3 Attribute data operations
- This chapter requires the following packages to be installed and attached:
- It also relies on spData, which loads datasets used in the code examples of this chapter:
Attribute data is non-spatial information associated with geographic (geometry) data. A bus stop provides a simple example: its position would typically be represented by latitude and longitude coordinates (geometry data), in addition to its name. The name is an attribute of the feature (to use Simple Features terminology) that bears no relation to its geometry.
Another example is the elevation value (attribute) for a specific grid cell in raster data. Unlike vector data, the raster data model stores the coordinate of the grid cell only indirectly: There is a less clear distinction between attribute and spatial information in raster data. Say, we are in the 3rd row and the 4th column of a raster matrix. To derive the corresponding coordinate, we have to move from the origin three cells in x-direction and four cells in y-direction with the cell resolution defining the distance for each x- and y-step. The raster header gives the matrix a spatial dimension which we need when plotting the raster or when we want to combine two rasters, think, for instance, of adding the values of one raster to another (see also next chapter).
The focus of this chapter is manipulating geographic objects based on attributes such as the name of a bus stop and elevation.
For vector data this means operations such as subsetting and aggregation (see sections 3.2.1 and 3.2.2).
These non-spatial operations have spatial equivalents:
[ operator in base R, for example, works equally for subsetting objects based on their attribute and spatial objects, as we will see in Chapter 4.
This is good news: skills developed here are cross-transferable, meaning that this chapter lays the foundation for Chapter 4, which extends the methods presented here to the spatial world.
Sections 3.2.3 and 3.2.4 demonstrate how to join data onto simple feature objects using a shared ID and how to create new variables, respectively.
Raster attribute data operations are covered in Section 3.3, which covers creating continuous and categorical raster layers and extracting cell values from one layer and multiple layers (raster subsetting). Section 3.3.2 provides an overview of ‘global’ raster operations which can be used to characterize entire raster datasets.
3.2 Vector attribute manipulation
Geographic vector data in R are well-support by
sf, a class which extends the
sf objects have one column per attribute variable (such as ‘name’) and one row per observation, or feature (e.g. per bus station).
sf objects also have a special column to contain geometry data, usually named
geometry column is special because it is a list-colum, which can contain multiple geographic entities (points, lines, polygons) per row.
In Chapter 2 we saw how to perform generic methods such as
sf also provides methods that allow
sf objects to behave like regular data frames:
Many of these functions, including
rbind() (for binding rows of data together) and
$<- (for creating new columns) were developed for data frames.
A key feature of
sf objects is that they store spatial and non-spatial data in the same way, as columns in a
data.frame (the geometry column is typically called
The geometry column of
sf objects is typically called
geometry but any name can be used.
The following command, for example, creates a geometry column named g:
This enables geometries imported from spatial databases to have a variety of names such as
st_sf(data.frame(n = world$name_long), g = world$geom)
sf objects also support
tbl classes used in the tidyverse, allowing ‘tidy’ data analysis workflows for spatial data.
Thus sf enables the full power of R’s data analysis capabilities to be unleashed on geographic data.
Before using these capabilities it’s worth re-capping how to discover the basic properties of vector data objects.
Let’s start by using base R functions to get a measure of the
Our dataset contains ten non-geographic columns (and one geometry list-column) with almost 200 rows representing the world’s countries.
Extracting the attribute data of an
sf object is the same as removing its geometry:
This can be useful if the geometry column causes problems, e.g. by occupying large amounts of RAM, or to focus the attention on the attribute data.
For most cases, however, there is no harm in keeping the geometry column because non-spatial data operations on
sf objects only change an object’s geometry when appropriate (e.g. by dissolving borders between adjacent polygons following aggregation).
This means that proficiency with attribute data in
sf objects equates to proficiency with data frames in R.
For many applications, the tidyverse package dplyr offers the most effective and intuitive approach for working with data frames.
Tidyverse compatibility is an advantage of sf over its predecessor sp, but there are some pitfalls to avoid (see the supplementary
tidyverse-pitfalls vignette at geocompr.github.io for details).
3.2.1 Vector attribute subsetting
Base R subsetting functions include
dplyr subsetting functions include
Both sets of functions preserve the spatial components of attribute data in
[ operator can subset both rows and columns.
You use indices to specify the elements you wish to extract from an object, e.g.
object[i, j], with
j typically being numbers or logical vectors —
FALSEs — representing rows and columns (they can also be character strings, indicating row or column names).
j empty returns all rows or columns, so
world[1:5, ] returns the first five rows and all columns.
The examples below demonstrate subsetting with base R.
The results are not shown; check the results on your own computer:
A demonstration of the utility of using
logical vectors for subsetting is shown in the code chunk below.
This creates a new object,
small_countries, containing nations whose surface area is smaller than 10,000 km2:
sel_area is a logical vector that shows that only seven countries match the query.
A more concise command, that omits the intermediary object, generates the same result:
The base R function
subset() provides yet another way to achieve the same result:
Base R functions are mature and widely used.
However, the more recent dplyr approach has several advantages.
It enables intuitive workflows.
It is fast, due to its C++ backend.
This is especially useful when working with big data as well as dplyr’s database integration.
The main dplyr subsetting functions are
raster and dplyr packages have a function called
select(). When using both packages, the function in the most recently attached package will be used, ‘masking’ the incumbent function. This can generate error messages containing text like:
unable to find an inherited method for function ‘select’ for signature ‘“sf”’. To avoid this error message, and prevent ambiguity, we use the long-form function name, prefixed by the package name and two colons (usually omitted from R scripts for concise code):
select() selects columns by name or position.
For example, you could select only two columns,
pop, with the following command (note the sticky
geom column remains):
select() also allows subsetting of a range of columns with the help of the
Omit specific columns with the
select() lets you subset and rename columns at the same time, for example:
This is more concise than the base R equivalent:
select() also works with ‘helper functions’ for advanced subsetting operations, including
num_range() (see the help page with
?select for details).
All dplyr functions including
select() always return a dataframe-like object.
To extract a single vector, one has to explicitly use the
The subsetting operator in base R (see
?[), by contrast, tries to return objects in the lowest possible dimension.
This means selecting a single column returns a vector in base R.
To turn off this behavior, set the
drop argument to
Due to the sticky geometry column, selecting a single attribute from an sf-object with the help of
[() returns also a dataframe.
$ will give back a vector.
slice() is the row-equivalent of
The following code chunk, for example, selects the 3rd to 5th rows:
filter() is dplyr’s equivalent of base R’s
It keeps only rows matching given criteria, e.g. only countries with a very high average of life expectancy:
The standard set of comparison operators can be used in the
filter() function, as illustrated in Table 3.1:
||Not equal to|
||Greater/Less than or equal|
||Logical operators: And, Or, Not|
dplyr works well with the ‘pipe’ operator
%>%, which takes its name from the Unix pipe
| (Grolemund and Wickham 2016).
It enables expressive code: the output of a previous becomes the first argument of the next function, enabling chaining.
This is illustrated below, in which the
world dataset is subset by columns (
continent) and the first five rows (result not shown).
The above chunk shows how the pipe operator allows commands to be written in a clear order:
the above run from top to bottom (line-by-line) and left to right.
The alternative to
%>% is nested function calls, which is harder to read:
3.2.2 Vector attribute aggregation
Aggregation operations summarize datasets by a ‘grouping variable’, typically an attribute column (spatial aggregation is covered in the next chapter).
An example of attribute aggregation is calculating the number of people per continent based on country-level data (one row per country).
world dataset contains the necessary ingredients: the columns
continent, the population and the grouping variable respectively.
The aim is to find the
sum() of country populations for each continent.
This can be done with the base R function
aggregate() as follows:
The result is a non-spatial data frame with six rows, one per continent, and two columns reporting the name and population of each continent (see Table 3.2 with results for the top 3 most populous continents).
aggregate() is a generic function which means that it behaves differently depending on its inputs.
sf provides a function that can be called directly with
sf:::aggregate() that is activated when a
by argument is provided, rather than using the
~ to refer to the grouping variable:
As illustrated above, an object of class
sf is returned this time.
world_agg2 which is a spatial object containing 6 polygons representing the columns of the world.
summarize() is the dplyr equivalent of
It usually follows
group_by(), which specifies the grouping variable, as illustrated below:
This approach is flexible and gives control over the new column names. This is illustrated below: the command calculates the Earth’s population (~7 billion) and number of countries (result not shown):
In the previous code chunk
n are column names in the result.
n() were the aggregating functions.
The result is an
sf object with a single row representing the world (this works thanks to the geometric operation ‘union’, as explained in section 5.2.6).
Let’s combine what we’ve learned so far about dplyr by chaining together functions to find the world’s 3 most populous continents (with
dplyr::n() ) and the number of countries they contain (the result of this command is presented in Table 3.2):
vignette(package = "dplyr")and Chapter 5 of R for Data Science.
3.2.3 Vector attribute joining
Combining data from different sources is a common task in data preparation.
Joins do this by combining tables based on a shared ‘key’ variable.
dplyr has multiple join functions including
inner_join() — see
vignette("two-table") for a full list.
These function names follow conventions used in the database language SQL (Grolemund and Wickham 2016, Chapter 13); using them to join non spatial datasets to
sf objects is the focus of this section.
dplyr join functions work the same on data frames and
sf objects, the only important difference being the
geometry list column.
The result of data joins can be either an
The most common type of attribute join on spatial data takes an
sf object as the first argument and adds columns to it from a
data.frame specified as the second argument.
To illustrate the utility of joins, imagine that you are researching global coffee production.
You have managed to extract data hidden-away in a PDF document supplied by the International Coffee Organization (ICO).
The results are stored in a data frame called
coffee_data which has 3 columns:
name_long) containing the names of coffee-producing nations and the other two reporting coffee production statistics for 2016 and 2017 (see
?coffee_data for further information).
We will use a ‘left join’ (meaning the left-hand dataset is kept intact) to merge this dataset with the pre-existing
The result of the code chunk below is a new
The resulting simple features object is the same as the orignal
world object but has two new variables (with column indeces 11 and 12) reporting coffee production by year.
This can be plotted as a map, as illustrated in Figure 3.1, generated with the
plot() function below:
For joining to work a ‘key variable’ must be supplied in both datasets.
By default dplyr uses all variables with matching names.
In this case both
world objects contained a variable called
name_long, explaining the message
Joining, by = "name_long".
In the majority of cases where variable names are not the same you have two options:
- Rename the key variable in one of the objects so they match.
- Use the
byargument to specify the joining variables.
The latter approach is demonstrated below on a renamed version of
Note that the name in the original object is kept, meaning that
world_coffee and the new object
world_coffee2 are identical.
Another feature of the result is that it has the same number of rows as the original dataset.
Although there are only 47 rows of data in
coffee_data all 177 the country records are kept intact in
rows in the original dataset with no match are assigned
NA values for the new coffee production variables.
What if we only want to keep countries that have a match in the key variable?
In that case an inner join can be used:
Note that the result of
inner_join() has only 45 rows compared with 47 in
What happened to the remaining rows?
We can identify the rows that did not match using the
setdiff() function as follows:
The result shows that
Others accounts for one row not present in the
world dataset and that the name of the
Democratic Republic of the Congo accounts for the other:
it has been abbreviated, causing the join to miss it.
The following command uses string matching (regex) to confirm what
Congo, Dem. Rep. of should be:
To fix this issue we will create a new version of
coffee_data and update the name.
inner_join()ing the updated data frame returns a result with all 46 coffee producing nations:
It is also possible to join in the other direction: starting with a non-spatial dataset and adding variables from a simple features object.
This is demonstrated below, which starts with the
coffee_data object and adds variables from the original
In contrast with the previous joins, the result is not another simple feature object, but a data frame in the form of a tidyverse tibble:
the output of a join tends to match its first argument:
sfobject. The geometry column can only be used for creating maps and spatial operations if R ‘knows’ it is a spatial object, defined by a spatial package such as sf. Fortunately non-spatial data frames with a geometry list column (like
coffee_world) can be coerced into an
sfobject as follows:
This section covers the majority joining use cases. For more information we recommend Grolemund and Wickham (2016), the join vignette in the geocompkg package that accompanies this book, and documentation of the data.table package.16 Another type of join is a spatial join, covered in the next chapter (section 4.2.3).
3.2.4 Creating attributes and removing spatial information
Often, we would like to create a new column based on already existing columns.
For example, we want to calculate population density for each country.
For this we need to divide a population column, here
pop, by an area column , here
area_km2 with unit area in square km.
Using base R, we can type:
Alternatively, we can use one of dplyr functions -
mutate() adds new columns at the penultimate position in the
sf object (the last one is reserved for the geometry):
The difference between
transmute() is that the latter skips all other existing columns (except for the sticky geometry column):
unite() pastes together existing columns.
For example, we want to combine the
region_un columns into a new column named
Additionally, we can define a separator (here: a colon
:) which defines how the values of the input columns should be joined, and if the original columns should be removed (here:
separate() function does the opposite of
unite(): it splits one column into multiple columns using either a regular expression or character positions.
The two functions
set_names() are useful for renaming columns.
The first replaces an old name with a new one.
The following command, for example, renames the lengthy
name_long column to simply
set_names() changes all column names at once, and requires a character vector with a name matching each column.
This is illustrated below, which outputs the same
world object, but with very short names:
It is important to note that attribute data operations preserve the geometry of the simple features.
As mentioned at the outset of the chapter, it can be useful to remove the geometry.
To do this, you have to explicitly remove it because
sf explicitly makes the geometry column sticky.
This behavior ensures that data frame operations do not accidentally remove the geometry column.
Hence, an approach such as
select(world, -geom) will be unsuccessful and you should instead use
3.3 Manipulating raster objects
In contrast to the vector data model underlying simple features (which represents points, lines and polygons as discrete entities in space), raster data represent continuous surfaces. This section shows how raster objects work, by creating them from scratch, building on section 2.2.1. Because of their unique structure, subsetting and other operations on raster datasets work in a different way, as demonstrated in section 3.3.1.
The following code recreates the raster dataset used in section 2.2.3, the result of which is illustrated in Figure 3.2.
This demonstrates how the
raster() function works to create an example raster named
elev (representing elevations).
The result is a raster object with 6 rows and 6 columns (specified by the
ncol arguments), and a minimum and maximum spatial extent in x and y direction (
vals argument sets the values that each cell contains: numeric data ranging from 1 to 36 in this case.
Raster objects can also contain categorical values of class
factor variables in R.
The following code creates a raster representing grain sizes (Figure 3.2):
rasterobjects can contain values of class
factor, but not
character. To use character values they must first be converted into an appropriate class, for example using the function
levelsargument was used in the preceding code chunk to create an ordered factor: clay < silt < sand in terms of grain size. See the Data structures chapter of Wickham (2014a) for further details on classes.
raster objects represent categorical variables as integers, so
grain[1, 1] returns a number that represents a unique identifier, rather than “clay”, “silt” or “sand”.
The raster object stores the corresponding look-up table or “Raster Attribute Table” (RAT) as a data frame in a new slot named
attributes, which can be viewed with
?ratify() for more information).
Use the function
levels() for retrieving and adding new factor levels to the attribute table:
This behavior demonstrates that raster cells can only possess one value, an identifier which can be used to look up the attributes in the corresponding attribute table (stored in a slot named
This is illustrated in command below, which returns the grain size and wetness of cell IDs 1, 11 and 35, we can run:
3.3.1 Raster subsetting
Raster subsetting is done with the base R operator
[, which accepts a variety of inputs:
- Row-column indexing
- Cell IDs
- Another raster object
Here, we only show the first two options since these can be considered non-spatial operations. If we need a spatial object to subset another or the output is a spatial object, we refer to this as spatial subsetting. Therefore, the latter two options will be shown in the next chapter (see section 4.3.1 in the next chapter).
The first two subsetting options are demonstrated in the commands below —
both return the value of the top left pixel in the raster object
elev (results not shown):
To extract all values or complete rows, you can use
For multi-layered raster objects
brick, this will return the cell value(s) for each layer.
stack(elev, grain) returns a matrix with one row and two columns — one for each layer.
For multi-layer raster objects another way to subset is with
raster::subset(), which extracts layers from a raster stack or brick. The
$ operators can also be used:
Cell values can be modified by overwriting existing values in conjunction with a subsetting operation.
The following code chunk, for example, sets the upper left cell of
elev to 0:
Leaving the square brackets empty is a shortcut version of
values() for retrieving all values of a raster.
Multiple cells can also be modified in this way:
3.3.2 Summarizing raster objects
raster contains functions for extracting descriptive statistics for entire rasters.
Printing a raster object to the console by typing its name, returns minimum and maximum values of a raster.
summary() provides common descriptive statistics (minimum, maximum, interquartile range and number of
Further summary operations such as the standard deviation (see below) or custom summary statistics can be calculated with
cellStats()functions with a raster stack or brick object, they will summarize each layer separately, as can be illustrated by running:
Raster value statistics can be visualized in a variety of ways.
Specific functions such as
pairs() work also with raster objects, as demonstrated in the histogram created with the command below (not shown):
In case a visualization function does not work with raster objects, one can extract the raster data to be plotted with the help of
Descriptive raster statistics belong to the so-called global raster operations. These and other typical raster processing operations are part of the map algebra scheme which are covered in the next chapter (section 4.3.2).
Some function names clash between packages (e.g.
select, as discussed in a previous note). In addition to not loading packages by referring to functions verbosely (e.g.
dplyr::select()) another way to prevent function names clashes is by unloading the offending package with
detach(). The following command, for example, unloads the raster package (this can also be done in the package tab which resides by default in the right-bottom pane in RStudio):
detach(“package:raster”, unload = TRUE, force = TRUE). The
force argument makes sure that the package will be detached even if other packages depend on it. This, however, may lead to a restricted usability of packages depending on the detached package, and is therefore not recommended.
For these exercises we will use the
us_states_df datasets from the spData package:
us_states is a spatial object (of class
sf), containing geometry and a few attributes (including name, region, area, and population) of states within the contiguous United States.
us_states_df is a data frame (of class
data.frame) containing the name and additional variables (including median income and poverty level, for years 2010 and 2015) of US states, including Alaska, Hawaii and Puerto Rico.
The data comes from the US Census Bureau, and is documented in
- Create a new object called
us_states_namethat contains only the
NAMEcolumn from the
us_statesobject. What is the class of the new object and what makes it geographic?
- Select columns from the
us_statesobject which contain population data. Obtain the same result using a different command (bonus: try to find three ways of obtaining the same result). Hint: try to use helper functions, such as
starts_withfrom dplyr (see
- Find all states with the following characteristics (bonus find and plot them):
- Belong to the Midwest region.
- Belong to the West region, have an area below 250,000 km2 and in 2015 a population greater than 5,000,000 residents (hint: you may need to use the function
- Belong to the South region, had an area larger than 150,000 km2 or a total population in 2015 larger than 7,000,000 residents.
- What was the total population in 2015 in the
us_statesdataset? What was the minimum and maximum total population in 2015?
- How many states are there in each region?
- What was the minimum and maximum total population in 2015 in each region? What was the total population in 2015 in each region?
- Add variables from
us_states, and create a new object called
us_states_stats. What function did you use and why? Which variable is the key in both datasets? What is the class of the new object?
us_states_dfhas two more variables than
us_states. How can you find them? (hint: try to use the
- What was the population density in 2015 in each state? What was the population density in 2010 in each state?
- How much has population density changed between 2010 and 2015 in each state? Calculate the change in percentages and map them.
- Change the columns names in
us_statesto lowercase. (Hint: helper functions -
us_states_dfcreate a new object called
us_states_sel. The new object should have only two variables -
geometry. Change the name of the
- Calculate the change in median income between 2010 and 2015 for each state. Bonus: what was the minimum, average and maximum median income in 2015 for each region? What is the region with the largest increase of the median income?
- Create a raster from scratch with nine rows and columns and a resolution of 0.5 decimal degrees (WGS84). Fill it with random numbers. Extract the values of the four corner cells.
- What is the most common class of our example raster
- Plot the histogram and the boxplot of the
data(dem, package = "RQGIS")raster.
Grolemund, Garrett, and Hadley Wickham. 2016. R for Data Science. 1 edition. O’Reilly Media.
Wickham, Hadley. 2014a. Advanced R. CRC Press.