05-geometry-operations.Rmd

# Geometry operations {#geometry-operations}

## Prerequisites {-}

- This chapter uses the same packages as Chapter \@ref(spatial-operations) but with the addition of **spDataLarge**, which was installed in Chapter \@ref(spatial-class):

```{r 05-geometry-operations-1, message=FALSE}
library(sf)
library(terra)
library(dplyr)
library(spData)
library(spDataLarge)
```

## Introduction

So far the book has explained the structure of geographic datasets (Chapter \@ref(spatial-class)), and how to manipulate them based on their non-geographic attributes (Chapter \@ref(attr)) and spatial relations (Chapter \@ref(spatial-operations)).
This chapter focuses on manipulating the geographic elements of geographic objects, for example by simplifying and converting vector geometries, cropping raster datasets, and converting vector objects into rasters and from rasters into vectors.
After reading it --- and attempting the exercises at the end --- you should understand and have control over the geometry column in `sf` objects and the extent and geographic location of pixels represented in rasters in relation to other geographic objects.

Section \@ref(geo-vec) covers transforming vector geometries with 'unary' and 'binary' operations.
Unary operations work on a single geometry in isolation, including simplification (of lines and polygons), the creation of buffers and centroids, and shifting/scaling/rotating single geometries using 'affine transformations' (Sections \@ref(simplification) to \@ref(affine-transformations)).
Binary transformations modify one geometry based on the shape of another, including clipping and geometry unions\index{vector!union}, covered in Sections \@ref(clipping) and \@ref(geometry-unions), respectively.
Type transformations (from a polygon to a line, for example) are demonstrated in Section \@ref(type-trans).

Section \@ref(geo-ras) covers geometric transformations on raster objects.
This involves changing the size and number of the underlying pixels, and assigning them new values.
It teaches how to change the resolution (also called raster aggregation and disaggregation), the extent and the origin of a raster.
These operations are especially useful if one would like to align raster datasets from diverse sources.
Aligned raster objects share a one-to-one correspondence between pixels, allowing them to be processed using map algebra operations, described in Section \@ref(map-algebra). 
The interaction between raster and vector objects is covered in Chapter \@ref(raster-vector). 
It presents how raster values can be 'masked' and 'extracted' by vector geometries.
Importantly it also shows how to 'polygonize' rasters and 'rasterize' vector datasets, making the two data models more interchangeable.

## Geometric operations on vector data {#geo-vec}

This section is about operations that in some way change the geometry of vector (`sf`) objects.
It is more advanced than the spatial data operations presented in the previous chapter (in Section \@ref(spatial-vec)), because here we drill down into the geometry:
the functions discussed in this section work on objects of class `sfc` in addition to objects of class `sf`.

### Simplification

\index{vector!simplification} 
Simplification is a process for generalization of vector objects (lines and polygons) usually for use in smaller scale maps.
Another reason for simplifying objects is to reduce the amount of memory, disk space and network bandwidth they consume:
it may be wise to simplify complex geometries before publishing them as interactive maps. 
The **sf** package provides `st_simplify()`, which uses the GEOS implementation of the Douglas-Peucker algorithm to reduce the vertex count.
`st_simplify()` uses the `dTolerance` to control the level of generalization in map units [see @douglas_algorithms_1973 for details].
Figure \@ref(fig:seine-simp) illustrates simplification of a `LINESTRING` geometry representing the river Seine and tributaries.
The simplified geometry was created by the following command:

```{r 05-geometry-operations-2}
seine_simp = st_simplify(seine, dTolerance = 2000)  # 2000 m
```

```{r seine-simp, echo=FALSE, fig.cap="Comparison of the original and simplified geometry of the seine object.", warning=FALSE, fig.scap="Simplification in action.", message=FALSE, fig.asp=0.5}
library(tmap)
p_simp1 = tm_shape(seine) + tm_lines() +
  tm_title("Original data")
p_simp2 = tm_shape(seine_simp) + tm_lines() +
  tm_title("st_simplify")
tmap_arrange(p_simp1, p_simp2, ncol = 2)
```

The resulting `seine_simp` object is a copy of the original `seine` but with fewer vertices.
This is apparent, with the result being visually simpler (Figure \@ref(fig:seine-simp), right) and consuming less memory than the original object, as verified below:

```{r 05-geometry-operations-3}
object.size(seine)
object.size(seine_simp)
```

\index{vector!simplification} 
Simplification is also applicable for polygons.
This is illustrated using `us_states`, representing the contiguous United States.
As we show in Chapter \@ref(reproj-geo-data), GEOS assumes that the data is in a projected CRS and this could lead to unexpected results when using a geographic CRS.
Therefore, the first step is to project the data into some adequate projected CRS, such as US National Atlas Equal Area (EPSG = 2163) (on the top left in Figure \@ref(fig:us-simp)):

```{r 05-geometry-operations-4}
us_states2163 = st_transform(us_states, "EPSG:2163")
```

`st_simplify()` works equally well with projected polygons:

```{r 05-geometry-operations-5}
us_states_simp1 = st_simplify(us_states2163, dTolerance = 100000)  # 100 km
```

A limitation with `st_simplify()` is that it simplifies objects on a per-geometry basis.
This means the 'topology' is lost, resulting in overlapping and 'holey' areal units illustrated in Figure \@ref(fig:us-simp) (right top panel).
`ms_simplify()` from **rmapshaper** provides an alternative that overcomes this issue.
By default it uses the Visvalingam algorithm, which overcomes some limitations of the Douglas-Peucker algorithm [@visvalingam_line_1993].
<!-- https://bost.ocks.org/mike/simplify/ -->
The following code chunk uses this function to simplify `us_states2163`.
The result has only 1% of the vertices of the input (set using the argument `keep`) but its number of objects remains intact because we set `keep_shapes = TRUE`:^[
Simplification of multipolygon objects can remove small internal polygons, even if the `keep_shapes` argument is set to TRUE. To prevent this, you need to set `explode = TRUE`. This option converts all mutlipolygons into separate polygons before its simplification.
]

```{r 05-geometry-operations-6, warning=FALSE, message=FALSE}
# proportion of points to retain (0-1; default 0.05)
us_states_simp2 = rmapshaper::ms_simplify(us_states2163, keep = 0.01,
                                          keep_shapes = TRUE)
```

\index{vector!simplification}
An alternative to simplification is smoothing the boundaries of polygon and linestring geometries, which is implemented in the **smoothr** package\index{smoothr (package)}. 
Smoothing interpolates the edges of geometries and does not necessarily lead to fewer vertices, but can be especially useful when working with geometries that arise from spatially vectorizing a raster (a topic covered in Chapter \@ref(raster-vector)).
**smoothr** implements three techniques for smoothing: a Gaussian kernel regression, Chaikin's corner cutting algorithm, and spline interpolation, which are all described in the package vignette and [website](https://strimas.com/smoothr/). 
Note that similar to `st_simplify()`, the smoothing algorithms don't preserve 'topology'.
The workhorse function of **smoothr** is `smooth()`, where the `method` argument specifies what smoothing technique to use.
Below is an example of using Gaussian kernel regression to smooth the borders of US states by using `method=ksmooth`.
The `smoothness` argument controls the bandwidth of the Gaussian that is used to smooth the geometry and has a default value of 1.

```{r 05-geometry-operations-6b, warning=FALSE}
us_states_simp3 = smoothr::smooth(us_states2163, 
                                  method = "ksmooth", smoothness = 6)
```

Finally, the visual comparison of the original dataset with the simplified and smoothed versions is shown in (Figure \@ref(fig:us-simp)). 
Differences can be observed between the outputs of the Douglas-Peucker (`st_simplify`), Visvalingam (`ms_simplify`), and Gaussian kernel regression (`smooth(method=ksmooth`) algorithms.

```{r us-simp, echo=FALSE, fig.cap="Polygon simplification in action, comparing the original geometry of the contiguous United States with simplified versions, generated with functions from sf (top-right), rmapshaper (bottom-left), and smoothr (bottom-right) packages.", warning=FALSE, fig.scap="Polygon simplification in action."}
library(tmap)
p_ussimp1 = tm_shape(us_states2163) + tm_polygons() + tm_title("Original data")
p_ussimp2 = tm_shape(us_states_simp1) + tm_polygons() + tm_title("st_simplify")
p_ussimp3 = tm_shape(us_states_simp2) + tm_polygons() + tm_title("ms_simplify")
p_ussimp4 = tm_shape(us_states_simp3) + tm_polygons() + tm_title('smooth(method = "ksmooth")')
tmap_arrange(p_ussimp1, p_ussimp2, p_ussimp3, p_ussimp4,
             ncol = 2, nrow = 2)
```

### Centroids

\index{vector!centroids} 
Centroid operations identify the center of geographic objects.
Like statistical measures of central tendency (including mean and median definitions of 'average'), there are many ways to define the geographic center of an object.
All of them create single point representations of more complex vector objects.

The most commonly used centroid operation is the *geographic centroid*.
This type of centroid operation (often referred to as 'the centroid') represents the center of mass in a spatial object (think of balancing a plate on your finger).
Geographic centroids have many uses, for example to create a simple point representation of complex geometries, or to estimate distances between polygons.
They can be calculated with the **sf** function `st_centroid()` as demonstrated in the code below, which generates the geographic centroids of regions in New Zealand and tributaries to the River Seine, illustrated with black points in Figure \@ref(fig:centr).

```{r 05-geometry-operations-7, warning=FALSE}
nz_centroid = st_centroid(nz)
seine_centroid = st_centroid(seine)
```

Sometimes the geographic centroid falls outside the boundaries of their parent objects (think of a doughnut).
In such cases *point on surface* operations can be used to guarantee the point will be in the parent object (e.g., for labeling irregular multipolygon objects such as island states), as illustrated by the red points in Figure \@ref(fig:centr).
Notice that these red points always lie on their parent objects.
They were created with `st_point_on_surface()` as follows:^[
A description of how `st_point_on_surface()` works is provided at https://gis.stackexchange.com/a/76563/20955.
]

```{r 05-geometry-operations-8, warning=FALSE}
nz_pos = st_point_on_surface(nz)
seine_pos = st_point_on_surface(seine)
```

```{r centr, warning=FALSE, echo=FALSE, fig.cap="Centroids (black points) and 'points on surface' (red points) of New Zealand's regions (left) and the Seine (right) datasets.", fig.scap="Centroid vs point on surface operations."}
p_centr1 = tm_shape(nz) + tm_borders() +
  tm_shape(nz_centroid) + tm_symbols(shape = 1, col = "black", size = 0.5) +
  tm_shape(nz_pos) + tm_symbols(shape = 1, col = "red", size = 0.5)  
p_centr2 = tm_shape(seine) + tm_lines() +
  tm_shape(seine_centroid) + tm_symbols(shape = 1, col = "black", size = 0.5) +
  tm_shape(seine_pos) + tm_symbols(shape = 1, col = "red", size = 0.5)  
tmap_arrange(p_centr1, p_centr2, ncol = 2)
```

Other types of centroids exist, including the *Chebyshev center* and the *visual center*.
We will not explore these here but it is possible to calculate them using R, as we'll see in Chapter \@ref(algorithms).

### Buffers

\index{vector!buffers} 
Buffers are polygons representing the area within a given distance of a geometric feature:
regardless of whether the input is a point, line or polygon, the output is a polygon.
Unlike simplification (which is often used for visualization and reducing file size) buffering tends to be used for geographic data analysis.
How many points are within a given distance of this line?
Which demographic groups are within travel distance of this new shop?
These kinds of questions can be answered and visualized by creating buffers around the geographic entities of interest.

Figure \@ref(fig:buffs) illustrates buffers of different sizes (5 and 50 km) surrounding the river Seine and tributaries.
These buffers were created with commands below, which show that the command `st_buffer()` requires at least two arguments: an input geometry and a distance, provided in the units of the CRS (in this case meters):

```{r 05-geometry-operations-9}
seine_buff_5km = st_buffer(seine, dist = 5000)
seine_buff_50km = st_buffer(seine, dist = 50000)
```

```{r buffs, echo=FALSE, fig.cap="Buffers around the Seine dataset of 5 km (left) and 50 km (right). Note the colors, which reflect the fact that one buffer is created per geometry feature.", fig.show='hold', out.width="75%", fig.scap="Buffers around the seine dataset."}
p_buffs1 = tm_shape(seine_buff_5km) + tm_polygons(fill =   "name") +
  tm_shape(seine) + tm_lines() +
  tm_title("5 km buffer") +
  tm_layout(legend.show = FALSE)
p_buffs2 = tm_shape(seine_buff_50km) + tm_polygons(fill =   "name") +
  tm_shape(seine) + tm_lines() +
  tm_title("50 km buffer") +
  tm_layout(legend.show = FALSE)
tmap_arrange(p_buffs1, p_buffs2, ncol = 2)  
```

```{block2 05-geometry-operations-10, type = "rmdnote"}
The third and final argument of `st_buffer()` is `nQuadSegs`, which means 'number of segments per quadrant' and is set by default to 30 (meaning circles created by buffers are composed of $4 \times 30 = 120$ lines).
This argument rarely needs to be set.
Unusual cases where it may be useful include when the memory consumed by the output of a buffer operation is a major concern (in which case it should be reduced) or when very high precision is needed (in which case it should be increased).
```

```{r nQuadSegs, eval=FALSE, echo=FALSE}
# Demonstrate nQuadSegs
seine_buff_simple = st_buffer(seine, dist = 50000, nQuadSegs = 3)
plot(seine_buff_simple, key.pos = NULL, main = "50 km buffer")
plot(seine, key.pos = NULL, lwd = 3, pal = rainbow, add = TRUE)
seine_points = st_cast(seine[1, ], "POINT")
buff_single = st_buffer(seine_points[1, ], 50000, 2)
buff_points = st_cast(buff_single, "POINT")
plot(st_geometry(buff_single), add = TRUE)
```

### Affine transformations

\index{vector!affine transformation} 
Affine transformation is any transformation that preserves lines and parallelism.
However, angles or length are not necessarily preserved.
Affine transformations include, among others, shifting (translation), scaling and rotation.
Additionally, it is possible to use any combination of these.
Affine transformations are an essential part of geocomputation.
For example, shifting is needed for labels placement, scaling is used in non-contiguous area cartograms (see Section \@ref(other-mapping-packages)), and many affine transformations are applied when reprojecting or improving the geometry that was created based on a distorted or wrongly projected map.
The **sf** package implements affine transformation for objects of classes `sfg` and `sfc`.

```{r 05-geometry-operations-11}
nz_sfc = st_geometry(nz)
```

Shifting moves every point by the same distance in map units.
It could be done by adding a numerical vector to a vector object.
For example, the code below shifts all y-coordinates by 100,000 meters to the north, but leaves the x-coordinates untouched (left panel of Figure \@ref(fig:affine-trans)). 

```{r 05-geometry-operations-12}
nz_shift = nz_sfc + c(0, 100000)
```

Scaling enlarges or shrinks objects by a factor.
It can be applied either globally or locally.
Global scaling increases or decreases all coordinates values in relation to the origin coordinates, while keeping all geometries topological relations intact.
It can be done by subtraction or multiplication of a`sfg` or `sfc` object.

```{r 05-geometry-operations-13, echo=FALSE,eval=FALSE}
nz_scale0 = nz_sfc * 0.5
```

Local scaling treats geometries independently and requires points around which geometries are going to be scaled, e.g., centroids.
In the example below, each geometry is shrunk by a factor of two around the centroids (middle panel in Figure \@ref(fig:affine-trans)).
To achieve that, each object is firstly shifted in a way that its center has coordinates of `0, 0` (`(nz_sfc - nz_centroid_sfc)`). 
Next, the sizes of the geometries are reduced by half (`* 0.5`).
Finally, each object's centroid is moved back to the input data coordinates (`+ nz_centroid_sfc`). 

```{r 05-geometry-operations-14}
nz_centroid_sfc = st_centroid(nz_sfc)
nz_scale = (nz_sfc - nz_centroid_sfc) * 0.5 + nz_centroid_sfc
```

Rotation of two-dimensional coordinates requires a rotation matrix:

$$
R =
\begin{bmatrix}
\cos \theta & -\sin \theta \\  
\sin \theta & \cos \theta \\
\end{bmatrix}
$$

It rotates points in a clockwise direction.
The rotation matrix can be implemented in R as:

```{r 05-geometry-operations-15}
rotation = function(a){
  r = a * pi / 180 #degrees to radians
  matrix(c(cos(r), sin(r), -sin(r), cos(r)), nrow = 2, ncol = 2)
} 
```

The `rotation` function accepts one argument `a` - a rotation angle in degrees.
Rotation could be done around selected points, such as centroids (right panel of Figure \@ref(fig:affine-trans)).
See `vignette("sf3")` for more examples.

```{r 05-geometry-operations-16}
nz_rotate = (nz_sfc - nz_centroid_sfc) * rotation(30) + nz_centroid_sfc
```

```{r affine-trans, echo=FALSE, fig.cap="Illustrations of affine transformations: shift, scale and rotate.", warning=FALSE, eval=TRUE, fig.scap="Illustrations of affine transformations."}
st_crs(nz_shift) = st_crs(nz_sfc)
st_crs(nz_scale) = st_crs(nz_sfc)
st_crs(nz_rotate) = st_crs(nz_sfc)
p_at1 = tm_shape(nz_sfc) + tm_polygons() +
  tm_shape(nz_shift) + tm_polygons(fill = "red") +
  tm_title("Shift")
p_at2 = tm_shape(nz_sfc) + tm_polygons() +
  tm_shape(nz_scale) + tm_polygons(fill = "red") +
  tm_title("Scale")
p_at3 = tm_shape(nz_sfc) + tm_polygons() +
  tm_shape(nz_rotate) + tm_polygons(fill = "red") +
  tm_title("Rotate")
tmap_arrange(p_at1, p_at2, p_at3, ncol = 3)
```

```{r 05-geometry-operations-17, echo=FALSE,eval=FALSE}
nz_scale_rotate = (nz_sfc - nz_centroid_sfc) * 0.25 * rotation(90) + nz_centroid_sfc
```
```{r 05-geometry-operations-18, echo=FALSE,eval=FALSE}
shearing = function(hx, hy){
  matrix(c(1, hy, hx, 1), nrow = 2, ncol = 2)
}
nz_shear = (nz_sfc - nz_centroid_sfc) * shearing(1.1, 0) + nz_centroid_sfc
```
```{r 05-geometry-operations-19, echo=FALSE,eval=FALSE}
plot(nz_sfc)
plot(nz_shear, add = TRUE, col = "red")
```

Finally, the newly created geometries can replace the old ones with the `st_set_geometry()` function: 

```{r 05-geometry-operations-20}
nz_scale_sf = st_set_geometry(nz, nz_scale)
```

### Clipping {#clipping}

\index{vector!clipping} 
\index{spatial!subsetting} 
Spatial clipping is a form of spatial subsetting that involves changes to the `geometry` columns of at least some of the affected features.

Clipping can only apply to features more complex than points: 
lines, polygons and their 'multi' equivalents.
To illustrate the concept we will start with a simple example:
two overlapping circles with a center point one unit away from each other and a radius of one (Figure \@ref(fig:points)).

```{r points, fig.cap="Overlapping circles.", fig.asp=0.4, crop = TRUE, echo=-1}
op = par(mar = rep(0, 4))
b = st_sfc(st_point(c(0, 1)), st_point(c(1, 1))) # create 2 points
b = st_buffer(b, dist = 1) # convert points to circles
plot(b, border = "grey")
text(x = c(-0.5, 1.5), y = 1, labels = c("x", "y"), cex = 3) # add text
```

```{r, echo=FALSE}
par(op)
```

Imagine you want to select not one circle or the other, but the space covered by both `x` *and* `y`.
This can be done using the function `st_intersection()`\index{vector!intersection}, illustrated using objects named `x` and `y` which represent the left- and right-hand circles (Figure \@ref(fig:circle-intersection)).

```{r circle-intersection, fig.cap="Overlapping circles with a gray color indicating intersection between them.", fig.asp=0.4, fig.scap="Overlapping circles showing intersection types.", crop = TRUE, echo=-1}
op = par(mar = rep(0, 4))
x = b[1]
y = b[2]
x_and_y = st_intersection(x, y)
plot(b, border = "grey")
plot(x_and_y, col = "lightgrey", border = "grey", add = TRUE) # intersecting area
```

```{r, echo=FALSE}
par(op)
```

The subsequent code chunk demonstrates how this works for all combinations of the 'Venn' diagram representing `x` and `y`, inspired by [Figure 5.1](http://r4ds.had.co.nz/transform.html#logical-operators) of the book *R for Data Science* [@grolemund_r_2016].

```{r venn-clip, echo=FALSE, fig.cap="Spatial equivalents of logical operators.", warning=FALSE}
source("code/05-venn-clip.R")
```

### Subsetting and clipping

\index{vector!clipping} 
\index{spatial!subsetting} 
Clipping objects can change their geometry but it can also subset objects, returning only features that intersect (or partly intersect) with a clipping/subsetting object.
To illustrate this point, we will subset points that cover the bounding box of the circles `x` and `y` in Figure \@ref(fig:venn-clip).
Some points will be inside just one circle, some will be inside both and some will be inside neither.
`st_sample()` is used below to generate a *simple random* distribution of points within the extent of circles `x` and `y`, resulting in output illustrated in Figure \@ref(fig:venn-subset), raising the question: how to subset the points to only return the point that intersects with *both* `x` and `y`?

```{r venn-subset, fig.cap="Randomly distributed points within the bounding box enclosing circles x and y. The point that intersects with both objects x and y is highlighted.", fig.height=6, fig.width=9, fig.scap="Randomly distributed points within the bounding box. Note that only one point intersects with both x and y, highlighted with a red circle.", echo=FALSE, echo=-1}
op = par(mar = rep(0, 4))
bb = st_bbox(st_union(x, y))
box = st_as_sfc(bb)
set.seed(2017)
p = st_sample(x = box, size = 10)
p_xy1 = p[x_and_y]
plot(box, border = "grey", lty = 2)
plot(x, add = TRUE, border = "grey")
plot(y, add = TRUE, border = "grey")
plot(p, add = TRUE)
plot(p_xy1, cex = 3, col = "red", add = TRUE)
text(x = c(-0.5, 1.5), y = 1, labels = c("x", "y"), cex = 2)
```

```{r, echo=FALSE}
par(op)
```

```{r venn-subset-to-show, eval=FALSE}
bb = st_bbox(st_union(x, y))
box = st_as_sfc(bb)
set.seed(2017)
p = st_sample(x = box, size = 10)
x_and_y = st_intersection(x, y)
```

The code chunk below demonstrates three ways to achieve the same result.
We can use the intersection\index{vector!intersection} of `x` and `y` (represented by `x_and_y` in the previous code chunk) as a subsetting object directly, as shown in the first line in the code chunk below.
We can also find the *intersection* between the input points represented by `p` and the subsetting/clipping object `x_and_y`, as demonstrated in the second line in the code chunk below.
This second approach will return features that partly intersect with `x_and_y` but with modified geometries for spatially extensive features that cross the border of the subsetting object.
The third approach is to create a subsetting object using the binary spatial predicate `st_intersects()`, introduced in the previous chapter.
The results are identical (except superficial differences in attribute names), but the implementation differs substantially:

```{r 05-geometry-operations-21}
p_xy1 = p[x_and_y] # way #1
p_xy2 = st_intersection(p, x_and_y) # way #2
sel_p_xy = st_intersects(p, x, sparse = FALSE)[, 1] & 
  st_intersects(p, y, sparse = FALSE)[, 1] # way #3
p_xy3 = p[sel_p_xy]
```

```{r 05-geometry-operations-22, echo=FALSE, eval=FALSE}
# test if objects are identical:
identical(p_xy1, p_xy2)
identical(p_xy2, p_xy3)
identical(p_xy1, p_xy3)
waldo::compare(p_xy1, p_xy2) # the same except attribute names
waldo::compare(p_xy2, p_xy3) # the same except attribute names

# An alternative way to sample from the bb
bb = st_bbox(st_union(x, y))
pmulti = st_multipoint(pmat)
box = st_convex_hull(pmulti)
```

Although the example above is rather contrived and provided for educational rather than applied purposes, and we encourage the reader to reproduce the results to deepen your understanding for handling geographic vector objects in R, it raises an important question: which implementation to use?
Generally, more concise implementations should be favored, meaning the first approach above.
We will return to the question of choosing between different implementations of the same technique or algorithm in Chapter \@ref(algorithms).

### Geometry unions

\index{vector!union} 
\index{aggregation!spatial} 
As we saw in Section \@ref(vector-attribute-aggregation), spatial aggregation can silently dissolve the geometries of touching polygons in the same group.
This is demonstrated in the code chunk below in which 49 `us_states` are aggregated into four regions using base and **dplyr**\index{dplyr (package)} functions (see results in Figure \@ref(fig:us-regions)):

```{r 05-geometry-operations-23}
regions = aggregate(x = us_states[, "total_pop_15"], by = list(us_states$REGION),
                    FUN = sum, na.rm = TRUE)
regions2 = us_states |> 
  group_by(REGION) |>
  summarize(pop = sum(total_pop_15, na.rm = TRUE))
```

```{r 05-geometry-operations-24, echo=FALSE}
# st_join(buff, africa[, "pop"]) |>
#   summarize(pop = sum(pop, na.rm = TRUE))
# summarize(africa[buff, "pop"], pop = sum(pop, na.rm = TRUE))
```

```{r us-regions, fig.cap="Spatial aggregation on contiguous polygons, illustrated by aggregating the population of US states into regions, with population represented by color. Note the operation automatically dissolves boundaries between states.", echo=FALSE, warning=FALSE, out.width="100%", fig.scap="Spatial aggregation on contiguous polygons."}
source("code/05-us-regions.R", print.eval = TRUE)
```

What is going on in terms of the geometries?
Behind the scenes, both `aggregate()` and `summarize()` combine the geometries and dissolve the boundaries between them using `st_union()`.
This is demonstrated in the code chunk below which creates a united western US: 

```{r 05-geometry-operations-25}
us_west = us_states[us_states$REGION == "West", ]
us_west_union = st_union(us_west)
```

The function can take two geometries and unite them, as demonstrated in the code chunk below which creates a united western block incorporating Texas (challenge: reproduce and plot the result):

```{r 05-geometry-operations-26, message=FALSE}
texas = us_states[us_states$NAME == "Texas", ]
texas_union = st_union(us_west_union, texas)
```

```{r 05-geometry-operations-27, echo=FALSE, eval=FALSE}
plot(texas_union)
# aim: experiment with st_union
us_south2 = st_union(us_west[1, ], us_west[6, ])
plot(us_southhwest)
```

### Type transformations {#type-trans}

\index{vector!geometry casting} 
Geometry casting is a powerful operation that enables transformation of the geometry type.
It is implemented in the `st_cast()` function from the **sf** package.
Importantly, `st_cast()` behaves differently on single simple feature geometry (`sfg`) objects, simple feature geometry column (`sfc`) and simple features objects.

Let's create a multipoint to illustrate how geometry casting works on simple feature geometry (`sfg`) objects:

```{r 05-geometry-operations-28}
multipoint = st_multipoint(matrix(c(1, 3, 5, 1, 3, 1), ncol = 2))
```

In this case, `st_cast()` can be useful to transform the new object into a linestring or a polygon (Figure \@ref(fig:single-cast)):

```{r 05-geometry-operations-29}
linestring = st_cast(multipoint, "LINESTRING")
polyg = st_cast(multipoint, "POLYGON")
```

```{r single-cast, echo = FALSE, fig.cap="Examples of a linestring and a polygon casted from a multipoint geometry.", warning=FALSE, fig.asp=0.3, fig.scap="Examples of casting operations."}
p_sc1 = tm_shape(st_sfc(multipoint, crs = "+proj=merc")) + tm_symbols(shape = 1,
                                                  col = "black",
                                                  size = 0.5) +
  tm_title("MULTIPOINT") +
  tm_layout(inner.margins = c(0.05, 0.05, 0.05, 0.05))
p_sc2 = tm_shape(st_sfc(linestring, crs = "+proj=merc")) + tm_lines() +
  tm_title("LINESTRING")
p_sc3 = tm_shape(st_sfc(polyg, crs = "+proj=merc")) + tm_polygons(border.col = "black") +
  tm_title("POLYGON")
tmap_arrange(p_sc1, p_sc2, p_sc3, ncol = 3)
```

Conversion from multipoint to linestring is a common operation that creates a line object from ordered point observations, such as GPS measurements or geotagged media.
This, in turn, allows to perform spatial operations such as the calculation of the length of the path traveled.
Conversion from multipoint or linestring to polygon is often used to calculate an area, for example from the set of GPS measurements taken around a lake or from the corners of a building lot.

The transformation process can be also reversed using `st_cast()`:

```{r 05-geometry-operations-30}
multipoint_2 = st_cast(linestring, "MULTIPOINT")
multipoint_3 = st_cast(polyg, "MULTIPOINT")
all.equal(multipoint, multipoint_2)
all.equal(multipoint, multipoint_3)
```

```{block2 05-geometry-operations-31, type='rmdnote'}
For single simple feature geometries (`sfg`), `st_cast()` also provides geometry casting from non-multi-types to multi-types (e.g., `POINT` to `MULTIPOINT`) and from multi-types to non-multi-types.
However, when casting from multi-types to non-multi-types only the first element of the old object would remain in the output object.
```

```{r 05-geometry-operations-32, include=FALSE}
cast_all = function(xg) {
  lapply(c("MULTIPOLYGON", "MULTILINESTRING", "MULTIPOINT", "POLYGON", "LINESTRING", "POINT"), 
         function(x) st_cast(xg, x))
}
t = cast_all(multipoint)
t2 = cast_all(polyg)
```

Geometry casting of simple features geometry column (`sfc`) and simple features objects works the same as for `sfg` in most of the cases. 
One important difference is the conversion between multi-types to non-multi-types.
As a result of this process, multi-objects of `sfc` or `sf` are split into many non-multi-objects.

Table \@ref(tab:sfs-st-cast) shows possible geometry type transformations on simple feature objects.
Single simple feature geometries (represented by the first column in the table) can be transformed into multiple geometry types, represented by the columns in Table \@ref(tab:sfs-st-cast).
Some transformations are not possible: you cannot convert a single point into a multilinestring or a polygon, for example, explaining why the cells `[1, 4:5]` in the table contain NA.
Some transformations split single features input into multiple sub-features, 'expanding' `sf` objects (adding new rows with duplicate attribute values).
When a multipoint geometry consisting of five pairs of coordinates is tranformed into a 'POINT' geometry, for example, the output will contain five features.

```{r sfs-st-cast, echo=FALSE}
sfs_st_cast = read.csv("extdata/sfs-st-cast.csv")
abbreviate_geomtypes = function(geomtypes) {
  geomtypes_new = gsub(pattern = "POINT", replacement = "POI", x = geomtypes)
  geomtypes_new = gsub(pattern = "POLYGON", replacement = "POL", x = geomtypes_new)
  geomtypes_new = gsub(pattern = "LINESTRING", replacement = "LIN", x = geomtypes_new)
  geomtypes_new = gsub(pattern = "MULTI", replacement = "M", x = geomtypes_new)
  geomtypes_new = gsub(pattern = "GEOMETRYCOLLECTION", replacement = "GC", x = geomtypes_new)
  geomtypes_new
}
sfs_st_cast$input_geom = abbreviate_geomtypes(sfs_st_cast$input_geom)
names(sfs_st_cast) = abbreviate_geomtypes(names(sfs_st_cast))
names(sfs_st_cast)[1] = ""
knitr::kable(sfs_st_cast, 
             caption = paste("Geometry casting on simple feature geometries", 
                             "(see Section 2.1) with input type by row and", 
                             "output type by column"),
             caption.short = "Geometry casting on simple feature geometries.",
             booktabs = TRUE) |> 
  kableExtra::add_footnote("Note: Values like (1) represent the number of features; NA means the operation is not possible. Abbreviations: POI, LIN, POL and GC refer to POINT, LINESTRING, POLYGON and GEOMETRYCOLLECTION. The MULTI version of these geometry types is indicated by a preceding M, e.g., MPOI is the acronym for MULTIPOINT.", notation = "none")
```

Let's try to apply geometry type transformations on a new object, `multilinestring_sf`, as an example (on the left in Figure \@ref(fig:line-cast)):

```{r 05-geometry-operations-33}
multilinestring_list = list(matrix(c(1, 4, 5, 3), ncol = 2), 
                            matrix(c(4, 4, 4, 1), ncol = 2),
                            matrix(c(2, 4, 2, 2), ncol = 2))
multilinestring = st_multilinestring(multilinestring_list)
multilinestring_sf = st_sf(geom = st_sfc(multilinestring))
multilinestring_sf
```

You can imagine it as a road or river network. 
The new object has only one row that defines all the lines.
This restricts the number of operations that can be done, for example it prevents adding names to each line segment or calculating lengths of single lines.
The `st_cast()` function can be used in this situation, as it separates one mutlilinestring into three linestrings:

```{r 05-geometry-operations-34}
linestring_sf2 = st_cast(multilinestring_sf, "LINESTRING")
linestring_sf2
```

```{r line-cast, echo=FALSE, fig.cap="Examples of type casting between MULTILINESTRING (left) and LINESTRING (right).", warning=FALSE, fig.scap="Examples of type casting."}
p_lc1 = tm_shape(multilinestring_sf) + tm_lines(lwd = 3) +
  tm_title("MULTILINESTRING")
linestring_sf2$name = c("Riddle Rd", "Marshall Ave", "Foulke St")
p_lc2 = tm_shape(linestring_sf2) + tm_lines(lwd = 3, col = "name", col.scale = tm_scale(values = "Set2")) +
  tm_title("LINESTRING") +
  tm_layout(legend.show = FALSE)
tmap_arrange(p_lc1, p_lc2, ncol = 2)
```

The newly created object allows for attributes creation (see more in Section \@ref(vec-attr-creation)) and length measurements:

```{r 05-geometry-operations-35}
linestring_sf2$name = c("Riddle Rd", "Marshall Ave", "Foulke St")
linestring_sf2$length = st_length(linestring_sf2)
linestring_sf2
```

## Geometric operations on raster data {#geo-ras}

\index{raster!manipulation} 
Geometric raster operations include the shift, flipping, mirroring, scaling, rotation or warping of images.
These operations are necessary for a variety of applications including georeferencing, used to allow images to be overlaid on an accurate map with a known CRS [@liu_essential_2009].
A variety of georeferencing techniques exist, including:

- Georectification based on known [ground control points](https://www.qgistutorials.com/en/docs/3/georeferencing_basics.html)
- Orthorectification, which also accounts for local topography
- Image [registration](https://en.wikipedia.org/wiki/Image_registration) is used to combine images of the same thing but shot from different sensors by aligning one image with another (in terms of coordinate system and resolution)

R is rather unsuitable for the first two points since these often require manual intervention which is why they are usually done with the help of dedicated GIS software (see also Chapter \@ref(gis)).
On the other hand, aligning several images is possible in R and this section shows among others how to do so.
This often includes changing the extent, the resolution and the origin of an image.
A matching projection is of course also required but is already covered in Section \@ref(reproj-ras).

In any case, there are other reasons to perform a geometric operation on a single raster image.
For instance, in Chapter \@ref(location) we define metropolitan areas in Germany as 20 km^2^ pixels with more than 500,000 inhabitants. 
The original inhabitant raster, however, has a resolution of 1 km^2^ which is why we will decrease (aggregate) the resolution by a factor of 20 (see Section \@ref(define-metropolitan-areas)).
Another reason for aggregating a raster is simply to decrease run-time or save disk space.
Of course, this approach is only recommended if the task at hand allows a coarser resolution of raster data.

### Geometric intersections

\index{raster!intersection} 
In Section \@ref(spatial-raster-subsetting) we have shown how to extract values from a raster overlaid by other spatial objects.
To retrieve a spatial output, we can use almost the same subsetting syntax.
The only difference is that we have to make clear that we would like to keep the matrix structure by setting the `drop` argument to `FALSE`.
This will return a raster object containing the cells whose midpoints overlap with `clip`.

```{r 05-geometry-operations-36}
elev = rast(system.file("raster/elev.tif", package = "spData"))
clip = rast(xmin = 0.9, xmax = 1.8, ymin = -0.45, ymax = 0.45,
            resolution = 0.3, vals = rep(1, 9))
elev[clip, drop = FALSE]
```

For the same operation we can also use the `intersect()` and `crop()` command.

### Extent and origin

\index{raster!merging} 
When merging or performing map algebra on rasters, their resolution, projection, origin and/or extent have to match. Otherwise, how should we add the values of one raster with a resolution of 0.2 decimal degrees to a second raster with a resolution of 1 decimal degree?
The same problem arises when we would like to merge satellite imagery from different sensors with different projections and resolutions. 
We can deal with such mismatches by aligning the rasters.

In the simplest case, two images only differ with regard to their extent.
Following code adds one row and two columns to each side of the raster while setting all new values to `NA` (Figure \@ref(fig:extend-example)).

```{r extend-example0}
elev = rast(system.file("raster/elev.tif", package = "spData"))
elev_2 = extend(elev, c(1, 2))
```

```{r extend-example, fig.cap = "Original raster (left) and the same raster (right) extended by one row on the top and bottom and two columns on the left and right.", fig.scap="Extending rasters.", echo=FALSE, fig.asp=0.5}
source("code/05-extend-example.R", print.eval = TRUE)
```

Performing an algebraic operation on two objects with differing extents in R, the **terra** package returns an error.

```{r 05-geometry-operations-37, error=TRUE}
elev_3 = elev + elev_2
```

However, we can align the extent of two rasters with `extend()`. 
Instead of telling the function how many rows or columns should be added (as done before), we allow it to figure it out by using another raster object.
Here, we extend the `elev` object to the extent of `elev_2`. 
The values of the newly added rows and columns are set to `NA`.

```{r 05-geometry-operations-38}
elev_4 = extend(elev, elev_2)
```

\index{raster!origin} 
The origin of a raster is the cell corner closest to the coordinates (0, 0).
The `origin()` function returns the coordinates of the origin.
In the below example a cell corner exists with coordinates (0, 0), but that is not necessarily the case.

```{r 05-geometry-operations-39}
origin(elev_4)
```

If two rasters have different origins, their cells do not overlap completely which would make map algebra impossible.
To change the origin -- use `origin()`.^[
If the origins of two raster datasets are just marginally apart, it sometimes is sufficient to simply increase the `tolerance` argument  of `terra::terraOptions()`.
]
Figure \@ref(fig:origin-example) reveals the effect of changing the origin in this way.

```{r}
# change the origin
origin(elev_4) = c(0.25, 0.25)
```

```{r origin-example, fig.cap="Rasters with identical values but different origins.", echo=FALSE}
elev_poly = st_as_sf(as.polygons(elev, dissolve = FALSE))
elev4_poly = st_as_sf(as.polygons(elev_4, dissolve = FALSE))

tm_shape(elev4_poly) +
  tm_grid() +
  tm_polygons(fill = "elev", lwd = 0.5) +
  tm_shape(elev_poly) +
  tm_polygons(fill = "elev") +
  tm_layout(frame = FALSE, legend.show = FALSE,
            inner.margins = c(0.1, 0.12, 0, 0))
```

Note that changing the resolution (next section) frequently also changes the origin.

### Aggregation and disaggregation

\index{raster!aggregation} 
\index{raster!disaggregation} 
Raster datasets can also differ with regard to their resolution. 
To match resolutions, one can either decrease  (`aggregate()`) or increase (`disagg()`) the resolution of one raster.^[
Here we refer to spatial resolution.
In remote sensing the spectral (spectral bands), temporal (observations through time of the same area) and radiometric (color depth) resolution are also important.
Check out the `tapp()` example in the documentation for getting an idea on how to do temporal raster aggregation.
]
As an example, we here change the spatial resolution of `dem` (found in the **spDataLarge** package) by a factor of 5 (Figure \@ref(fig:aggregate-example)).
Additionally, the output cell value should correspond to the mean of the input cells (note that one could use other functions as well, such as `median()`, `sum()`, etc.):

```{r 05-geometry-operations-40}
dem = rast(system.file("raster/dem.tif", package = "spDataLarge"))
dem_agg = aggregate(dem, fact = 5, fun = mean)
```

```{r aggregate-example, fig.cap = "Original raster (left). Aggregated raster (right).", echo=FALSE}
p_ar1 = tm_shape(dem) +
  tm_raster(col.scale = tm_scale_continuous()) +
  tm_title("A. Original") + 
  tm_layout(frame = FALSE, legend.show = FALSE)
p_ar2 = tm_shape(dem_agg) +
  tm_raster(col.scale = tm_scale_continuous()) +
  tm_title("B. Aggregated") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tmap_arrange(p_ar1, p_ar2, ncol = 2)
```

\index{raster!disaggregation}
The `disagg()` function increases the resolution of raster objects. 
It comes with two methods on how to compute the values of the newly created cells: the default method (`method = "near"`) simply gives all output cells the value of the input cell, and hence duplicates values, which translates into a 'blocky' output.
The `bilinear` method uses the four nearest pixel centers of the input image (salmon colored points in Figure \@ref(fig:bilinear)) to compute an average weighted by distance (arrows in Figure \@ref(fig:bilinear).
The value of the output cell is represented by a square in the upper left corner in Figure \@ref(fig:bilinear)).

```{r 05-geometry-operations-41}
dem_disagg = disagg(dem_agg, fact = 5, method = "bilinear")
identical(dem, dem_disagg)
```

```{r bilinear, echo = FALSE, fig.width=8, fig.height=10, fig.cap="The distance-weighted average of the four closest input cells determine the output when using the bilinear method for disaggregation.", fig.scap="Bilinear disaggregation in action.", warning=FALSE}
source("code/05-bilinear.R", print.eval = TRUE)
# knitr::include_graphics("https://user-images.githubusercontent.com/1825120/146619205-3c0c2e3f-9e8b-4fda-b014-9c342a4befbb.png")
```

Comparing the values of `dem` and `dem_disagg` tells us that they are not identical (you can also use `compareGeom()` or `all.equal()`).
However, this was hardly to be expected, since disaggregating is a simple interpolation technique.
It is important to keep in mind that disaggregating results in a finer resolution; the corresponding values, however, are only as accurate as their lower resolution source.

### Resampling

\index{raster!resampling}
The above methods of aggregation and disaggregation are only suitable when we want to change the resolution of our raster by the aggregation/disaggregation factor. 
However, what to do when we have two or more rasters with different resolutions and origins?
This is the role of resampling -- a process of computing values for new pixel locations.
In short, this process takes the values of our original raster and recalculates new values for a target raster with custom resolution and origin (Figure \@ref(fig:resampl0)).

```{r resampl0, echo=FALSE, fig.cap="Resampling of an original (input) raster into a target raster with custom resolution and origin."}
target_rast = rast(xmin = 794650, xmax = 798250, 
                   ymin = 8931750, ymax = 8935350,
                   resolution = 150, crs = "EPSG:32717")
target_rast_p = st_as_sf(as.polygons(target_rast))
dem_resampl1 = resample(dem, target_rast, method = "near")

tm1 = tm_shape(dem) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("Original raster") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tm2 = tm_shape(dem) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("Target raster") +
  tm_layout(frame = FALSE, legend.show = FALSE) +
  tm_shape(target_rast_p) +
  tm_borders()
tm3 = tm_shape(dem) +
  tm_raster(col_alpha = 0) +
  tm_shape(dem_resampl1) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("Resampled raster") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tmap_arrange(tm1, tm2, tm3, nrow = 1)
```

\index{raster!resampling}
There are several methods for estimating values for a raster with different resolutions/origins, as shown in Figure \@ref(fig:resampl).
The main resampling methods include:

- Nearest neighbor: assigns the value of the nearest cell of the original raster to the cell of the target one. This is a fast simple technique that is usually suitable for resampling categorical rasters.
- Bilinear interpolation: assigns a weighted average of the four nearest cells from the original raster to the cell of the target one (Figure \@ref(fig:bilinear)). This is the fastest method that is appropriate for continuous rasters.
- Cubic interpolation: uses values of the 16 nearest cells of the original raster to determine the output cell value, applying third-order polynomial functions. Used for continuous rasters and results in a smoother surface compared to bilinear interpolation, but is computationally more demanding.
- Cubic spline interpolation: also uses values of the 16 nearest cells of the original raster to determine the output cell value, but applies cubic splines (piecewise third-order polynomial functions). Used for continuous rasters.
- Lanczos windowed sinc resampling: uses values of the 36 nearest cells of the original raster to determine the output cell value. Used for continuous rasters.^[
More detailed explanation of this method can be found at https://gis.stackexchange.com/a/14361/20955.
]

The above explanation highlights that only *nearest neighbor* resampling is suitable for categorical rasters, while all methods can be used (with different outcomes) for continuous rasters.
Please note also, that the methods gain both in complexity and processing time from top to bottom.

To apply resampling, the **terra** package provides a `resample()` function.
It accepts an input raster (`x`), a raster with target spatial properties (`y`), and a resampling method (`method`).

We need a raster with target spatial properties to see how the `resample()` function works.
For this example, we create `target_rast`, but you would often use an already existing raster object.

```{r 05-geometry-operations-42}
target_rast = rast(xmin = 794650, xmax = 798250, 
                   ymin = 8931750, ymax = 8935350,
                   resolution = 300, crs = "EPSG:32717")
```

Next, we need to provide our two raster objects as the first two arguments and one of the resampling methods described above.

```{r 05-geometry-operations-42b}
dem_resampl = resample(dem, y = target_rast, method = "bilinear")
```

Figure \@ref(fig:resampl) shows a comparison of different resampling methods on the `dem` object.

```{r resampl, echo=FALSE, fig.cap="Visual comparison of the original raster and five different resampling methods."}
dem_resampl1 = resample(dem, target_rast, method = "near")
dem_resampl2 = resample(dem, target_rast, method = "bilinear")
dem_resampl3 = resample(dem, target_rast, method = "cubic")
dem_resampl4 = resample(dem, target_rast, method = "cubicspline")
dem_resampl5 = resample(dem, target_rast, method = "lanczos")

library(tmap)
tm1 = tm_shape(dem) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("Original raster") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tm2 = tm_shape(dem_resampl1) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("near") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tm3 = tm_shape(dem_resampl2) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("bilinear") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tm4 = tm_shape(dem_resampl3) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("cubic") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tm5 = tm_shape(dem_resampl4) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("cubicspline") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tm6 = tm_shape(dem_resampl5) +
  tm_raster(col.scale = tm_scale(breaks = seq(200, 1100, by = 150))) +
  tm_title("lanczos") +
  tm_layout(frame = FALSE, legend.show = FALSE)
tmap_arrange(tm1, tm2, tm3, tm4, tm5, tm6)
```

The `resample()` function also has some additional resampling methods, including `sum`, `min`, `q1`, `med`, `q3`, `max`, `average`, `mode`, and `rms`.
All of them calculate a given statistic based on the values of all non-NA contributing grid cells.
For example, `sum` is useful when each raster cell represents a spatially extensive variable (e.g., number of people).
As an effect of using `sum`, the resampled raster should have the same total number of people as the original one.

\index{raster!resampling}
As you will see in section \@ref(reproj-ras), raster reprojection is a special case of resampling when our target raster has a different CRS than the original raster.

\index{GDAL}
```{block2 type='rmdnote'}
Most geometry operations in **terra** are user-friendly, rather fast, and work on large raster objects.
However, there could be some cases, when **terra** is not the most performant either for extensive rasters or many raster files, and some alternatives should be considered.

The most established alternatives come with the GDAL library.
It contains several utility functions, including:

- `gdalinfo` - lists various information about a raster file, including its resolution, CRS, bounding box, and more
- `gdal_translate` - converts raster data between different file formats
- `gdal_rasterize` - converts vector data into raster files
- `gdalwarp` - allows for raster mosaicing, resampling, cropping, and reprojecting

All of the above functions are written in C++, but can be called in R using `sf::gdal_utils()`, the **gdalUtilities** package or via system commands (see section \@ref(gdal)).
Importantly, all of these functions expect a raster file path as an input and often return their output as a raster file (for example, `gdalUtilities::gdal_translate("my_file.tif", "new_file.tif", t_srs = "EPSG:4326")`)
This is very different from the usual **terra** approach, which expects `SpatRaster` objects as inputs.
```

## Exercises

```{r, echo=FALSE, results='asis'}
res = knitr::knit_child('_05-ex.Rmd', quiet = TRUE, options = list(include = FALSE, eval = FALSE))
cat(res, sep = '\n')
```