update intervals. Depending on the camera, these can range from once every second to once every five minutes. The setting of this interval may have significant implications with respect to both coordinates and uncertainty.

Underwater digital cameras only update their position when the diver or snorkeler takes the camera above the surface long enough for the GPS to fix its position.

Using a large sample size (n>20,000) of GPS benchmarks in a variety of terrains in the United States, Wang et al. 2017 found that elevations in the Google Earth terrain model had a boundary of error interval at 95 per cent (BE95) of ±44 m, with worst-case scenarios around 200 m. The same study found that Google Earth terrain model had a BE95 of ±6 m along highways. Though we find no data for elsewhere in the world at this time, we recommend using the values extracted from the work of Wang et al. 2017 as estimates of elevational uncertainty when the source is the Google Earth terrain model. A second study using Google Earth to determine elevation in three regions of Egypt (El-Ashmawy 2016) on flat, medium, and steep terrains concluded that elevation data is more accurate in flat areas or areas with small height difference, with an accuracy of approximately 1.85 m (RMSE) and an error range of less than 3.72 m (and in some findings less than 1 m). Increasing the difference in height leads to decrease in the obtained accuracy with the RMSE rising to 5.69 m in steep terrain.

A locality consisting of an offset from a feature without a heading may arise as a result of an error when recording the locality in the field or through data entry. If the feature is small (such as a trig point) then the overall uncertainty will be largely due to the offset. With larger features (such as a town, or a lake), both the offset from, and the extent of the feature may contribute significantly to the overall uncertainty. The original collection catalogues or labels may contain information that can make the locality more specific. If not, a "Distance only" locality (e.g. "5 km from Lake Vättern, Sweden") might be envisioned as a band running around the reference feature at a distance given in the locality description. The problem is, we don’t know what was being used as the reference – some place in the lake, or some place on the edge — nor do we know if the offset was perpendicular to an edge or at some oblique angle to it. Because of these confounding factors, it is recommended to treat the locality as if it was a feature enlarged on all sides by the combination of all the sources of uncertainty (see Offset – Distance only in Georeferencing Quick Reference Guide (Zermoglio et al. 2020)).

A locality with a heading from a feature, but with no distance (e.g. "East of Albuquerque"), is particularly ambiguous and very subjective to georeference. With no additional information to constrain the distance , there is no clear indication of how far one might have to go to reach the location – to the next nearest feature; the next nearest feature of equivalent size, to a place where there is a major change in biome (such as a coast), or just keep going?

Note that seldom is such locality information given alone. For example, the locality may have administrative geography information (e.g. ‘East of Albuquerque, Bernalillo County, New Mexico’). This gives you a stopping point (e.g. the county border), and should allow you to georeference the locality (see Offset – Heading only in Georeferencing Quick Reference Guide (Zermoglio et al. 2020)). In any case, it is highly recommended not to record locality descriptions in this way.

A locality that contains an offset in a given direction to or from a feature is treated here as an "offset at a heading". There are several variations on such localities. One difficulty in determining a georeference for this type of locality description is knowing how the offset was determined – for example, by air, or along a path such as a road or river. Therefore, whenever a locality with an offset at a heading is described, be sure to be explicit about what is intended.

It is not uncommon for marine locality descriptions to use an azimuth – a heading toward a target feature, for example, "25° to Waipapa Point Lighthouse". In these cases the referenced feature is the starting point, and the heading from there should be 180 degrees on the compass away from the compass reading given in the locality description. This is known as a "back azimuth" or "backsighting".

Where the sense of the offset cannot be determined from the locality description or additional information and there is no obvious major path that can be followed in the rough direction and distance given, then it is best to assume the collector measured the distance by air. Whatever the decision, document the assumption in the georeference remarks (e.g. ‘Assumed "by air" – no roads E out of Yuma’, or ‘Assumed "by road" on Hwy. 80’) and georeference accordingly (see Offset – Distance at a Heading and Offset – Distance along a Path in Georeferencing Quick Reference Guide (Zermoglio et al. 2020)).

The addition of an adverbial modifier to the distance part of a locality description (such as "about 25 km"), while an honest observation, should not affect the determination of the geographic coordinates or the maximum uncertainty. Treat the uncertainty due to distance precision normally (see §3.4.6).

Sometimes it is convenient to describe a locality as a distance along a curvilinear feature — a path such as a road, river, trail, etc. (see Offset – Distance along a Path in Georeferencing Quick Reference Guide (Zermoglio et al. 2020)). One advantage of a description of this kind is that it avoids the uncertainty due to an imprecise heading. It might also be easy to register, such as when tracking distance with the odometer of a car while driving. However, a disadvantage is that it may not be quite as easy to determine the same location afterwards from maps alone during the georeferencing process. One reason is that you have to trace the facsimile of the path on a map. The map may have errors, loss of resolution due to map scale, inconsistencies with conditions at the time of the event, or might not even be present. There is also a difference between distance on the topographic surface and distance on a map, though for most normal situations (along roads and navigable waterways) the difference is <1% (see Offset – Distance along a Path in Georeferencing Quick Reference Guide (Zermoglio et al. 2020)). Worse, the path may have changed over time, making it even more difficult to find the exact locality retrospectively.

If the locality references a river, such as in the example "16 mi downstream from St Louis on the left bank of the Mississippi River", it is reasonable to assume that the offset is along the river. In this example, the locality is on the east side of the river, in Illinois, rather than on the west side, in Missouri, as the reference to "left bank" is conventionally taken to be in the orientation looking downstream.

This type of locality refers to rectilinear distances in two orthogonal directions from a feature, for example, "2 mi E and 1.5 mi N of Kandy" (see Offset – Distance along Orthogonal Directions in Georeferencing Quick Reference Guide (Zermoglio et al. 2020) and Figure 12). This way of describing a locality can be very effective, as it tends to remove one of the potentially largest sources of uncertainty, the ever-expanding uncertainty of direction with distance. Using orthogonal directions removes all directional uncertainty, as orthogonality implies directly in the orthogonal directions "by air". It is for this reason that this locality type is highly recommended for locality descriptions.

The depth of the benthic surface in large water bodies is called bathymetry or bathymetric depth. It is usually recorded in one of two ways – as a gridded surface (Digital Terrain Model), or as contours. The accuracy of the bathymetry depends on how it was determined, and is generally much more accurate near the coasts, or in harbors, than it is in the deeper ocean.

Since 2003, the most commonly used global coverage of bathymetry has been the One Minute General Bathymetric Chart of the Oceans (GEBCO 2019a), however, in 2019, a much finer, and more detailed, 15 arc-second grid coverage was released (GEBCO 2019b). The 3,732,480,000 grids (86,400 rows by 43,200 columns) cover from 89°59'52.5'' N, 179°59'52.5'' W to 89°59'52.5'' S, 179°59'52.5'' E, with elevation given for each pixel center. There are many criteria that determine the vertical accuracy of these grids, including the presence of steep canyons, water depth and turbidity (affects instrument penetration and acoustic beams get wider, the deeper they go), and methodology (satellite, single beam echo sounders (SES), multibeam echo sounders (MES), airborne laser (LADS), Light Detection and Ranging (LIDAR), etc.) (Wolf et al. 2019).

Bathymetric contours have generally only been available for harbors, coastal and near inshore areas, in some places extending to the edges of the continental slope. Where bathymetric contours (also called depth contours or isobaths) do exist, they are generally quite coarse (except in areas like the North Sea, and in harbors), and get wider apart as the depth increases. For example, the 2009 bathymetric contours for Australia are at 20 m, 40 m, 100 m, 200 m and 400 m. In some harbors, the contour interval is as small as one meter (Data.gov.au 2018). In 2019, the GEBCO_2019 global bathymetric contour dataset was derived from the GEBCO_2019 15 arc-second grid mentioned above. At large scales (1:5,000,000 and closer), the contour interval is 500 m; at medium scales (1:5,000,000 to 1:30,000,000) the contour interval is 1000 m; and at small scales (1:30,000,000 and greater), the contour interval is 2000 meters. Supplementary contours are shown in shallow waters (less than 500 m) (NCEI-NOAA 2019).

Very few studies have been carried out on the accuracy of either the bathymetric grids or contours – especially with GEBCO_2019 as the dataset has only recently been published. The authors have not been able to find any definitive information on accuracies that we can report on a general basis, but the contour intervals give an indication of the uncertainty inherent in the grids. In coastal, near inshore areas, harbors, and inland reservoirs and lakes, more intensive and different bathymetric surveys have generally been carried out (see the Bathymetric Data Viewer (NCEI 2020)) and accuracy studies have been conducted in some of these areas. In shallow-water areas there is less interference due to water depth and higher sound wave frequencies can be used for multibeam bathymetric surveying. The accuracy is much better than in other deeper-water areas, and thus these studies cannot be extrapolated to the broader ocean. For contours, as with land maps, uncertainty in the elevation is half the contour interval.

There are three methods for determining depth that are generally used by divers, i.e. dive computers, dive watches and depth gauges. All work on ambient pressure to determine the depth. Dive computers need to be calibrated before dives and set depending on the water density – i.e. saltwater or freshwater, etc. — and, if calibrated correctly, are reported by manufacturers to be accurate to within 0.3 meters.

A study of 47 brands of dive computers at depths of 10 m, 20 m, 30 m, 40 m and 50 m in both seawater and freshwater showed that the majority of depth estimates were in the ± 1 meter range, and that if the salinity is known and the instrument is properly calibrated, accuracies of around 1 per cent could or should be expected (Azzopardi & Sayer 2012). The accuracy of diver-held depth gauges are of a similar order. Dive watches are generally thought less accurate, but with reports for some watches of depth accuracy, at depths of up to 100 meters, as ± 1 per cent of displayed value + 0.3 meter (when used at constant temperature). Accuracy can be influenced by changes in ambient temperature and water salinity.

In cave systems and underground mines, determining the geographic position on the surface (known as ground zero) can be done with radiolocation or Electromagnetic Cave-to-Surface (ECMS) Mapping System (Sogade et al. 2004), which uses electromagnetic wave technology. This requires a levelled radio loop in the location within the cave and a receiver above ground to determine the location underground. The surface location can then be determined using a GPS/GNSS receiver, as usual. With a levelled antenna, an experienced operator can determine a ground zero with an accuracy of one meter for a 50 m depth (2%) (Gibson 1996, Gibson 2002), however, more recent radiolocation beacons have increased the horizontal accuracy to about 0.5 to 1 per cent (Goldsheider & Drew 2014, Buecher 2016). Fortunately, many caves and mines have already been extensively mapped, so where maps are available, these may be used to determine locations.

A second method, using the cave mouth, is probably more commonly used, is easier to determine, but is less accurate and has a much greater uncertainty. The cave mouth, tunnel opening, mine shaft entrance, etc., are the most obvious locations to begin with. These locations can easily be obtained using a GPS unit, but be aware of the likely reduced accuracy of the GPS unit if the cave entrance is within a deep valley where good GNSS reception may be reduced. It is documenting the location of the event from that position that is much more difficult, especially where detailed cave maps don’t exist. At its crudest level, one may estimate the cave extent and determine the corrected center of that extent. From there you can determine a geographic radial as noted elsewhere in this document (see §2.3.3). Just recording the location of the cave entrance, and using a large radius for the uncertainty is not ideal but may be a last resort. If doing this however, make sure that your locality description includes as much additional information as possible – such as estimated distance from the cave entrance, direction, and if possible, a ‘depth’. For georeferencing in Caves, see Feature – Cave in Georeferencing Quick Reference Guide (Zermoglio et al. 2020).

Traditionally, cavers have recorded the depth in a cave as the depth below the surface, however, in this document and for the purposes of recording biological observations, we use elevation (above mean sea level or geoid) for a position at the floor of the cave.

The distance below ground zero can be determined using the same radiolocation equipment as for determining the ground zero itself (see above). The accuracy of the distance below ground zero, calculated using these methods is around 5-10 per cent (Gibson 1996, Gibson 2002) for depths up to about 50 meters. As above, however, recent beacons have improved the accuracy to about 10 per cent for depths of up to 300 meters below the surface (NOT Engineers 2019). Uneven surface terrain can add to the uncertainties by up to a further 3 per cent and in very deep caves, mines, etc. where there are heavy ore bodies present, and where there are fault lines, this method is far less reliable for determining depth with errors increasing up to 20 per cent. In those conditions radiolocation may not be suitable for determining the distance below the surface.

From these figures, it is possible to determine the elevation of the floor of the cave by taking the elevation at ground zero and deducting the calculated distance below that point (see Figure 10). Note that when determining elevation in a cave, the accuracy mentioned above is additional to the elevation uncertainty determined for the elevation at ground zero.

Using detailed cave maps may provide a better (and cheaper) alternative to other methods, and you should choose the best method for your purpose, but be sure to document how the elevation was determined. Cave maps can usually be obtained by contacting local speleological or cave clubs.

The water depth within a subterranean water body (lake, river, sinkhole, etc.) is recorded as for other water bodies and is measured from the surface of the water body (see Figure 9B). The elevation of the surface of the water body is determined as for the floor of the cave in Figure 10.

Determining the distance above (and below) a surface (as documented elsewhere) is treated the same within a cave system (see Figure 9B, Figure 10). As above, the elevation of the cave floor has been determined, so a troglobiont (e.g. an animal) on the roof of the cave is given as meters above the floor of the cave whose elevation has been determined as above ("X" in Figure 10).

A workflow covering all the georeferencing activities can be a valuable instrument, not only for improving the efficiency of the whole georeferencing process, but also for incorporating checks and balances, and improving the quality of the resulting product. The type of workflow may be determined by the nature of the data, the way the original data are stored or documented, the nature of the desired end product, and even by the general preferences of those involved. In the following subsections we propose a generic workflow that covers all of the major aspects of georeferencing projects. Note that some of the steps presented might not apply to every project, and one must take into account priorities as discussed in §3.1. This first section, §3.1.1 outlines a recommended georeferencing project workflow in four phases. Subsequent sections deal with the details of some of the steps presented in the outline.

Based on an assessment of a variety of large-scale georeferencing projects that had efficiency as well as data quality in mind (e.g. §3.1.2), we recommend the following generic outline for a georeferencing project workflow using either the point-radius method or the shape method, or a mix of the two. This workflow can be used for projects that involve a single individual or a large collaboration, though some steps may apply more in one case than in another. Note that some of the actions included in different phases might happen simultaneously depending on the type and scale of the project.

One of the major contributions of the Mammal Networked Information System (MaNIS) project (Stein & Wieczorek 2004) was the design and implementation of a set of georeferencing guidelines (Wieczorek 2001) and online resources for a collaborative georeferencing workflow. The same basic workflow was implemented with great success for the sister projects HerpNET and the Ornithological Information System (ORNIS). Between the three projects, more than 1.2 million localities were georeferenced for 4.5 million vertebrate occurrence records. The basic workflow was more or less as follows:

It may be possible to use a look-up system that searches for similar localities that have already been georeferenced. For example, if you have a record with the locality "10 km NW of Campinas", you can search for all records with locality "Campinas" and see if any records that mean the same thing as "10 km NW of Campinas" have been georeferenced previously. Note that it is always worth verifying the georeference on a map — this can easily be done using software such as Google Maps, Google Earth, etc. Checking this way can reduce errors such as neglecting to add the minus (−) sign to a coordinate in the western or southern hemispheres.

An extension of this method could use the benefits of a distributed data system such as GBIF.org. A search could be conducted to see if the locality has already been georeferenced by another institution. At present, we quite often find that duplicates of occurrence records have been given significantly different georeferences by different institutions. Presumably this would not happen if best practices were followed, or if georeferencing is done by the original institution before distributing duplicates.

A preliminary study (Wieczorek pers. comm.) of roughly 33.1 million occurrences for 38.7 thousand plant taxa in GBIF from 15 April 2019 (GBIF 2019) showed that the records were associated with 7.2 million distinct locations, of which 25.7 per cent (30.9 per cent of occurrences) already had georeferences (i.e. decimalLatitude, decimalLongitude, geodeticDatum and coordinateUncertaintyInMeters). Of those without georeferences, exact matches (on geography plus locality fields, all turned into upper case) from other locations in GBIF could be found for 2.5 per cent of distinct locations (11.4 per cent of occurrences).

In the case where multiple possible georeferences are found using a lookup on previously existing georeferenced locations, the problem is knowing which of the several georeferences, if any, to choose.

If the georeference is not fully documented following best practices (including being reproducible), we recommend that existing georeferences not be used (or used only with extreme caution). Even if the georeference is documented, it should be checked visually on a map to be sure that it makes sense, just as for any new georeference.

Each institution will have needs for different resources in order to georeference their location data. The basics, however, include:

One of the most important preparation steps for efficient georeferencing is to have an effective way to handle the data. This section will help you decide if your data capture framework will need modification or not, and to what extent.

Some georeferencing projects (e.g. MaPSTeDI (Murphy et al. 2004)) used a separate working database for data entry operators so that the main data were not modified and day-to-day use of the database was not hindered. This also meant that the working database could be designed optimally for data entry, rather than trying to accommodate other database management and searching requirements. The data from the working database can be checked for quality, and then integrated into the main database from time to time. Such a way of operating is institution dependent, and may be worth considering.

What are the fields you need in your database to best store georeferencing information? This may seem obvious, but it is surprising how often a database is created and finalized before it is determined exactly what the database is supposed to hold. Be sure not to lump together dissimilar data into one field. Always atomize the data into separate fields with very specific definitions and rules for their content.

It is also of benefit to name the fields unambiguously, as users tend to go by the field names rather than looking at the field definitions. Thus, 'latitude_in_degrees' is a better name than 'latitude' for a field that is supposed to contain latitudes in decimal degrees, while 'verbatim_latitude' is better name for a field that is supposed to contain the latitude in the format given in the source. The names and definitions of fields in Darwin Core (Wieczorek et al. 2012b) were created specifically with this principle of clarity in mind. In order to take advantage of a community standard set of definitions, it is not a bad idea to use the term names from Darwin Core as field names in the database if the semantics of the two are the same.

Note, however, that the georeferencing results might benefit from additional fields that are not described in Darwin Core (e.g. 'feature_radial', 'radialUnits') in order to make it easier to reproduce the georeference and thus test its veracity. It is often tempting to include fields for the georeferenced coordinates and ignore any additional fields; however, you (or those who follow after you) are sure to regret this minimalist approach, because it severely limits the usability of the data. A Location occupies a physical extent, not just a point. The associated information on methods used to determine the georeference, the extent, radial, and uncertainty associated with the georeference are important pieces of information for the end user, as well as for managing and improving the quality of your information. The fields that are needed can be divided into two categories: the first consists of the fields associated with the textual description of the location, and the second consists of the fields associated with the spatially enabled interpretation as a georeference and the georeferencing process.

A reference worth checking before developing your own database system is the Herbarium Information Standards and Protocols for Interchange of Data (Conn 1999, Neish et al. 2007), which, although set up for data interchange between herbaria, is applicable to most data from natural history collections.

Many institutions separate locality descriptions into their component parts; feature name, distance, direction, etc., and store this information in separate fields in their databases. If this division of locality information is done, it is important not to replace the verbatim free-text locality field (the data as written on the label or in the field notebook), but instead add additional fields. This is because any transformation of data has the potential to lose information and to introduce errors, and the written format of the description may be the only original source available. The original information should never be overwritten or deleted.

Location-related fields to consider for georeferencing include all of the geography, locality, elevation, depth, and georeference terms in the Location class of Darwin Core (see location and §5.1) as well as the following fields that can have an influence on the georeference:

One of the key ways of making sure that data are standardized and accurate is to ensure, to the extent possible, that data are put in the correct field and that only data of an appropriate type can be put into each field by design. This is done by applying constraints on the data fields – for example, only allowing values between +90 and −90 in the field for decimal latitude. Many of the errors found when checking databases could have been easily avoided if the database had been set up correctly in the first place. The use of pick lists is essential where the field should contain only values from a restricted list of terms.

More complex constraints may also be possible. With ecological or survey data for example, one could set boundary limits between the starting locality and ending locality of a transect. For example, if your methodology always uses 1 km or shorter transects, then the database could include a boundary limit that flagged whenever an attempt was made to place these two points more than 1 km apart.

For more information on constraints, see various sections under §3.4.1.

Good user-friendly interfaces are essential to make georeferencing efficient and rapid, and to cut down on operator errors. The design should take into consideration the specific details of the georeferencing workflow, and optimize simultaneously for both overall efficiency, and consistency of the data entry process. This will improve accuracy and cut down on errors. The layout should be friendly, easy to use, and easy on the eyes. Where possible (and the software allows it) a number of different views of the data should be presented. These views can place emphasis on different aspects of the data and help data entry operator proficiency by allowing different ways of entering the data and by presenting a changing view for the operator.

In the same way, macros and scripts can help with automated and semi-automated procedures, reducing the need for tedious (and time-consuming) repetition. For example, if the data are being entered from a number of collections by one collector, taken at the same time from the same location, the information that is repeated from record to record should be able to be entered using just one or two keystrokes.

If maps are being used to assist in determining georeferences, a view that sorts the data geographically may also make the process more efficient by allowing the data operator to see all the records that may fall on one map sheet. Finally, it is also important to decide which fields the data entry operators should see when they are georeferencing. Fields such as date of collection, collector, specimen ID, taxonomy, habitat, and formation (for paleontological records) are very helpful for georeferencers to see along with the more obvious locality data.

Standard methodologies, in-house standards, and guidelines can help lead to consistency throughout the database and cut down on errors. A set of standards and guidelines should be established before any georeferencing begins (see §2.17). They should remain flexible enough to cater for new data and changes in processes over time, though careful thought beforehand can minimize the need for methodological changes, which might lead to inconsistencies where earlier efforts are lacking compared to those produced under newer protocols. Standards and guidelines in the following areas can improve the quality of the data and the efficiency of data entry:

Determining and documenting your institution’s own georeferencing best practice in manuals, for example that suit the circumstances of that institute (including language, local software and resources, etc.) can help maintain consistency as well as assist in training and data quality recording. As an example, see Escobar et al. 2015, where an internal document for the Alexander von Humboldt Institute in Colombia has been developed and put into practice. See also §2.17.

One of the greatest sources of georeferencing error is the data entry process. It is important that this process is made user-friendly and set up so that many errors cannot occur (e.g. through the use of pick lists, field constraints, etc.). The selection and training of data entry operators (see under §6.6) can make a big difference to the final quality of the georeferenced data. As mentioned earlier, the provision of good guidelines and standards can help in the training process and allow for data entry operators to reinforce their training over time.

Locality descriptions are often given in free text and encompass a wide range of content in a vast array of formats. An important part of the georeferencing process is to have a consistent way to interpret the text into spatial forms that can be operated on analytically. To do this, look for the parts of the description that can be interpreted independently, called locality clauses, each of which can be categorized into a locality type (see §3.2.2) that uses a specific set of rules to georeference (Wieczorek et al. 2004).

There is a lot of variation in the way clauses are written and the types of features they reference, but there are actually very few basic locality types, though these may have many variations depending on the feature type referenced. The Georeferencing Quick Reference Guide (Zermoglio et al. 2020) was written specifically to explain how to georeference all of the most common variations of locality types and feature types (Wieczorek et al. 2004):

A full locality description may contain multiple clauses. The goal of a georeference is to describe the location where all of the clauses are true simultaneously. In GIS terms, this would be the intersection of the shapes for all the clauses in the locality description. As humans, we would choose the clause that is most specific and georeference based on that, using the information from the other clauses to filter from among multiple possibilities. For example, a locality written as

should be georeferenced with a locality type "geographic feature only" with subtype Feature – with Obvious Spatial Extent as in Georeferencing Quick Reference Guide (Zermoglio et al. 2020) based on the bridge as the feature. Of course, the second clause helps us to determine which bridge (something we wouldn’t be able to do without that second clause), but beyond that the second clause contributes nothing to the boundaries of the feature, nor to the uncertainty in the final georeference.

If the more specific part of the locality cannot be unambiguously identified, then the next less specific part of the locality ("4 km N of Somerset" in the example above) should be georeferenced. In a case such as this, annotate in the georeference remarks with something like "unable to find the bridge georeferenced '4 km N of Somerset'".

Some locality descriptions give information about the nature of the offset (‘by road’, ‘by river’, ‘by air’, ‘up the valley’, etc.). Having this information simplifies the choice of offset-based locality type as §2.9.3 or §2.9.4.

It is sometimes possible to infer the nature of the offset path from additional supporting evidence in the locality description. For example, the locality

suggests a measurement by road since the final coordinates by that path are nearer to the lake than going 58 km NW in a straight line. At other times, you may have to consult detailed supplementary sources, such as field notes, collectors’ itineraries (see §3.2.4.4), diaries, or sequential collections made on the same day, to determine this information.

If any of the clauses in the locality description is classified as one of the three locality types, ‘dubious’, ‘cannot be located’, or ‘demonstrably inaccurate’, then the locality should not be georeferenced. Instead, an annotation should be made to the locality record giving the reason why it is not being georeferenced. See also Difficult Localities in Georeferencing Quick Reference Guide (Zermoglio et al. 2020).

Regardless of the method to be used (shape, bounding box, or point-radius), the georeferencing protocols for nearly every locality type begin with the identification of the features of reference in the locality description and the determination of the geographic boundaries of their extents. This is usually the most critical and time-consuming part of the protocols. It is best to use a visual reference to determine boundaries. If a feature name search on a visual source does not reveal the feature of interest, it is a good idea to use coordinates from a gazetteer to find the feature on a map, and then use the map to find the boundaries:

Use information from other clauses, such as administrative geography, information from other location fields such as elevation, and environmental information (e.g. terrestrial, freshwater aquatic, marine, taxon-specific) to constrain the extent as appropriate (see §3.2.4 and §3.1.6).

There are many ways that a location can be constrained beyond what the geography and locality descriptions alone suggest. Doing so relies on applying additional location information, such as elevation or depth, lithostratigraphic information for fossils, or information outside the location information, such as environmental constraints for a particular species. There are important implications about workflow and effort that need to be considered when applying additional constraints. For example, if taxon constraints are going to be applied, the georeferencing cannot be done strictly on location information, which means it has to be done on occurrence records, or on an index combining location and taxon. This would be much slower than georeferencing based on location alone. A good compromise would be to georeference in multiple stages, with the first stage based on location information, and a subsequent stage including the rest of the occurrence information, and perhaps a final stage of review by collectors to be able to set dwc:georeferenceVerificationStatus to "verified by collector" – the best status a georeference can possibly have.

Based on the aspirations for georeferencing methods described in the previous paragraph, the point method, consisting of only coordinates, or coordinates in a coordinate reference system, is insufficient to be useful except to center a point on a map (and even that potentially incorrectly without the coordinate reference system). The point method does not give any indication of scale, though the mistake is often made to try to represent scale and/or uncertainties in the precision of the coordinates. For these reasons, the point method is NOT recommended as the end product of a georeferencing workflow.

The result of the point-radius method (Wieczorek et al. 2004) is a geographic coordinate (the "corrected center"), its geodetic datum, and a maximum uncertainty distance as a radius. The length of the radius must be large enough so that a circle centered on the corrected center and based on that radius encompasses all of the uncertainties in the interpretation of the location. The point-radius is a very simple representation of the location that contains all of the places that the locality description might refer to, but may also circumscribe areas that do not match the locality description. That’s OK. The point-radius circle can also be intersected with other spatially enabled information to constrain the effective area within the circle, such as elevation, to derive a shape representation of the locality. For example, calculate the intersection of a point-radius circle with the shape of the matching elevation contours in a geographic information system to get a shape that better matches the described locality. Similarly, one could calculate the intersection of an exposed geological formation with a point-radius georeference to refine the latter into a shape. The detailed recommended protocols for georeferencing using the point-radius method are given in the Georeferencing Quick Reference Guide (Zermoglio et al. 2020).

The result of the bounding box method (Wieczorek et al. 2004) is a set of two coordinates, one for each of two corners diagonally opposed on the bounding box along with their coordinate reference system. The corners define the minimum and maximum values of the coordinates, within which the whole of the location and its uncertainties is contained. Like the point-radius method, the bounding box method results in a very simple representation of the location that contains all of the places that the locality description might refer to, but may also contain areas that do not match the locality description.

Unlike the point-radius method, this method has no scalar maximum uncertainty distance to be able to easily understand or filter on the size of the enclosed region, though one can be calculated using half the distance between the two corners as given by Vincenty’s formulae (Vincenty 1975, Vincenty 1976). Thus, a bounding box georeference can be turned into a point-radius georeference by using the distance just described as the geographic radial, and from that finding the corrected center, which will not be equal to the geographic center of the bounding box, except where the bounding box spans equal distances north and south of the equator or is based on a metric grid.

A point-radius georeference can be turned into a bounding box georeference by using the geographic radial from the corrected center of the point-radius to determine the coordinates of the east-west and north-south extremes of the bounding box.

Like the point-radius circle, the bounding box can also be intersected with other spatially enabled information to constrain the effective area within.

The shape method (also called the polygon method by some (Yost 2015)) of determining uncertainty is a conceptually simple method that delineates a locality using geometries with one or more polygons, buffered points, or buffered polylines. A combination of these shapes can represent a town, park, river, junction, or any other feature or combination of features found on a map. While simple to describe, the task of generating these shapes must account for all the uncertainties, and that can be difficult. Except for the simplest locality types, creating shapes is impractical without the aid of digital maps, GIS software (for buffering, clipping, etc.), and expertise, all of which can be relatively expensive. Also, except for a bounding box, which is an extremely simple example, storing a shape in a database can be considerably more complicated than storing a single pair of coordinates with a scalar uncertainty distance as in the point-radius method. Darwin Core (Wieczorek et al. 2012b) offers the field dwc:footprintWKT, in which a geometry can be stored in the Well-Known Text format (ISO 2016) accompanied by the coordinate reference system in the field dwc:footprintSRS. Particular challenges to making this method practical for georeferencing natural history collections data include assembling freely accessible digital cartographic resources and developing tools for automation of the georeferencing process (Yost n.d.). This is because, not only does the geometry of the feature usually need to be created (unless it is an administrative boundary or other shape available in a spatial data layer), but also all the points in the feature geometry have to be used in combination with the uncertainties to arrive at a final shape that includes the location with its uncertainties and nothing more. Note that GEOLocate (Rios 2019) does produce an "error polygon" (Biedron & Famoso 2016) in addition to a point-radius, but how this is done is not documented in detail.

Of all the methods discussed in this document, the shape method has the potential to generate the most specific digital spatial descriptions of localities, leaving out areas that are not viable as part of the location. A point-radius can be easily derived from a final shape by using the corrected center for the coordinates and the geographic radial of the georeference (not just the feature) for the maximum uncertainty distance. See Figure 15 for one example of where a point-radius may be refined by using the shape method. See also §2.3.3.

Other shape-based methods have been proposed that use probabilistic approaches (Guo et al. 2008, Liu et al. 2009). Since these methods are even more difficult than the shape method, and there are currently no tools available to take advantage of these methods, we do not discuss them further in this document.

The first step in determining the coordinates for a locality description is to identify the most specific feature within the locality description. Coordinates may be retrieved from gazetteers, geographic name databases, maps, or from other locality descriptions that have coordinates or shapes. We use the term ‘feature’ to refer to not only traditional named places, but also to places that may not have proper names, such as road junctions, stream confluences, highway mile pegs, and cells in grid systems (e.g. Quarter Degree Square Cells, see [Quarter Degree Squares]). The source and precision of the coordinates should be recorded so that the validity of the georeferenced locality can be checked. The original coordinate system and the geodetic datum should also be recorded. This information helps to determine sources and the maximum uncertainty distance, especially with respect to the original coordinate precision.

How do we take into account the uncertainty due to the shape of the feature? The method that results in the least uncertainty is to find the smallest enclosing circle (Matoušek et al. 1996) that contains all of the points on the geographic boundary of the feature. If the center of the circle does not fall on or within the boundary of the feature, choose the point nearest to the center that is on the boundary. This is known as the corrected center. The distance from the corrected center to the farthest point on the geographic boundary of the feature is called the geographic radial. The geographic radial is the uncertainty due to the extent of the feature (see Figure 4).

Every feature occupies a finite space, or ‘extent’. The extents of features are an important source of uncertainty. Points of reference for features may change over time – post offices and courthouses are relocated, towns change in size, the courses of rivers change, etc. Moreover, there is no guarantee that the person who recorded the locality information paid attention to any specific convention when reporting a locality as an offset from a feature. For example,

may have been measured from the post office, the civic plaza, or from the bus station on the eastern side of the heavily populated part of town, or anywhere else in Bariloche, which is actually quite large. When calculating an offset, we generally have no way of knowing where the person who recorded the locality started to measure the distance. The determination of the boundaries of a feature are discussed in §3.2.3.

It is also worth noting that the extent of a feature may have changed over time, so the date of the recording may also be important when calculating an extent and thus the geographic radial. In many cases (especially for populated places), the current extent of a feature will be greater than its historical extent and the uncertainty may be somewhat overestimated if current maps are used.

If the locality described is an irregular shape (e.g. a winding road or river), there are two ways of calculating the "center" coordinates and determining the radial. The first is to measure along the vector (line) and determine the midpoint as the location of the feature. This is not always easy, so the second method is to determine the geographic center (i.e. the midpoint of the extremes of latitude and longitude) of the feature. This method describes a point where the uncertainty due to the extent of the feature is minimized (what we are calling the corrected center). The radial is then determined as the distance from the determined position to the furthest point at the extremes of the vector. If the geographic center of the shape is used and it does not lie within the locality described (e.g. the geographic center of a segment of a river does not actually lie on the river), then the point nearest the geographic center that lies within the shape (corrected center) is the preferred reference for the feature and represents the point from which the geographic radial should be calculated (see Figure 4).

When documenting the georeferencing process, it is recommended that the feature, its extent, radial, and the source of the information (including its date) all be recorded. For details on georeferencing, see Geographic Feature Only in Georeferencing Quick Reference Guide (Zermoglio et al. 2020).

Geographic coordinates can be expressed in a number of different coordinate formats. Decimal degrees provide the most convenient coordinates to use for georeferencing for no more profound reason than a locality can be described with only four attributes – decimal latitude, decimal longitude, datum, and uncertainty (Wieczorek 2001).

There are many ways of finding coordinates for a location, including using a gazetteer, a GPS, aerial photogrammetry, digital maps, or paper maps of many different types, and scales.

Geographic coordinates should always be recorded using as many digits as possible; the precision of the coordinates should be captured separately from the coordinates themselves, preferably as a distance, which conserves its meaning regardless of location and coordinate transformations. Recording coordinates with insufficient precision can result in unnecessary uncertainties. The magnitude of the uncertainty is a function of not only the precision with which the data are recorded, but also of the datum and the coordinates themselves. This is a direct result of the fact that a degree does not correspond to the same distance everywhere on the surface of the earth.

Table 3 shows examples of the contributions to uncertainty for different levels of precision in coordinates using the WGS84 reference ellipsoid. Calculations are based on the same degree of imprecision in both coordinates and are given for several different latitudes. Approximate calculations can be made based on this table, however, more accurate calculations can be obtained using the Georeferencing Calculator (Wieczorek & Wieczorek 2020) – see further discussion below.

From Table 3, it can be seen that an observation recorded in degrees, minutes, and seconds (DMS) has a minimum uncertainty of between 32 and 44 metres.

It is important to record the datum used for the coordinate source (GPS, map sheet, gazetteer) if it is known, or to record the fact that it is not known. Coordinates without a coordinate reference system are ambiguous. Geographic coordinates with a datum constitute a coordinate reference system (see §2.5), but seldom do natural history collections have complete coordinate reference system information. Even with a GPS being used to record coordinates in the field, the geodetic datum is typically ignored.

The ambiguity from a missing datum varies geographically and adds greatly to the error inherent in the georeferencing. Differences between datums may cause an error in true location from a few centimeters up to kilometers (Wieczorek 2019). Note that the difference between datums is not a simple function that can be calculated on the fly. The values have to be pre-calculated comparing all datums to a reference datum of choice (e.g. WGS84) at every point of interest over the earth’s surface and stored in a way that can be looked up by geographic coordinates. The Georeferencing Calculator (Wieczorek & Wieczorek 2020) is capable of doing such a lookup (see §3.4.9). In the absence of looking up the actual value by coordinates, the worst case scenario of 5359 m (Wieczorek 2019) can be used.

The calculation of uncertainty from the precision in which a direction is recorded depends on the distance from the starting reference feature. The uncertainty will increase with increasing distance from the source. For simple determinations of angular precision due to direction – see Table 4.

Using the example

and if we ignore distance imprecision, uncertainty due to the direction imprecision (Figure 11) is encompassed by an arc centered 10 km (d) from the center of Lambert Centre (at x,y) at a heading of 45 degrees (θ), extending 22.5 degrees in either direction from that point. At this scale the distance (e) from the center of the arc to the furthest extent of the arc (at x′,y′) at a heading of 22.5 degrees (θ′) from the center of Lambert Centre can be approximated by the Pythagorean Theorem,

where x=dcos(θ), y=dsin(θ), x′=dcos(θ′), and y′=dsin(θ′). The uncertainty in the above example would be 3.90 km.

This shows just one simple example. For details and formulae for calculating more complicated uncertainties, see Wieczorek 2001 and Wieczorek et al. 2004. Because of the complicated nature of these calculations, it is best to use the Georeferencing Calculator (Wieczorek & Wieczorek 2020) – see §3.4.9.

Precision can be difficult to gauge from a locality description as it is seldom, if ever, explicitly recorded. Further, a database record may not reflect, or may reflect incorrectly, the precision inherent in the original measurements, especially if the locality description in the database has undergone normalization, reformatting, or secondary interpretation of the original locality description.

There are a number of ways of calculating uncertainty from distances. In this document, we recommend a conservative approach, which assumes that many records have undergone a certain amount of interpretation or transformation when being entered into the database. Thus, a record of "10¼ mi" may be entered into the database as 10.25 mi. The precision implied in the value 10.25 is thus a false precision and the real precision should not be assumed to be between 10.24 and 10.26 or between 10.2 and 10.3. The method of Wieczorek et al. 2004, adapted here, bases the estimate of uncertainty on the fractional part of the distance, calculated by dividing 1 by the fractional denominator. The uncertainty would just be half of the precision. For example, 10.5 mi N of Bakersfield could reasonably be expected to mean 10½ mi with a precision of half a mile between 10.25 and 10.75 mi, or 10.5 with an uncertainty of 0.25 mi.

For distance measurements that are positive integer powers of 10, the precision should be ten to the next lower power. This calculation differs from Wieczorek et al. 2004, which recommended that the precision should be based on ten to the same power. Upon reconsideration, that seems excessive (see Table 5). This same reasoning can be used for precision in verbatim elevations and depths. Recommended values for uncertainty related to offset precisions are shown in Table 5.

Precision can also be masked or lost when measurements are converted, such as from feet to meters, or from miles to kilometers.

Figure 12 shows an example of two orthogonal distances measured from a feature, each with the uncertainty due to distance precision. If we ignore all sources of uncertainty except those arising from distance precision, the uncertainty is a bounding box centered on the point 8 km E and 6 km N of the corrected center of the feature. Each of the distance measurements demonstrates a precision of 1 km. Thus, each side of the box is a total of 1 km in length (0.5 km uncertainty in each cardinal direction from the center). Since we are characterizing the precision as a single distance measurement (1 km), we need the circle that circumscribes the above-mentioned bounding box to get the uncertainty due to the combined distance precisions. The radius of this circle is half the length of the distance precision bounding box, which is equal to one half the square root of two times the distance precision. So, for the above example the uncertainty associated with only the distance precision is one half the square root of two, or 0.707 km.

When combining uncertainties from different sources, it is not as simple as taking the average or adding them together. Uncertainties inherent in the location of the feature, in its extent, in the direction of the offset, and the distance of the offset, are just four sources that need to be combined to get an overall uncertainty. A detailed discussion of the calculations involved can be found in Wieczorek 2001 and Wieczorek et al. 2004. For a practical way of calculating uncertainties in locality descriptions, we recommend the Georeferencing Calculator (Wieczorek & Wieczorek 2020). To understand how each source of uncertainty contributes to the net overall uncertainty, see Understanding Uncertainty Contributions in the Georeferencing Calculator Manual (Bloom et al. 2020).

The Georeferencing Quick Reference Guide (Zermoglio et al. 2020) is a practical guide for georeferencing giving step-by-step instructions on how to georeference a wide variety of locality types (see §3.2) following the best practices in this document and with specific reference on what to enter into the Georeferencing Calculator (Wieczorek & Wieczorek 2020).

The Georeferencing Calculator (Wieczorek & Wieczorek 2020) (Figure 13) is a tool to aid in georeferencing descriptive localities such as those found in museum-based natural history collections. It was originally designed for the Mammal Networked Information System (MaNIS) Project and has since been adopted by many other georeferencing initiatives. The current version and its Georeferencing Calculator Manual (Bloom et al. 2020) have been extensively upgraded to include new features and to bring it in line with this document.

The application makes calculations adapted from the methods originally described in the Georeferencing Guidelines (Wieczorek 2001) and later formalized in a peer-reviewed publication (Wieczorek 2004). We recommend its use generally by all natural history institutions to calculate uncertainty in location data without the need for a detailed understanding of the complicated underlying algorithms. The more institutions that use this one method, the more consistent will be the quality of data across and between institutions, making it easier for users to evaluate the quality of the data. We recommend reading both of the above-mentioned publications and the Georeferencing Calculator Manual (Bloom et al. 2020) for an understanding of the calculations involved and an understanding of how the Calculator works.

The Calculator can work online or locally in a browser (latest release available on GitHub). The source code is freely and openly available on GitHub.

One of the major sources of error in georeferencing is at the stage of data entry. Errors can be reduced by the establishment of good data entry procedures – use of pick lists, field constraints, etc. Many of these issues should also be addressed as part of the database design (see §3.1.1.4). Good database design can reduce many of the errors associated with data entry. However, once these are in place and working, then regular checks need to be carried out on the data entry operators and on the process of data entry.

One method developed for the MaPSTeDI project (Murphy et al. 2004) is to first check the accuracy of the georeferencing. This process involves checking a certain number of each georeferencer’s records. Based on various trials, it is recommended that the first 200 records that a new georeferencer completes be checked for accuracy. Not only is this initial checking beneficial to the accuracy of the data, but also it is essential to allow the georeferencer to improve and learn through feedback from making mistakes. We recommend the following protocol for checking data quality:

The second purpose of quality checking is to allow georeferencers to refer difficult or confusing records to the quality checker for help or advice. The quality checker will then resolve these ‘problem records’ as well as possible. Checking problem records can be like detective work. Historical records often have locality descriptions with features that do not appear on modern maps or in gazetteers. To find these localities, it is often necessary to consult several different sources of information. These sources include, but are not limited to catalogue books, field notes, other records with similar localities, other collections, scientific and other publications, websites, online databases, speciality gazetteers, and historical maps. Bits of information from several places can often be used to establish the correct coordinates for a historical locality.

In addition, some problem records do not make sense because of contradictions or missing or garbled information. These problem records may be the result of mistakes in data entry made in either the paper catalogue or the database. It may also be necessary to consult the curatorial staff or even the original collector. If georeferencing information is able to be found for difficult locations, then it is worthwhile documenting them for future use, or even publishing the results of your searches, as seldom are unusual localities orphans and you, or others, may come across the same locality again at a later date.

Data validation (checking for errors) can be a time-consuming process; however, it is one of the most important processes you can carry out with your data. It is not practical to check every record individually, so the use of batch processing techniques and outlier detection procedures, etc., is essential. Fortunately, a number of these have been developed and are available in software products or online (see georeferencing.org and Chapman 2005b). The information in those resources are not repeated here. We recommend that you incorporate some of those methodologies into your own working practices.

There are many methods of checking for errors in georeferenced data. These can involve:

Some of these techniques are incorporated into a number of programs including Biogeo (Robertson et al. 2016), CoordinateCleaner (Zizka et al. 2019) and the stand-alone GIS software DIVA-GIS (Hijmans et al. 2012). See also georeferencing.org.

The Data Quality Interest Group of TDWG established a Task Group in 2014 to develop a set of Core Tests and Assertions for checking and validating the quality of species occurrence data (specimens and observation, etc.). The resulting 101 tests based on Darwin Core terms will be coded and available for use in 2020 (Chapman et al. 2020). Currently (as of February 2020), there are 12 validation tests related to coordinates and datums and another seven related to geography. In addition there are seven amendment tests that can be used to improve the quality of the data.

When making corrections to your database, we strongly recommend that you always add and never replace or delete. For this to happen you will usually require additional fields in the database. For example, you may have ‘original’ or ‘verbatim’ georeference fields in addition to the main georeference fields. Additionally, the database may require a number of ‘Remarks/Notes/Comments’ fields. Fields that can be valuable are those that describe validation checking that has been carried out – even (and often especially) if that checking has led to confirmation of the georeference. These fields may include information on what checks were carried out, by whom, when and with what results. Be sure to update the equivalent of dwc:georeferenceVerificationStatus and associated fields (dwc:georeferencedBy, dwc:georeferencedDate) whenever changes are made to the georeference.

‘Truth in Labelling’ is an important consideration with respect to documenting data quality. This is especially so where data are being made available to a wider audience, for example, through GBIF. We recommend that documentation of the data and their quality be upfront and honest. Error is an inescapable characteristic of any dataset, and it should be recognized as a fundamental attribute of those data. All databases have errors, and it is in no one’s interest to hide those errors (Chrisman 1991). On the contrary, revealing data actually exposes them to editing, validation and correction through user feedback, while hiding information almost guarantees that it will remain dirty and of little long-term value.