This vignette explains the data
model of stars
objects, illustrated using artificial and
real datasets.
stars
objects consist of
dim
) attributedimensions
of class
dimensions
that carries dimension metadatastars
A dimensions
object is a named list of
dimension
elements, each describing the semantics a
dimension of the data arrays (space, time, type etc). In addition to
that, a dimensions
object has an attribute called
raster
of class stars_raster
, which is a named
list with three elements:
dimensions
length 2 character; the dimension names that
constitute a spatial raster (or NA)affine
length 2 numeric; the two affine parameters of
the geotransform (or NA)curvilinear
a boolean indicating whether a raster is a
curvilinear raster (or NA)The affine
and curvilinear
values are only
relevant in case of raster data, indicated by dimensions
to
have non-NA values.
A dimension
object describes a single
dimension; it is a list with named elements
from
: (numeric length 1): the start index of the
arrayto
: (numeric length 1): the end index of the arrayoffset
: (numeric length 1): the start coordinate (or
time) value of the first pixel (i.e., a pixel/cell boundary)delta
: (numeric length 1): the increment, or cell
sizerefsys
: (character, or crs
): object
describing the reference system; e.g. the PROJ string, or string
POSIXct
or PCICt
(for 360 and 365 days/year
calendars), or object of class crs
(containing both EPSG
code and proj4string)point
: (logical length 1): boolean indicating whether
cells/pixels refer to areas/periods, or to points/instances (may be
NA)values
: one of
NULL
(missing),POSIXct
,
PCICt
, or sfc
),intervals
(a list with two vectors,
start
and end
, with interval start- and
end-values), orfrom
and to
will usually be 1 and the
dimension size, but from
may be larger than 1 in case a
sub-grid got was selected (or cropped).
offset
and delta
only apply to
regularly discretized dimensions, and are NA
if
this is not the case. If they are NA
, dimension values may
be held in the values
field. Rectilinear and curvilinear
grids need grid values in values
that can be either:
Alternatively, values
can contains a set of spatial
geometries encoded in an sfc
vector (“list-column”), in
which case we have a vector data
cube.
With a very simple file created from a 4 × 5 matrix
suppressPackageStartupMessages(library(stars))
m = matrix(1:20, nrow = 5, ncol = 4)
dim(m) = c(x = 5, y = 4) # named dim
(s = st_as_stars(m))
## stars object with 2 dimensions and 1 attribute
## attribute(s):
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## A1 1 5.75 10.5 10.5 15.25 20
## dimension(s):
## from to offset delta point x/y
## x 1 5 0 1 FALSE [x]
## y 1 4 0 1 FALSE [y]
we see that
from
and to
fields of each dimension
define a range that corresponds to the array dimension:When we plot this object, using the image
method for
stars
objects,
we see that (0, 0) is the origin of the grid (grid corner), and 1 the coordinate value increase from one index (row, col) to the next. It means that consecutive matrix columns represent grid lines, going from south to north. Grids defined this way are regular: grid cell size is constant everywhere.
Many actual grid datasets have y coordinates (grid rows) going from
North to South (top to bottom); this is realised with a negative value
for delta
. We see that the grid origin (0, 0) did not change:
An example is the GeoTIFF carried in the package, which, as probably
all data sources read through GDAL, has a negative delta
for the y
-coordinate:
Dimension tables of stars
objects carry a
raster
attribute:
str(attr(st_dimensions(s), "raster"))
## List of 4
## $ affine : num [1:2] 0 0
## $ dimensions : chr [1:2] "x" "y"
## $ curvilinear: logi FALSE
## $ blocksizes : NULL
## - attr(*, "class")= chr "stars_raster"
which is a list that holds
dimensions
: character, the names of raster dimensions
(if any), as opposed to e.g. spectral, temporal or other dimensionsaffine
: numeric, the affine parameterscurvilinear
: a logical indicating whether the raster is
curvilinearThese fields are needed at this level, because they describe
properties of the array at a higher level than individual dimensions do:
a pair of dimensions forms a raster, both affine
and
curvilinear
describe how x and y as a pair are
derived from grid indexes (see below) when this cannot be done on a
per-dimension basis.
With two affine parameters a1 and a2, x and y coordinates are derived from (1-based) grid indexes i and j, grid offset values ox and oy, and grid cell sizes dx and dy by
x = ox + (i − 1)dx + (j − 1)a1
y = oy + (i − 1)a2 + (j − 1)dy Clearly, when a1 = a2 = 0, x and y are entirely derived from their respective index, offset and cellsize.
Note that for integer indexes, the coordinates are that of the starting edge of a grid cell; to get the grid cell center of the top left grid cell (in case of a negative dy), use i = 1.5 and j = 1.5.
We can rotate grids by setting a1 and a2 to a non-zero value:
attr(attr(s, "dimensions"), "raster")$affine = c(0.1, 0.1)
plot(st_as_sf(s, as_points = FALSE), axes = TRUE, nbreaks = 20)
The rotation angle, in degrees, is
Sheared grids are obtained when the two rotation coefficients, a1 and a2, are unequal:
attr(attr(s, "dimensions"), "raster")$affine = c(0.1, 0.2)
plot(st_as_sf(s, as_points = FALSE), axes = TRUE, nbreaks = 20)
Now, the y-axis and x-axis have different rotation in degrees of respectively
Rectilinear grids have orthogonal axes, but do not have congruent (equally sized and shaped) cells: each axis has its own irregular subdivision.
We can define a rectilinear grid by specifying the cell boundaries, meaning for every dimension we specify one more value than the dimension size:
x = c(0, 0.5, 1, 2, 4, 5) # 6 numbers: boundaries!
y = c(0.3, 0.5, 1, 2, 2.2) # 5 numbers: boundaries!
(r = st_as_stars(list(m = m), dimensions = st_dimensions(x = x, y = y)))
## stars object with 2 dimensions and 1 attribute
## attribute(s):
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## m 1 5.75 10.5 10.5 15.25 20
## dimension(s):
## from to point values x/y
## x 1 5 FALSE [0,0.5),...,[4,5) [x]
## y 1 4 FALSE [0.3,0.5),...,[2,2.2) [y]
st_bbox(r)
## xmin ymin xmax ymax
## 0.0 0.3 5.0 2.2
image(r, axes = TRUE, col = grey((1:20)/20))
Would we leave out the last value, than stars
may come
up with a different cell boundary for the last cell, as this is
now derived from the width of the one-but-last cell:
x = c(0, 0.5, 1, 2, 4) # 5 numbers: offsets only!
y = c(0.3, 0.5, 1, 2) # 4 numbers: offsets only!
(r = st_as_stars(list(m = m), dimensions = st_dimensions(x = x, y = y)))
## stars object with 2 dimensions and 1 attribute
## attribute(s):
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## m 1 5.75 10.5 10.5 15.25 20
## dimension(s):
## from to point values x/y
## x 1 5 FALSE [0,0.5),...,[4,6) [x]
## y 1 4 FALSE [0.3,0.5),...,[2,3) [y]
st_bbox(r)
## xmin ymin xmax ymax
## 0.0 0.3 6.0 3.0
This is not problematic if cells have a constant width, in which case
the boundaries are reduced to an offset
and
delta
value, irrespective whether an upper boundary is
given:
x = c(0, 1, 2, 3, 4) # 5 numbers: offsets only!
y = c(0.5, 1, 1.5, 2) # 4 numbers: offsets only!
(r = st_as_stars(list(m = m), dimensions = st_dimensions(x = x, y = y)))
## stars object with 2 dimensions and 1 attribute
## attribute(s):
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## m 1 5.75 10.5 10.5 15.25 20
## dimension(s):
## from to offset delta point x/y
## x 1 5 0 1 FALSE [x]
## y 1 4 0.5 0.5 FALSE [y]
st_bbox(r)
## xmin ymin xmax ymax
## 0.0 0.5 5.0 2.5
Alternatively, one can also set the cell midpoints by
specifying arguments cell_midpoints
to the
st_dimensions
call:
x = st_as_stars(matrix(1:9, 3, 3),
st_dimensions(x = c(1, 2, 3), y = c(2, 3, 10), cell_midpoints = TRUE))
When the dimension is regular, this results in offset
being shifted back with half a delta
, or else in intervals
derived from the distances between cell centers. This should obviously
not be done when cell boundaries are specified.
Curvilinear grids are grids whose grid lines are not straight. Rather than describing the curvature parametrically, the typical (HDF5 or NetCDF) files in which they are found have two raster layers with the longitudes and latitudes for every corresponding pixel of remaining layers.
As an example, we will use a Sentinel 5P dataset available from
package starsdata
; this package can be installed with
The dataset is found here:
(s5p = system.file("sentinel5p/S5P_NRTI_L2__NO2____20180717T120113_20180717T120613_03932_01_010002_20180717T125231.nc", package = "starsdata"))
## [1] ""
We can construct the curvilinear stars
raster by calling
read_stars
on the right sub-array:
For this array, we can see the GDAL metadata under item
GEOLOCATION
:
which reveals where, in this dataset, the longitude and latitude arrays are kept.
nit.c = read_stars(subs[[6]])
threshold = units::set_units(9e+36, mol/m^2)
nit.c[[1]][nit.c[[1]] > threshold] = NA
nit.c
The curvilinear array has the actual arrays (raster layers, matrices) with longitude and latitude values read in its dimension table. We can plot this file:
plot(nit.c, breaks = "equal", reset = FALSE, axes = TRUE, as_points = TRUE,
pch = 16, logz = TRUE, key.length = 1)
maps::map('world', add = TRUE, col = 'red')
plot(nit.c, breaks = "equal", reset = FALSE, axes = TRUE, as_points = FALSE,
border = NA, logz = TRUE, key.length = 1)
maps::map('world', add = TRUE, col = 'red')
We can downsample the data by
(nit.c_ds = stars:::st_downsample(nit.c, 8))
plot(nit.c_ds, breaks = "equal", reset = FALSE, axes = TRUE, as_points = TRUE,
pch = 16, logz = TRUE, key.length = 1)
maps::map('world', add = TRUE, col = 'red')
which doesn’t look nice, but plotting the cells as polygons looks better:
plot(nit.c_ds, breaks = "equal", reset = FALSE, axes = TRUE, as_points = FALSE,
border = NA, logz = TRUE, key.length = 1)
maps::map('world', add = TRUE, col = 'red')
Another approach would be to warp the curvilinear grid to a regular grid, e.g. by