How to calculate the length of borders between countries in R? - r

Question

In some analyses, it makes sense to use border length as a measure of cultural distance between countries, the idea being that countries that share larger proportions of their borders are more culturally close. This then raises the question of how to compute this. We can grab a shapefile of the world from naturalearthdata.com which covers some 251 units (i.e. they are not all sovereign).
I looked over the methods in the Geocomputation with R ebook website and it seems like an intersection is closest to what we want, i.e. st_intersection(), while st_touches() finds the neighbors without giving any sense of the border length. However, when I try it out on two neighbors, Denmark and Germany, I get no overlap:
> suppressWarnings(library(sf))
Linking to GEOS 3.6.2, GDAL 2.2.3, PROJ 4.9.3; sf_use_s2() is TRUE
> world = read_sf("data/ne_10m_admin_0_countries/ne_10m_admin_0_countries.shp")
> #make valid otherwise get an error
> world = st_make_valid(world)
> #which countries touch each other by border (of the polygons)
> neighbor_ids = st_touches(
+ world$geometry,
+ world$geometry
+ )
> #Denmark Germany
> (germany_idx = which(world$ADMIN=="Germany"))
[1] 50
> (denmark_idx = which(world$ADMIN=="Denmark"))
[1] 71
> world$ADMIN[neighbor_ids[germany_idx][[1]]]
[1] "France" "Czechia" "Luxembourg" "Belgium" "Denmark" "Poland" "Austria" "Switzerland" "Netherlands"
> world$ADMIN[neighbor_ids[denmark_idx][[1]]]
[1] "Germany"
> #border intersection
> #Denmark Germany border as test
> st_intersection(
+ world$geometry[germany_idx],
+ world$geometry[denmark_idx]
+ )
Geometry set for 0 features
Bounding box: xmin: NA ymin: NA xmax: NA ymax: NA
CRS: 4326
How does one get the border lengths? According to Wikipedia, it should be 68 km.
It seems that what is needed is to tell st_intersection() to include the line at the border. By default, this 1 point overlap is ignored, I guess because it has a 0 area. This functionality is controlled by the ... which forwards to s2_options(). The right parameter is model, which defaults to "open", but should be "closed". Thus:
> #include the line
> st_intersection(
+ world$geometry[germany_idx],
+ world$geometry[denmark_idx],
+ model = "closed"
+ )
Geometry set for 1 feature
Geometry type: MULTILINESTRING
Dimension: XY
Bounding box: xmin: 8.660776 ymin: 54.80162 xmax: 9.437503 ymax: 54.9059
CRS: 4326
MULTILINESTRING ((9.436922 54.81014, 9.422143 5...
To get the length, just add on st_length():
> st_intersection(
+ world$geometry[germany_idx],
+ world$geometry[denmark_idx],
+ mode = "closed"
+ ) %>% st_length()
111563 [m]
Only the result is wrong! The scaling is off by factor of 1.64 or so.
Potential problems:
Is this an issue with the coastline paradox?
Some kind of incorrect setting? I the only setting for st_distance() is the size of the earth, which seems to be set correctly.
Bad shapefile? I downloaded a different one (I forgot the source), and it produced a result of 113469 m, which is slightly different but not remotely close to 68000 as Wikipedia gives.
Is it due to a water border? I plotted the border with tmap, and it looks fine.

Answer 1

This is an interesting problem; I believe the coastline paradox plays a role, but only a minor one. The chief issue seems to be driven by CRS.
Let me illustrate on three examples using the world dataset provided by GISCO (i.e. Eurostat). I like this dataset as it allows several levels of precision.
a rough map in EPSG:3035 (the official CRS for continental EU)
a fine map in EPSG:3035
the same fine map in EPSG:4326 / WGS84
Compare these with the official, i.e. wikipedia length of 68 kilometers.
The rough map is off by about 1/6th, which is to be expected given the low resolution. The fine map is quite close (7% off), and you could expect the actual length to increase yet more, as 1:1M is still a coarse map.
On the other hand the length of the same fine map as in previous example, but projected in WGS84, is off by a factor of two, as you observed.
library(sf)
library(dplyr)
library(giscoR)
# rough line / resolution 1:60 000 000
rough <- gisco_get_countries(resolution = "60",
epsg = 3035,
country = c("DE", "DK"))
plot(st_geometry(rough))
sf::st_intersection(rough[1, ], rough[2, ], model = "closed") %>%
mutate(border_length = st_length(.)) %>%
pull(border_length)
# 53141.03 [m]
# fine line / resolution 1:1 000 000
fine <- gisco_get_countries(resolution = "01",
epsg = 3035,
country = c("DE", "DK"))
plot(st_geometry(fine))
sf::st_intersection(fine[1, ], fine[2, ], model = "closed") %>%
mutate(border_length = st_length(.)) %>%
pull(border_length)
# 63795.4 [m]
# fine line in WGS84
fine_wgs <- gisco_get_countries(resolution = "01",
epsg = 4326,
country = c("DE", "DK"))
sf::st_intersection(fine_wgs[1, ], fine_wgs[2, ], model = "closed") %>%
mutate(border_length = st_length(.)) %>%
pull(border_length)
# 127223 [m]
EDIT (2022-09-12) on second thought this seems to be affected by the behavior of the S2 engine behind {sf} (turning it off via sf_use_s2(FALSE) leads to more reasonable length of borders even for data projected in WGS84).
I will raise it as an issue with {sf} maintainers, as it does not seem likely that this is an expected behaviour.