Skip to contents

3. Translation of OSM to Simple Features

Mark Padgham

2023-03-27

Source: vignettes/osm-sf-translation.Rmd
osm-sf-translation.Rmd

1. OpenStreetMap Data Structure

OpenStreetMap (OSM) data has a unique structure that is not directly reconcilable with other modes of representing spatial data, notably including the widely-adopted Simple Features (SF) scheme of the Open Geospatial Consortium (OGC). The three primary spatial objects of OSM are:

  1. nodes, which are directly translatable to spatial points

  2. ways, which may be closed, in which case they form polygons, or unclosed, in which case they are (non-polygonal) lines.

  3. relations which are higher-level objects used to specify relationships between collections of ways and nodes. While there are several recognised categories of relations, in spatial terms these may be reduced to a binary distinction between:

    • multipolygon relations, which specify relationships between an exterior polygon (through designating role='outer') and possible inner polygons (role='inner'). These may or may not be designated with type=multipolygon. Political boundaries, for example, often have type=boundary rather than explicit type-multipolygon. osmdata identifies multipolygons as those relation objects having at least one member with role=outer or role=inner.

    • In the absence of inner and outer roles, an OSM relation is assumed to be non-polygonal, and to instead form a collection of non-enclosing lines.

2. Simple Features Data Structure

The representation of spatial objects as Simple Features is described at length by the OGC, with this document merely reviewing relevant aspects. The SF system assumes that spatial features can be represented in one of seven distinct primary classes, which by convention are referred to in all capital letters. Relevant classes for OSM data are:

  1. POINT
  2. MULTIPOINT
  3. LINESTRING
  4. MULTILINESTRING
  5. POLYGON
  6. MULTIPOLYGON

(The seventh primary class is GEOMETRYCOLLECTION, which contains several objects with different geometries.) An SF (where that acronym may connote both singular and plural) consists of a sequence of spatial coordinates, which for OSM data are only ever XY coordinates represented as strings enclosed within brackets. In addition to coordinate data and associated coordinate reference systems, an SF may include any number of additional data which quantify or qualify the feature of interest. In the sf extension to R, for example, a single SF is represented by one row of a data.frame, with the geometry stored in a single column, and any number of other columns containing these additional data.

Simple Feature geometries are referred to in this vignette using all capital letters (such as POLYGON), while OSM geometries use lower case (such as polygon). Similarly, the Simple Features standard of the OGC is referred to as SF, while the R package of the same name is referred to as R::sf–upper case R followed by lower case sf. Much functionality of R::sf is determined by the underlying Geospatial Data Abstraction Library (GDAL; described below). Representations of data are often discussed here with reference to GDAL/sf, in which case it may always be assumed that the translation and representation of data are determined by GDAL and not directly by the creators of R::sf.

3. How osmdata translates OSM into Simple Features

3.1. OSM Nodes

OSM nodes translate directly into SF::POINT objects, with all OSM key-value pairs stored in additional data.frame columns.

3.2. OSM Ways

OSM ways may be either polygons or (non-polygonal) lines. osmdata translates these into SF::LINESTRING and SF::POLYGON objects, respectively. Although polygonal and non-polygonal ways may have systematically different key fields, they are conflated here to the single set of key values common to all way objects regardless of shape. This enables direct comparison and uniform operation on both SF::LINESTRING and SF::POLYGON objects.

3.3 OSM Relations

OSM relations comprising members with role=outer or role=inner are translated into SF::MULTIPOLYGON objects; otherwise they form SF::MULTILINESTRING objects. As in the preceding case of OSM ways, potentially systematic differences between OSM key fields for multipolygon and other relation objects are ignored in favour of returning identical key fields in both cases, whether or not value fields for those keys exist.

3.3(a) Multipolygon Relations

An OSM multipolygon is translated by osmdata into a single SF::MULTIPOLYGON object which has an additional column specifying num_members. The SF geometry thus consists of a list (an R::List object) of this number of polygons, the first of which is the outer polygon, with all subsequent members forming closed inner rings (either individually or in combination).

Each of these inner polygons are also represented as one or more OSM objects, which will generally include detailed data on the individual components not able to be represented in the single multipolygon representation. Each inner polygon is therefore additionally stored in the sf::MULTIPOLYGON data.frame along with all associated data. Thus the row containing a multipolygon of num_polygon polygons is followed by num_polygon - 1 rows containing the data for each inner polygon.

Note that OSM relation objects generally have fewer (or different) key-value pairs than do OSM way objects. In the OSM system, data describing the detailed properties of the constituent ways of a given OSM relation are stored with those ways rather than with the relation. osmdata follows this general principle, and stored the geometry of all ways of a relation with the relation itself (that is, as part of the MULTIPOLYGON or MULTILINESTRING object), while those ways are also stored themselves as LINESTRING (or potentially POLYGON) objects, from where their additional key-value data may be accessed.

3.3(b) Multilinestring Relations

OSM relations that are not multipolygons are translated into SF::MULTILINESTRING objects. Each member of any OSM relation is attributed a role, which may be empty. osmdata collates all ways within a relation according to their role attributes. Thus, unlike multipolygon relations which are always translated into a single sf::MULTIPOLYGON object, multilinestring relations are translated by omsdata into potentially several sf::MULTILINESTRING objects, one for each unique role.

This is particularly useful because relations are often used to designated extended highways (for example, designated bicycle routes or motorways), yet these often exist in primary and alternative forms, with these categories specified in roles. Separating these roles enables ready access to any desired role.

These multilinestring objects also have a column specifying num_members, as for multipolygons, with the primary member followed by num_members rows, one for each member of the multilinestring.

4. GDAL Translation of OSM into Simple Features

The R package sf provide an R implementation of Spatial Features, and provides a wrapper around GDAL for reading geospatial data. GDAL provides a ‘driver’ to read OSM data, and thus sf can also be used to read OSM data in R, as detailed in the main osmdata vignette. However, the GDAL translation of OSM data differs in several important ways from the osmdata translation.

The primary difference is that GDAL only returns unique objects of each spatial (SF) type. Thus sf::POINT objects consist of only those points that are not otherwise members of some ‘higher’ object (line, polygon, or relation objects). Although a given set of OSM data may actually contain a great many points, attempting to load these with

sf::st_read (file, layer = 'points')

will generally return surprisingly few points.

4.1. OSM Nodes

Apart from the numerical difference arising through osmdata returning an sf::POINTS structure containing all nodes within a given set of OSM data, while sf::st_read (file, layer='points') returns only those points not represented in other structure, the representation of points remains otherwise broadly similar. The only other major difference is that osmdata retains all key-value pairs present in a given set of OSM data, whereas GDAL/sf only retains a select few of these. Moreover, the keys returned by GDAL/sf are pre-defined and invariant, meaning that data returned from sf::st_read (...) may often contain key columns in the resultant data.frame which contain no (non-NA) data. This difference is illustrated in an example repeated here from the
main osmdata vignette, with the same principles applying to all of the following classes of OSM data.

The following three lines define a query and download the resultant data to an XML file.

q <- opq (bbox = 'Trentham, Australia')
q <- add_osm_feature (q, key = 'name') # any named objects
osmdata_xml (q, 'trentham.osm')

These data may then be converted into SF representations using either R::sf or osmdata, with OSM keys being the column names of the resultant data.frame objects.

names (sf::st_read ('trentham.osm', layer = 'points', quiet = TRUE))
##  [1] "osm_id"     "name"       "barrier"    "highway"    "ref"       
##  [6] "address"    "is_in"      "place"      "man_made"   "other_tags"
## [11] "geometry"
names (osmdata_sf (q, 'trentham.osm')$osm_points)
##  [1] "osm_id"           "name"             "X_description_"   "X_waypoint_"     
##  [5] "addr.city"        "addr.housenumber" "addr.postcode"    "addr.street"     
##  [9] "amenity"          "barrier"          "denomination"     "foot"            
## [13] "ford"             "highway"          "leisure"          "note_1"          
## [17] "phone"            "place"            "railway"          "railway.historic"
## [21] "ref"              "religion"         "shop"             "source"          
## [25] "tourism"          "waterway"         "geometry"

osmdata returns far more key fields than does GDAL/sf. More importantly, however, GDAL/sf returns pre-defined key fields regardless of whether they contain any data:

addr <- sf::st_read ('trentham.osm', layer = 'points', quiet = TRUE)$address
all (is.na (addr))
## [1] TRUE

In contrast, osmdata returns only those key fields which contain data (and so excludes address in the above example).

4.2. OSM Ways

As for points, GDAL/sf only returns those ways that are not represented or contained in ‘higher’ objects (OSM relations interpreted as SF::MULTIPOLYGON or SF::MULTILINESTRING objects). osmdata returns all ways, and thus enables, for example, examination of the full attributes of any member of a multigeometry object. This is not possible with the GDAL/sf translation. As for points, the only additional difference between osmdata and GDAL/sf is that osmdata retains all key-value pairs, whereas GDAL retains only a select few.

4.3 OSM Relations

Translation of OSM relations into Simple Features differs more significantly between osmdata and GDAL/sf.

4.3(a) Multipolygon Relations

As indicated above, multipolygon relations are translated in broadly comparable ways by both osmdata and sf/GDAL. Note, however, the way members of an OSM relation may be specified in arbitrary order, and the multipolygonal way may not necessarily be traced through simply following the segments in the order returned by sf/GDAL.

4.3(b) Multilinestring Relations

Linestring relations are simply read by GDAL directly in terms of the their constituent ways, resulting in a single SF::MULTILINESTRING object that contains exactly the same number of lines as the ways in the OSM relation, regardless of their role attributes. Note that roles are frequently used to specify alternative multi-way routes through a single OSM relation. Such distinctions between primary and alternative are erased with GDAL/sf reading.

5 Examples

5.1 Routing

Navigable paths, routes, and ways are all tagged within OSM as highway, readily enabling an overpass query to return only ways that can be used for routing purposes. Routes are nevertheless commonly assembled within OSM relations, particularly where they form major, designated transport ways such as long-distance foot or bicycle paths or major motorways.

5.1(a) Routing with sf/GDAL

A query for key=highway translated through GDAL/sf will return those ways not part of any ‘higher’ structure as SF::LINESTRING objects, but components of an entire transport network might also be returned as:

  1. SF::MULTIPOLYGON objects, holding all single ways which form simple polygons (that is, in which start and end points are the same);
  2. SF::MULTIPOLYGON objects holding all single (non-polygonal) ways which combine to form an OSM multipolygon relation (that is, in which the collection of ways ultimately forms a closed role=outer polygon).
  3. SF::MULTILINESTRING objects holding all single (non-polygonal) ways which combine to form an OSM relation that is not a multipolygon.

Translating these data into a single form usable for routing purposes is not simple. A particular problem that is extremely difficult to resolve is reconciling the SF::MULTIPOLYGON objects with the geometry of the SF::LINESTRING objects. Highway components contained in SF::MULTIPOLYGON objects need to be re-connected with the network represented by the SF::LINESTRING objects, yet the OSM identifiers of the MULTIPOLYGON components are removed by sf/GDAL, preventing these components from being directly re-connected. The only way to ensure connection would be to re-connect those geographic points sharing identical coordinates. This would require code too long and complicated to be worthwhile demonstrating here.

5.1(b) Routing with osmdata

osmdata retains all of the underlying ways of ‘higher’ structures (SF::MULTIPOLYGON or SF::MULTILINESTRING objects) as SF::LINESTRING or SF::POLYGON objects. The geometries of the latter objects duplicate those of the ‘higher’ relations, yet contain additional key-value pairs corresponding to each way. Most importantly, the OSM ID values for all members of a relation are stored within that relation, readily enabling the individual ways (LINESTRING or POLYGON objects) to be identified from the relation (MULTIPOLYGON or MULTILINESTRING object).

The osmdata translation thus readily enables a singularly complete network to be reconstructed by simply combining the SF::LINESTRING layer with the SF::POLYGON layer. These layers will always contain entirely independent members, and so will always be able to be directly combined without duplicating any objects.