7. Spherical geometry in sf using s2geometry

Edzer Pebesma and Dewey Dunnington

For a better version of the sf vignettes see https://r-spatial.github.io/sf/articles/


This vignette describes what spherical geometry implies, and how package sf uses the s2geometry library (https://s2geometry.io) for geometrical measures, predicates and transformations. The code in this vignette only run if s2 has been installed, e.g. with


and is available. After sf has been attached, you can check if s2 is available as follows:

s2_available = !inherits(try(sf_use_s2(TRUE), silent=TRUE), "try-error")
## [1] TRUE

Attach the s2 package as follows:


Most of the package’s functions start with s2_ in the same way that most sf function names start with st_.

Projected and geographic coordinates

Spatial coordinates either refer to projected (or Cartesian) coordinates, meaning that they are associated to points on a flat space, or to unprojected or geographic coordinates, when they refer to angles (latitude, longitude) pointing to locations on a sphere (or ellipsoid). The flat space is also referred to as \(R^2\), the sphere as \(S^2\)

Package sf implements simple features, a standard for point, line, and polygon geometries where geometries are built from points (nodes) connected by straight lines (edges). The simple feature standard does not say much about its suitability for dealing with geographic coordinates, but the topological relational system it builds upon (DE9-IM) refer to \(R^2\), the two-dimensional flat space.

Yet, more and more data are routinely served or exchanged using geographic coordinates. Using software that assumes an \(R^2\), flat space may work for some problems, and although sf up to version 0.9-x had some functions in place for spherical/ellipsoidal computations (from package lwgeom, for computing area, length, distance, and for segmentizing), it has also happily warned the user that it is doing \(R^2\), flat computations with such coordinates with messages like

although coordinates are longitude/latitude, st_intersects assumes that they are planar

hinting to the responsibility of the user to take care of potential problems. Doing this however leaves ambiguities, e.g. whether LINESTRING(-179 0,179 0)

and whether it is

Starting with sf version 1.0, sf uses the new package s2 (Dunnington, Pebesma, Rubak 2020) for spherical geometry, which has functions for computing pretty much all measures, predicates and transformations on the sphere. This means:

The s2 package is really a wrapper around the C++ s2geometry library which was written by Google, and which is used in many of its products (e.g. Google Maps, Google Earth Engine, Bigquery GIS) and has been translated in several programming other languages.

Fundamental differences

Compared to geometry on \(R^2\), and DE9-IM, the s2 package brings a few fundamentally new concepts, which are discussed first.

Polygons on \(S^2\) divide the sphere in two parts

On the sphere (\(S^2\)), any polygon defines two areas; when following the exterior ring, we need to define what is inside, and the definition is the left side of the enclosing edges. This also means that we can flip a polygon (by inverting the edge order) to obtain the other part of the globe, and that in addition to an empty polygon (the empty set) we can have the full polygon (the entire globe).

Simple feature geometries should obey a ring direction too: exterior rings should be counter clockwise, interior (hole) rings should be clockwise, but in some sense this is obsolete as the difference between exterior ring and interior rings is defined by their position (exterior, followed by zero or more interior). sf::read_sf has an argument check_ring_dir that checks, and corrects, ring directions and many (legacy) datasets have wrong ring directions. With wrong ring directions, many things still work.

For \(S^2\), ring direction is essential. For that reason, st_as_s2 has an argument oriented = FALSE, which will check and correct ring directions, assuming that all exterior rings occupy an area smaller than half the globe:

Here is an example where the oceans are computed as the difference from the full polygon representing the entire globe,

and the countries, and shown in an orthographic projection:

(Note that the WKT printing of the full polygon g is not valid WKT, and can not be used as input as it doesn’t follow the simple feature standard; it seems to reflect the internal representation of the full polygon, and would require a WKT extension).

We can now calculate the proportion of the Earth’s surface covered by oceans:

Semi-open polygon boundaries

Polygons in s2geometry can be

In principle the DE9-IM model deals with interior, boundary and exterior, and intersection predicates are sensitive to this (the difference between contains and covers is all about boundaries). DE9-IM however cannot uniquely assign points to polygons when polygons form a polygon coverage (no overlaps, but shared boundaries). This means that if we would count points by polygon, and some points fall on shared polygon boundaries, we either miss them (contains) or we count them double (covers, intersects); this might lead to bias and require post-processing. Using SEMI-OPEN non-overlapping polygons guarantees that every point is assigned to maximally one polygon in an intersection. This corresponds to e.g. how this would be handled in a grid (raster) coverage, where every grid cell (typically) only contains its upper-left corner and its upper and left sides.

Bounding cap, bounding rectangle

Computing the minimum and maximum values over coordinate ranges, as sf does with st_bbox(), is of limited value for spherical coordinates because due the the spherical space, the area covered is not necessarily covered by the coordinate range. Two examples:

S2 has two alternatives: the bounding cap and the bounding rectangle:

The cap reports a bounding cap (circle) as a mid point (lat, lng) and an angle around this point. The bounding rectangle reports the _lo and _hi bounds of lat and lng coordinates. Note that for Fiji, lng_lo being higher than lng_hi indicates that the region covers (crosses) the antimeridian.

Switching between S2 and GEOS

The two-dimensional \(R^2\) library that was formerly used by sf is GEOS, and sf can be instrumented to use GEOS or sf. First we will ask if s2 is being used by default:

## [1] TRUE

then we can switch it of (and use GEOS) by


and switch it on (and use s2) by







All unary and binary predicates are available in s2, except for st_relate with a pattern. In addition, when using the s2 predicates

st_intersects(nc[1:3,], nc[1:3,]) # self-intersections + neighbours
## although coordinates are longitude/latitude, st_intersects assumes that they are planar
## Sparse geometry binary predicate list of length 3, where the predicate was `intersects'
##  1: 1, 2
##  2: 1, 2, 3
##  3: 2, 3
st_intersects(nc[1:3,], nc[1:3,], s2_model = "semi-open") # only self-intersections
## although coordinates are longitude/latitude, st_intersects assumes that they are planar
## Sparse geometry binary predicate list of length 3, where the predicate was `intersects'
##  1: 1, 2
##  2: 1, 2, 3
##  3: 2, 3


st_intersection, st_union, st_difference and st_sym_difference are available as s2 equivalents. N-ary instersection and difference are not (yet) present; cascaded union is present; unioning by feature does not work with s2.


Buffers can be calculated for features with geographic coordinates as follows, using an unprojected object representing the UK as an example:

The plots above show that you can adjust the level of spatial precision in the results of s2 buffer operations with the max_cells argument, set to 1000 by default. Deciding on an appropriate value is a balance between excessive detail increasing computational resources (represented by uk_buffer2, bottom left) and excessive simplification (bottom right). Note that buffers created with s2 always follow s2 cell boundaries, they are never smooth. Hence, choosing a large number for max_cells leads to seemingly smooth but, zoomed in, very complex buffers.

To achieve a similar result you could first transform the result and then use the st_buffer() function from sf. A simple benchmark shows the computational efficiency of the s2 geometry engine in comparison with transforming and then creating buffers:

The result of the previous benchmarks emphasise the point that there are trade-offs between geographic resolution and computational resources, something that web developers working on geographic services such as Google Maps understand well. In this case the default setting of 1000 cells, which runs slightly faster than the default transform -> buffer workflow, is probably appropriate given the low resolution of the input geometry representing the UK.

st_buffer or st_is_within_distance?

As discussed in the sf issue tracker, deciding on workflows and selecting appropriate levels of level of geographic resolution can be an iterative process. st_buffer as powered by GEOS, for \(R^2\) data, are smooth and (nearly) exact. st_buffer as powered by \(S^2\) is rougher, complex, non-smooth, and may need tuning. An common pattern where st_buffer is used is this:

When this is the case, and you are working with geographic coordinates, it may pay off to not compute buffers, but instead directly work with st_is_within_distance to select, for each feature of x, all features of y that are within a certain distance d from x. The \(S^2\) version of this function uses spatial indexes, so is fast for large datasets.