This chapter introduces a number of concepts associated with handling spatial and spatiotemporal data, pointing forward to later chapters where these concepts are discussed in more detail. It also introduces a number of open source technologies that form the foundation of all spatial data science language implementations.

1.1 A first map

The typical way to graph spatial data is by creating a map. Let us consider a simple map, shown in Figure 1.1.

Code
library(tidyverse)
library(sf)
system.file("gpkg/nc.gpkg", package="sf") |>
    read_sf() -> nc
nc.32119 <- st_transform(nc, 'EPSG:32119')
nc.32119 |>
    select(BIR74) |>
    plot(graticule = TRUE, axes = TRUE)
Figure 1.1: A first map: birth counts 1974-78, North Carolina counties

A number of graphical elements are present here, in this case:

  • polygons are drawn with a black outline and filled with colours chosen according to a variable BIR74, whose name is in the title
  • a legend key explains the meaning of the colours, and has a certain colour palette and colour breaks, values at which colour changes
  • the background of the map shows curved lines with constant latitude or longitude (graticule)
  • the axis ticks show the latitude and longitude values

Polygons are a particular form of geometry; spatial geometries (points, lines, polygons, pixels) are discussed in detail in Chapter 3. Polygons consist of sequences of points, connected by straight lines. How point locations of spatial data are expressed, or measured, is discussed in Chapter 2. As can be seen from Figure 1.1, lines of equal latitude and longitude do not form straight lines, indicating that some form of projection took place before plotting; projections are also discussed in Chapter 2 and Section 8.1.

The colour values in Figure 1.1 are derived from numeric values of a variable, BIR74, which has a single value associated with each geometry or feature. Chapter 5 discusses such feature attributes, and how they can relate to feature geometries. In this case, BIR74 refers to birth counts, meaning counts over the region. This implies that the count does not refer to a value associated with every point inside the polygon, which the continuous colour might suggest, but rather measures an integral (sum) over the polygon.

Before plotting Figure 1.1 we had to read the data, in this case from a file (Section 7.1). Printing a data summary for the first three records of three attribute variables shows:

Code
nc |> select(AREA, BIR74, SID74) |> print(n = 3)
# Simple feature collection with 100 features and 3 fields
# Geometry type: MULTIPOLYGON
# Dimension:     XY
# Bounding box:  xmin: -84.3 ymin: 33.9 xmax: -75.5 ymax: 36.6
# Geodetic CRS:  NAD27
# # A tibble: 100 × 4
#    AREA BIR74 SID74                                             geom
#   <dbl> <dbl> <dbl>                               <MULTIPOLYGON [°]>
# 1 0.114  1091     1 (((-81.5 36.2, -81.5 36.3, -81.6 36.3, -81.6 36…
# 2 0.061   487     0 (((-81.2 36.4, -81.2 36.4, -81.3 36.4, -81.3 36…
# 3 0.143  3188     5 (((-80.5 36.2, -80.5 36.3, -80.5 36.3, -80.5 36…
# # ℹ 97 more rows

The printed output shows:

  • the (selected) dataset has 100 features (records) and 3 fields (attributes)
  • the geometry type is MULTIPOLYGON (Chapter 3)
  • it has dimension XY, indicating that each point will consist of 2 coordinate values
  • the range of \(x\) and \(y\) values of the geometry
  • the coordinate reference system (CRS) is geodetic, with coordinates in degrees longitude and latitude associated to the NAD27 datum (Chapter 2)
  • the three selected attribute variables are followed by a variable geom of type MULTIPOLYGON with unit degrees that contains the polygon information

More complicated plots can involve facet plots with a map in each facet, as shown in Figure 1.2.

Code
year_labels <- c("SID74" = "1974 - 1978", "SID79" = "1979 - 1984")
nc.32119 |> select(SID74, SID79) |>
    pivot_longer(starts_with("SID")) -> nc_longer
ggplot() + geom_sf(data = nc_longer, aes(fill = value), linewidth = 0.4) + 
  facet_wrap(~ name, ncol = 1, labeller = labeller(name = year_labels)) +
  scale_y_continuous(breaks = 34:36) +
  scale_fill_gradientn(colors = sf.colors(20)) +
  theme(panel.grid.major = element_line(color = "white"))
Figure 1.2: Facet maps of sudden infant death syndrome counts, 1974-78 and 1979-84, North Carolina counties

An interactive, leaflet-based map is obtained in Figure 1.3.

Code
library(mapview) |> suppressPackageStartupMessages()
mapviewOptions(fgb = FALSE)
nc.32119 |> mapview(zcol = "BIR74", legend = TRUE, col.regions = sf.colors)
Figure 1.3: Interactive map created with mapview: pan and zoom move the map and change scale; clicking a county pops up window with the available county properties.

1.2 Coordinate reference systems

In Figure 1.1, the grey lines denote the graticule, a grid with lines along constant latitude or longitude. Clearly, these lines are not straight, which indicates that a projection of the data was used for which the \(x\) and \(y\) axes do not align with longitude and latitude. In Figure 1.3 we see that the north boundary of North Carolina is plotted as a straight line again, indicating that another projection was used.

The ellipsoidal coordinates of the graticule of Figure 1.1 are associated with a particular datum (here: NAD27), which implicates a set of rules, what the shape of the Earth is and how it is attached to the Earth (to which point of the Earth is the origin associated, and how is it directed.) If one would measure coordinates with a GPS device (such as a mobile phone) it would typically report coordinates associated with the World Geodetic System 1984 (WGS84) datum, which can be around 30 m different from the identical coordinate values when associated with the North American Datum 1927 (NAD27).

Projections describe how we go back and forth between

  • ellipsoidal coordinates which are expressed as degrees latitude and longitude, pointing to locations on a shape approximating the Earth’s shape (ellipsoid or spheroid), and

  • projected coordinates which are coordinates on a flat, two-dimensional coordinate system, used when plotting maps.

Datums transformations are associated with moving from one datum to another. Both topics are covered by spatial reference systems, and are described in more detail in Chapter 2.

1.3 Raster and vector data

Polygon, point, and line geometries are examples of vector data: point coordinates describe the “exact” locations that can be anywhere. Raster data on the other hand describe data where values are aligned on a raster, meaning on a regularly laid out lattice of usually square pixels. An example is shown in Figure 1.4.

Code
library(stars)
par(mfrow = c(2, 2))
par(mar = rep(1, 4))
tif <- system.file("tif/L7_ETMs.tif", package = "stars")
x <- read_stars(tif)[,,,1]
image(x, main = "(a)")
image(x[,1:10,1:10], text_values = TRUE, border = 'grey', main = "(b)")
image(x, main = "(c)")
set.seed(131)
pts <- st_sample(st_as_sfc(st_bbox(x)), 3)
plot(pts, add = TRUE, pch = 3, col = 'blue')
image(x, main = "(d)")
plot(st_buffer(pts, 500), add = TRUE, pch = 3, border = 'blue', col = NA, lwd = 2)
Figure 1.4: Raster maps (Olinda, Atlantic coast of Brazil): Landsat-7 blue band, with colour values derived from data values (a), the top-left \(10 \times 10\) sub-image from (a) with numeric values shown (b), and overlayed by two different types of vector data: three sample points (c), and a 500 m radius around the points represented as polygons (d)

Vector and raster data can be combined in different ways; for instance we can query the raster at the three points of Figure 1.4(c) or compute an aggregate, such as the average, over arbitrary regions such as the circles shown in Figure 1.4(d).

Code
st_extract(x, pts) # query at points
aggregate(x, st_buffer(pts, 500), FUN = mean) |> st_as_sf() # aggregate over circles

Other raster-to-vector conversions are discussed in Section 7.6 and include:

  • converting raster pixels into point values
  • converting raster pixels into small polygons, possibly merging polygons with identical values (“polygonize”)
  • generating lines or polygons that delineate continuous pixel areas with a certain value range (“contour”)

Code
plot(st_rasterize(nc["BIR74"], dx = 0.1), col = sf.colors(), breaks = "equal")
Figure 1.5: Map obtained by rasterizing county births counts for the period 1974-78 shown in 1.1

Vector-to-raster conversions can be as simple as rasterizing polygons, as shown in Figure 1.5. Other, more general vector-to-raster conversions that may involve statistical modelling include:

  • interpolation of point values to points on a regular grid (Chapter 12)
  • estimating densities of points over a regular grid (Chapter 11)
  • area-weighted interpolation of polygon values to grid cells (Section 5.3)
  • direct rasterization of points, lines, or polygons (Section 7.6)

1.4 Raster types

Raster dimensions describe how the rows and columns relate to spatial coordinates. Figure 1.6 shows a number of different possibilities.

Code
x <- 1:5
y <- 1:4
d <- st_dimensions(x = x, y = y, .raster = c("x", "y"))
m <- matrix(runif(20),5,4)
r1 <- st_as_stars(r = m, dimensions = d)

r <- attr(d, "raster")
r$affine <- c(0.2, -0.2)
attr(d, "raster") = r
r2 <- st_as_stars(r = m, dimensions = d)

r <- attr(d, "raster")
r$affine <- c(0.1, -0.3)
attr(d, "raster") = r
r3 = st_as_stars(r = m, dimensions = d)

x <- c(1, 2, 3.5, 5, 6)
y <- c(1, 1.5, 3, 3.5)
d <- st_dimensions(x = x, y = y, .raster = c("x", "y"))
r4 <- st_as_stars(r = m, dimensions = d)

grd <- st_make_grid(cellsize = c(10,10), offset = c(-130,10), n = c(8,5), crs = st_crs('OGC:CRS84'))
r5 <- st_transform(grd, "+proj=laea +lon_0=-70 +lat_0=35")

par(mfrow = c(2,3), mar = c(0.1, 1, 1.1, 1))
r1 <- st_make_grid(cellsize = c(1,1), n = c(5,4), offset = c(0,0))
plot(r1, main = "regular")
plot(st_geometry(st_as_sf(r2)), main = "rotated")
plot(st_geometry(st_as_sf(r3)), main = "sheared")
plot(st_geometry(st_as_sf(r4, as_points = FALSE)), main = "rectilinear")
plot(st_geometry((r5)), main = "curvilinear")
Figure 1.6: Various raster geometry types

Regular rasters like those shown in Figure 1.6 have a constant, not necessarily square cell size and axes aligned with the \(x\) and \(y\) (Easting and Northing) axes. Other raster types include those where the axes are no longer aligned with \(x\) and \(y\) (rotated), where axes are no longer perpendicular (sheared), or where cell size varies along a dimension (rectilinear). Finally, curvilinear rasters have cell size and/or direction properties that are no longer independent from the other raster dimension.

When a raster that is regular in a given coordinate reference system is projected to another raster while keeping each raster cell intact, it changes shape and may become rectilinear (for instance when going from ellipsoidal coordinates to Mercator, as in Figure 1.3) or curvilinear (for instance when going from ellipsoidal coordinates to Lambert Conic Conformal, as used in Figure 1.1). When reverting this procedure, one can recover the exact original raster.

Creating a new, regular grid in the new projection is called raster (or image) reprojection or warping (Section 7.8). Warping is lossy, irreversible, and needs to be informed whether raster cells should be interpolated, averaged or summed, or whether resampling using nearest neighbours should be used. For such choices it matters whether cell values reflect a categorical or continuous variable (see also Section 1.6).

1.5 Time series, arrays, data cubes

A lot of spatial data is not only spatial, but also temporal. Just like any observation is associated with an observation location, it is associated with an observation time or period. The dataset on the North Carolina counties shown above contains disease cases counted over two time periods (shown in Figure 1.2). Although the original dataset has these variables in two different columns, for plotting them these columns had to be stacked first, while repeating the associated geometries - a form called tidy Wickham (2014). When we have longer time series associated with geometries, neither option - distributing time over multiple columns, or stacking columns while repeating geometries - works well, and a more effective way of storing such data would be a matrix or array, where one dimension refers to time, and the other(s) to space. The natural way for image or raster data is already to store them in matrices; time series of rasters then lead to a three-dimensional array. The general term for such data is a (spatiotemporal) data cube, where cube refers to arrays with any number of dimensions. Data cubes can refer to both raster and vector data, examples are given in Chapter 6.

1.6 Support

When we have spatial data with geometries that are not points but collections of points (multi-points, lines, polygons, pixels), then an attribute associated with these geometries has one of several different relationships to them. An attribute can have:

  • a constant value for every point of the geometry
  • a single value that is an aggregate over all points of the geometry
  • a value that is unique to only this geometry, describing its identity

An example of a constant is land use or bedrock type of a polygon. An example of an aggregate is the number of births of a county over a given period of time. An example of an identity is a county name.

The spatial area an attribute value refers to is called its support: aggregate properties have “block” (or area, or line) support, constant properties have “point” support (they apply to every point). Support matters when we manipulate the data. For instance, Figure 1.5 was derived from a variable that has polygon support: the number of births per county. Rasterizing these values gives pixels with values that remain associated with counties. The result of the rasterization is a meaningless map: the numeric values (“birth totals”) are not associated with the raster cells, and the county boundaries are no longer present. Totals of birth for the whole state or birth densities can no longer be recovered from the pixel values. Ignoring support can easily lead to meaningless results. Chapter 5 discusses this further.

Raster cell values may have point support or block support. An example of point support is elevation, when cells record the elevation of the point at the cell centre in a digital elevation model. An example of block (or cell) support is a satellite image pixel that gives the colour values averaged over (an area similar to) a pixel. Most file formats do not provide this information, yet it may be important to know when aggregating, regridding or warping rasters (Section 7.8), extracting values at point locations.

1.7 Spatial data science software

Although this book largely uses R and R packages for spatial data science, a number of these packages use software libraries that were not developed for R specifically. As an example, the dependency of R package sf on other R packages and system libraries is shown in Figure 1.7.

Figure 1.7: sf and its dependencies; arrows indicate strong dependency, dashed arrows weak dependency

The C or C++ libraries used (GDAL, GEOS, PROJ, liblwgeom, s2geometry, NetCDF, udunits2) are all developed, maintained, and used by (spatial) data science communities that are large and mostly different from the R community. By using these libraries, R users share how we understand what we are doing with these other communities. Because R, Python, and Julia provide interactive interfaces to this software, many users get closer to these libraries than do users of other software based on these libraries. The first part of this book describes many of the concepts implemented in these libraries, which is relevant to spatial data science in general.

GDAL

GDAL (Geospatial Data Abstraction Library) can be seen as the Swiss army knife of spatial data; besides for R it is being used in Python, QGIS, PostGIS, and more than 100 other software projects.

GDAL is a “library of libraries” – in order to read and write these data, it needs a large number of other libraries. It typically links to over 100 other libraries, each of which may provide access to a particular data file format, a database, a web service, or a particular compression codec.

Binary R packages distributed by CRAN contain only statically linked code: CRAN does not want to make any assumptions about presence of third-party libraries on the host system. As a consequence, when the sf package is installed in binary form from CRAN, it includes a copy of all the required external libraries as well as their dependencies, which may amount to 100 Mb.

PROJ

PROJ (or PR\(\phi\)J) is a library for cartographic projections and datum transformations: it converts spatial coordinates from one coordinate reference system to another. It comes with a large database of known projections and access to datum grids (high-precision, pre-calculated values for datum transformations). It aligns with an international standard for coordinate reference systems (Lott 2015). Chapter 2 deals with coordinate systems, and PROJ.

GEOS and s2geometry

GEOS (Geometry Engine Open Source) and s2geometry are two libraries for geometric operations. They are used to find measures (length, area, distance), and calculate predicates (do two geometries have any points in common?) or new geometries (which points do these two geometries have in common?). GEOS does this for flat, two-dimensional space (indicated by \(R^2\)), s2geometry does this for geometries on the sphere (indicated by \(S^2\)). Chapter 2 introduces coordinate reference systems, and Chapter 4 discusses more about the differences between working with these two spaces.

NetCDF, udunits2, liblwgeom

NetCDF (UCAR 2020) refers to a file format as well as a C library for reading and writing NetCDF files. It allows the definition of arrays of any dimensionality, and is widely used for spatial and spatiotemporal information, especially in the (climate) modelling communities. Udunits2 (UCAR 2014; Pebesma, Mailund, and Hiebert 2016; Pebesma et al. 2022) is a database and software library for units of measurement that allows the conversion of units, handles derived units, and supports user-defined units. The liblwgeom “library” is a software component of PostGIS (Obe and Hsu 2015) that contains several routines missing from GDAL or GEOS, including convenient access to GeographicLib routines (Karney 2013) that ship with PROJ.

1.8 Exercises

  1. List five differences between raster and vector data.
  2. In addition to those listed below Figure 1.1, list five further graphical components that are often found on a map.
  3. In your own words, why is the numeric information shown in Figure 1.5 misleading (or meaningless)?
  4. Under which conditions would you expect strong differences when doing geometrical operations on \(S^2\), compared to doing them on \(R^2\)?