Currently, there are 9 functions associated with the
sample
verb in the sgsR
package:
Algorithm | Description | Reference |
---|---|---|
sample_srs() |
Simple random | |
sample_systematic() |
Systematic | |
sample_strat() |
Stratified | Queinnec, White, & Coops (2021) |
sample_sys_strat() |
Systematic Stratified | |
sample_nc() |
Nearest centroid | Melville & Stone (2016) |
sample_clhs() |
Conditioned Latin hypercube | Minasny & McBratney (2006) |
sample_balanced() |
Balanced sampling | Grafström, A. Lisic, J (2018) |
sample_ahels() |
Adapted hypercube evaluation of a legacy sample | Malone, Minasny, & Brungard (2019) |
sample_existing() |
Sub-sampling an existing sample |
sample_srs
We have demonstrated a simple example of using the
sample_srs()
function in vignette("sgsR")
. We
will demonstrate additional examples below.
raster
The input required for sample_srs()
is a
raster
. This means that sraster
and
mraster
are supported for this function.
#--- perform simple random sampling ---#
sample_srs(
raster = sraster, # input sraster
nSamp = 200, # number of desired sample units
plot = TRUE
) # plot
#> Simple feature collection with 200 features and 0 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431110 ymin: 5337790 xmax: 438530 ymax: 5343210
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> geometry
#> 1 POINT (432130 5338050)
#> 2 POINT (437830 5341790)
#> 3 POINT (437310 5339490)
#> 4 POINT (432730 5340570)
#> 5 POINT (431670 5343070)
#> 6 POINT (433450 5340430)
#> 7 POINT (435010 5337910)
#> 8 POINT (431290 5338190)
#> 9 POINT (432350 5342410)
#> 10 POINT (432690 5338050)
sample_srs(
raster = mraster, # input mraster
nSamp = 200, # number of desired sample units
access = access, # define access road network
mindist = 200, # minimum distance sample units must be apart from one another
buff_inner = 50, # inner buffer - no sample units within this distance from road
buff_outer = 200, # outer buffer - no sample units further than this distance from road
plot = TRUE
) # plot
#> Simple feature collection with 200 features and 0 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431110 ymin: 5337710 xmax: 438490 ymax: 5343230
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> geometry
#> 1 POINT (433590 5342070)
#> 2 POINT (434630 5341790)
#> 3 POINT (436490 5341330)
#> 4 POINT (435150 5339590)
#> 5 POINT (437350 5342390)
#> 6 POINT (433550 5341590)
#> 7 POINT (434130 5341610)
#> 8 POINT (434710 5338930)
#> 9 POINT (432230 5341530)
#> 10 POINT (432590 5342970)
sample_systematic
The sample_systematic()
function applies systematic
sampling across an area with the cellsize
parameter
defining the resolution of the tessellation. The tessellation shape can
be modified using the square
parameter. Assigning
TRUE
(default) to the square
parameter results
in a regular grid and assigning FALSE
results in a
hexagonal grid.
The location of sample units can also be adjusted using the
locations
parameter, where centers
takes the
center, corners
takes all corners, and random
takes a random location within each tessellation. Random start points
and translations are applied when the function is called.
#--- perform grid sampling ---#
sample_systematic(
raster = sraster, # input sraster
cellsize = 1000, # grid distance
plot = TRUE
) # plot
#> Simple feature collection with 38 features and 0 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431487.6 ymin: 5337830 xmax: 438221.2 ymax: 5343217
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> geometry
#> 1 POINT (436900.3 5337882)
#> 2 POINT (437789.5 5338339)
#> 3 POINT (434664.4 5337856)
#> 4 POINT (435553.6 5338313)
#> 5 POINT (438221.2 5339686)
#> 6 POINT (432428.5 5337830)
#> 7 POINT (433317.7 5338287)
#> 8 POINT (434206.9 5338745)
#> 9 POINT (435096.1 5339202)
#> 10 POINT (435985.3 5339660)
#--- perform grid sampling ---#
sample_systematic(
raster = sraster, # input sraster
cellsize = 500, # grid distance
square = FALSE, # hexagonal tessellation
location = "random", # randomly sample within tessellation
plot = TRUE
) # plot
#> Simple feature collection with 169 features and 0 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431108.2 ymin: 5337770 xmax: 438521.8 ymax: 5343194
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> geometry
#> 1 POINT (438387.1 5343007)
#> 2 POINT (438239.2 5343090)
#> 3 POINT (438214.1 5342496)
#> 4 POINT (438521.8 5340596)
#> 5 POINT (438026.8 5342695)
#> 6 POINT (438094.6 5341920)
#> 7 POINT (438242.8 5341084)
#> 8 POINT (438452.3 5340391)
#> 9 POINT (438516.1 5339555)
#> 10 POINT (437617.6 5342993)
sample_systematic(
raster = sraster, # input sraster
cellsize = 500, # grid distance
access = access, # define access road network
buff_outer = 200, # outer buffer - no sample units further than this distance from road
square = FALSE, # hexagonal tessellation
location = "corners", # take corners instead of centers
plot = TRUE
)
#> Simple feature collection with 635 features and 0 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431149.7 ymin: 5337713 xmax: 438493.9 ymax: 5343209
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> geometry
#> 1 POINT (431886.5 5338179)
#> 2 POINT (431968.4 5337902)
#> 3 POINT (431149.7 5340670)
#> 4 POINT (431231.5 5340394)
#> 5 POINT (431968.4 5337902)
#> 6 POINT (431886.5 5338179)
#> 7 POINT (432003.4 5338665)
#> 8 POINT (431886.5 5338179)
#> 9 POINT (431231.5 5340394)
#> 10 POINT (431149.7 5340670)
sample_strat
The sample_strat()
contains two method
s to
perform sampling:
"Queinnec"
- Hierarchical sampling using a focal
window to isolate contiguous groups of stratum pixels, which was
originally developed by Martin Queinnec.
"random"
- Traditional stratified random sampling.
This method
ignores much of the functionality of the
algorithm to allow users the capability to use standard stratified
random sampling approaches without the use of a focal window to locate
contiguous stratum cells.
method = "Queinnec"
Queinnec, M., White, J. C., & Coops, N. C. (2021). Comparing airborne and spaceborne photon-counting LiDAR canopy structural estimates across different boreal forest types. Remote Sensing of Environment, 262(August 2020), 112510.
This algorithm uses moving window (wrow
and
wcol
parameters) to filter the input sraster
to prioritize sample unit allocation to where stratum pixels are
spatially grouped, rather than dispersed individuals across the
landscape.
Sampling is performed using 2 rules:
Rule 1 - Sample within spatially grouped stratum
pixels. Moving window defined by wrow
and
wcol
.
Rule 2 - If no additional sample units exist to
satisfy desired sample size(nSamp
), individual stratum
pixels are sampled.
The rule applied to a select each sample unit is defined in the
rule
attribute of output samples. We give a few examples
below:
#--- perform stratified sampling random sampling ---#
sample_strat(
sraster = sraster, # input sraster
nSamp = 200
) # desired sample size # plot
#> Simple feature collection with 200 features and 3 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431230 ymin: 5337750 xmax: 438530 ymax: 5343210
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata type rule geometry
#> x 1 new rule1 POINT (436250 5342970)
#> x1 1 new rule1 POINT (435830 5339590)
#> x2 1 new rule1 POINT (435650 5342590)
#> x3 1 new rule1 POINT (433830 5341550)
#> x4 1 new rule1 POINT (438230 5339230)
#> x5 1 new rule1 POINT (437750 5341150)
#> x6 1 new rule1 POINT (436290 5342770)
#> x7 1 new rule1 POINT (433550 5341550)
#> x8 1 new rule1 POINT (434290 5341090)
#> x9 1 new rule1 POINT (434230 5341350)
In some cases, users might want to include an existing
sample within the algorithm. In order to adjust the total number of
sample units needed per stratum to reflect those already present in
existing
, we can use the intermediate function
extract_strata()
.
This function uses the sraster
and existing
sample units and extracts the stratum for each. These sample units can
be included within sample_strat()
, which adjusts total
sample units required per class based on representation in
existing
.
#--- extract strata values to existing samples ---#
e.sr <- extract_strata(
sraster = sraster, # input sraster
existing = existing
) # existing samples to add strata value to
TIP!
sample_strat()
requires the sraster
input
to have an attribute named strata
and will give an error if
it doesn’t.
sample_strat(
sraster = sraster, # input sraster
nSamp = 200, # desired sample size
access = access, # define access road network
existing = e.sr, # existing sample with strata values
mindist = 200, # minimum distance sample units must be apart from one another
buff_inner = 50, # inner buffer - no sample units within this distance from road
buff_outer = 200, # outer buffer - no sample units further than this distance from road
plot = TRUE
) # plot
#> Simple feature collection with 400 features and 3 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431130 ymin: 5337730 xmax: 438530 ymax: 5343230
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata type rule geometry
#> 1 1 existing existing POINT (437570 5342770)
#> 2 1 existing existing POINT (433510 5340890)
#> 3 1 existing existing POINT (437630 5338990)
#> 4 1 existing existing POINT (433510 5341070)
#> 5 1 existing existing POINT (437790 5342710)
#> 6 1 existing existing POINT (435450 5342070)
#> 7 1 existing existing POINT (438270 5339550)
#> 8 1 existing existing POINT (435450 5342850)
#> 9 1 existing existing POINT (437110 5338070)
#> 10 1 existing existing POINT (433390 5341410)
The code in the example above defined the mindist
parameter, which specifies the minimum euclidean distance that new
sample units must be apart from one another.
Notice that the sample units have type
and
rule
attributes which outline whether they are
existing
or new
, and whether
rule1
or rule2
were used to select them. If
type
is existing (a user provided
existing
sample), rule
will be
existing as well as seen above.
sample_strat(
sraster = sraster, # input
nSamp = 200, # desired sample size
access = access, # define access road network
existing = e.sr, # existing samples with strata values
include = TRUE, # include existing sample in nSamp total
buff_outer = 200, # outer buffer - no samples further than this distance from road
plot = TRUE
) # plot
#> Simple feature collection with 200 features and 3 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431130 ymin: 5337730 xmax: 438530 ymax: 5343170
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata type rule geometry
#> 1 1 existing existing POINT (437570 5342770)
#> 2 1 existing existing POINT (433510 5340890)
#> 3 1 existing existing POINT (437630 5338990)
#> 4 1 existing existing POINT (433510 5341070)
#> 5 1 existing existing POINT (437790 5342710)
#> 6 1 existing existing POINT (435450 5342070)
#> 7 1 existing existing POINT (438270 5339550)
#> 8 1 existing existing POINT (435450 5342850)
#> 9 1 existing existing POINT (437110 5338070)
#> 10 1 existing existing POINT (433390 5341410)
The include
parameter determines whether
existing
sample units should be included in the total
sample size defined by nSamp
. By default, the
include
parameter is set as FALSE
.
method = "random
Stratified random sampling with equal probability for all cells
(using default algorithm values for mindist
and no use of
access
functionality). In essence this method perform the
sample_srs
algorithm for each stratum separately to meet
the specified sample size.
#--- perform stratified sampling random sampling ---#
sample_strat(
sraster = sraster, # input sraster
method = "random", # stratified random sampling
nSamp = 200, # desired sample size
plot = TRUE
) # plot
#> Simple feature collection with 200 features and 2 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431110 ymin: 5337730 xmax: 438530 ymax: 5343210
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata type geometry
#> x 1 new POINT (438510 5337970)
#> x1 1 new POINT (433190 5341630)
#> x2 1 new POINT (438070 5339550)
#> x3 1 new POINT (434710 5339050)
#> x4 1 new POINT (435470 5342050)
#> x5 1 new POINT (434850 5342830)
#> x6 1 new POINT (435650 5342270)
#> x7 1 new POINT (432670 5342810)
#> x8 1 new POINT (433750 5340830)
#> x9 1 new POINT (435950 5339430)
sample_sys_strat
sample_sys_strat()
function implements systematic
stratified sampling on an sraster
. This function uses the
same functionality as sample_systematic()
but takes an
sraster
as input and performs sampling on each stratum
iteratively.
#--- perform grid sampling on each stratum separately ---#
sample_sys_strat(
sraster = sraster, # input sraster with 4 strata
cellsize = 1000, # grid size
plot = TRUE # plot output
)
#> Warning: [readStart] source already open for reading
#> Processing strata : 1
#> Warning: [extract] source already open for reading
#> Processing strata : 2
#> Warning: [extract] source already open for reading
#> Processing strata : 3
#> Warning: [extract] source already open for reading
#> Processing strata : 4
#> Warning: [extract] source already open for reading
#> Simple feature collection with 30 features and 1 field
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431134.3 ymin: 5337735 xmax: 438391.3 ymax: 5343003
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata geometry
#> 1 1 POINT (437894.8 5340474)
#> 2 1 POINT (437542.2 5339104)
#> 3 1 POINT (435508.5 5342549)
#> 4 1 POINT (437189.5 5337735)
#> 5 1 POINT (433786.3 5341532)
#> 6 2 POINT (431762 5338570)
#> 7 2 POINT (432392.5 5339836)
#> 8 2 POINT (432387.4 5342998)
#> 9 2 POINT (433971.1 5341420)
#> 10 2 POINT (434288.8 5340471)
Just like with sample_systematic()
we can specify where
we want our samples to fall within our tessellations. We specify
location = "corners"
below. Note that the tesselations are
all saved to a list file when details = TRUE
should the
user want to save them.
sample_sys_strat(
sraster = sraster, # input sraster with 4 strata
cellsize = 500, # grid size
square = FALSE, # hexagon tessellation
location = "corners", # samples on tessellation corners
plot = TRUE # plot output
)
#> Processing strata : 1
#> Warning: [extract] source already open for reading
#> Processing strata : 2
#> Warning: [extract] source already open for reading
#> Processing strata : 3
#> Warning: [extract] source already open for reading
#> Processing strata : 4
#> Warning: [extract] source already open for reading
#> Simple feature collection with 1261 features and 1 field
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431105.8 ymin: 5337705 xmax: 438538.9 ymax: 5343224
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata geometry
#> 1 1 POINT (432962.5 5343057)
#> 2 1 POINT (432674.7 5343035)
#> 3 1 POINT (434401.7 5343168)
#> 4 1 POINT (432674.7 5343035)
#> 5 1 POINT (432549.9 5342775)
#> 6 1 POINT (432962.5 5343057)
#> 7 1 POINT (432962.5 5343057)
#> 8 1 POINT (434401.7 5343168)
#> 9 1 POINT (437730.8 5343172)
#> 10 1 POINT (437443 5343150)
This sampling approach could be especially useful incombination with
strat_poly()
to ensure consistency of sampling accross
specific management units.
#--- read polygon coverage ---#
poly <- system.file("extdata", "inventory_polygons.shp", package = "sgsR")
fri <- sf::st_read(poly)
#> Reading layer `inventory_polygons' from data source
#> `/tmp/RtmpdRhMcx/Rinst15736361282e/sgsR/extdata/inventory_polygons.shp'
#> using driver `ESRI Shapefile'
#> Simple feature collection with 632 features and 3 fields
#> Geometry type: MULTIPOLYGON
#> Dimension: XY
#> Bounding box: xmin: 431100 ymin: 5337700 xmax: 438560 ymax: 5343240
#> Projected CRS: UTM_Zone_17_Northern_Hemisphere
#--- stratify polygon coverage ---#
#--- specify polygon attribute to stratify ---#
attribute <- "NUTRIENTS"
#--- specify features within attribute & how they should be grouped ---#
#--- as a single vector ---#
features <- c("poor", "rich", "medium")
#--- get polygon stratification ---#
srasterpoly <- strat_poly(
poly = fri,
attribute = attribute,
features = features,
raster = sraster
)
#--- systematatic stratified sampling for each stratum ---#
sample_sys_strat(
sraster = srasterpoly, # input sraster from strat_poly() with 3 strata
cellsize = 500, # grid size
square = FALSE, # hexagon tessellation
location = "random", # randomize plot location
plot = TRUE # plot output
)
#> Processing strata : 1
#> Warning: [extract] source already open for reading
#> Processing strata : 2
#> Warning: [extract] source already open for reading
#> Processing strata : 3
#> Warning: [extract] source already open for reading
#> Simple feature collection with 163 features and 1 field
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431152.2 ymin: 5337716 xmax: 438498.4 ymax: 5343239
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata geometry
#> 1 1 POINT (431277.8 5343080)
#> 2 1 POINT (431366.7 5342305)
#> 3 1 POINT (432046.6 5342837)
#> 4 1 POINT (431991.3 5342485)
#> 5 1 POINT (432630.3 5342978)
#> 6 1 POINT (431795.6 5341897)
#> 7 1 POINT (432005.6 5342145)
#> 8 1 POINT (433789.7 5342817)
#> 9 1 POINT (432751.7 5341912)
#> 10 1 POINT (433339.8 5342370)
sample_nc
sample_nc()
function implements the Nearest Centroid
sampling algorithm described in Melville &
Stone (2016). The algorithm uses kmeans clustering where the number
of clusters (centroids) is equal to the desired sample size
(nSamp
).
Cluster centers are located, which then prompts the nearest neighbour
mraster
pixel for each cluster to be selected (assuming
default k
parameter). These nearest neighbours are the
output sample units.
#--- perform simple random sampling ---#
sample_nc(
mraster = mraster, # input
nSamp = 25, # desired sample size
plot = TRUE
)
#> K-means being performed on 3 layers with 25 centers.
#> Simple feature collection with 25 features and 4 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431410 ymin: 5337910 xmax: 438490 ymax: 5342970
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> zq90 pzabove2 zsd kcenter geometry
#> 4924 14.40 47.3 4.04 1 POINT (432590 5342970)
#> 38086 17.20 88.6 4.28 2 POINT (431890 5341190)
#> 99566 4.67 30.8 1.12 3 POINT (438050 5337910)
#> 15785 15.00 88.7 3.65 4 POINT (433470 5342390)
#> 28781 18.70 86.0 5.11 5 POINT (432290 5341690)
#> 83289 23.80 89.8 6.83 6 POINT (433290 5338770)
#> 89485 7.33 14.2 1.96 7 POINT (437850 5338450)
#> 4966 10.40 64.3 2.73 8 POINT (433430 5342970)
#> 39124 7.62 81.8 1.70 9 POINT (437730 5341150)
#> 41792 15.20 94.5 2.71 10 POINT (431410 5340990)
Altering the k
parameter leads to a multiplicative
increase in output sample units where total output samples = nSamp * k.
#--- perform simple random sampling ---#
samples <- sample_nc(
mraster = mraster, # input
k = 2, # number of nearest neighbours to take for each kmeans center
nSamp = 25, # desired sample size
plot = TRUE
)
#> K-means being performed on 3 layers with 25 centers.
Visualizing what the kmeans centers and sample units looks like is
possible when using details = TRUE
. The $kplot
output provides a quick visualization of where the centers are based on
a scatter plot of the first 2 layers in mraster
. Notice
that the centers are well distributed in covariate space and chosen
sample units are the closest pixels to each center (nearest
neighbours).
sample_clhs
sample_clhs()
function implements conditioned Latin
hypercube (clhs) sampling methodology from the clhs
package.
TIP!
A number of other functions in the sgsR
package help to
provide guidance on clhs sampling including calculate_pop()
and calculate_lhsOpt()
. Check out these functions to better
understand how sample numbers could be optimized.
The syntax for this function is similar to others shown above,
although parameters like iter
, which define the number of
iterations within the Metropolis-Hastings process are important to
consider. In these examples we use a low iter
value for
efficiency. Default values for iter
within the
clhs
package are 10,000.
sample_clhs(
mraster = mraster, # input
nSamp = 200, # desired sample size
plot = TRUE, # plot
iter = 100
) # number of iterations
The cost
parameter defines the mraster
covariate, which is used to constrain the clhs sampling. An example
could be the distance a pixel is from road access
(e.g. from calculate_distance()
see example below), terrain
slope, the output from calculate_coobs()
, or many
others.
#--- cost constrained examples ---#
#--- calculate distance to access layer for each pixel in mr ---#
mr.c <- calculate_distance(
raster = mraster, # input
access = access, # define access road network
plot = TRUE
) # plot
#> |---------|---------|---------|---------|=========================================
sample_balanced
The sample_balanced()
algorithm performs a balanced
sampling methodology from the stratifyR / SamplingBigData
packages.
#> Simple feature collection with 200 features and 0 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431130 ymin: 5337730 xmax: 438550 ymax: 5343230
#> Projected CRS: +proj=utm +zone=17 +ellps=GRS80 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs
#> First 10 features:
#> geometry
#> 1 POINT (434410 5343230)
#> 2 POINT (434170 5343210)
#> 3 POINT (431790 5343150)
#> 4 POINT (433830 5343110)
#> 5 POINT (435610 5343110)
#> 6 POINT (434630 5343090)
#> 7 POINT (437590 5343070)
#> 8 POINT (433170 5343030)
#> 9 POINT (432170 5343010)
#> 10 POINT (436910 5342990)
sample_balanced(
mraster = mraster, # input
nSamp = 100, # desired sample size
algorithm = "lcube", # algorithm type
access = access, # define access road network
buff_inner = 50, # inner buffer - no sample units within this distance from road
buff_outer = 200
) # outer buffer - no sample units further than this distance from road
#> Simple feature collection with 100 features and 0 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431390 ymin: 5337810 xmax: 438550 ymax: 5343210
#> Projected CRS: +proj=utm +zone=17 +ellps=GRS80 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs
#> First 10 features:
#> geometry
#> 1 POINT (434270 5343210)
#> 2 POINT (434050 5343190)
#> 3 POINT (435830 5343150)
#> 4 POINT (434010 5343010)
#> 5 POINT (431970 5342970)
#> 6 POINT (434590 5342970)
#> 7 POINT (436370 5342950)
#> 8 POINT (432270 5342930)
#> 9 POINT (434810 5342890)
#> 10 POINT (437850 5342870)
sample_ahels
The sample_ahels()
function performs the adapted
Hypercube Evaluation of a Legacy Sample (ahels) algorithm
usingexisting
sample data and an mraster
. New
sample units are allocated based on quantile ratios between the
existing
sample and mraster
covariate
dataset.
This algorithm was adapted from that presented in the paper below, which we highly recommend.
Malone BP, Minansy B, Brungard C. 2019. Some methods to improve the utility of conditioned Latin hypercube sampling. PeerJ 7:e6451 DOI 10.7717/peerj.6451
This algorithm:
Determines the quantile distributions of existing
sample units and mraster
covariates.
Determines quantiles where there is a disparity between sample units and covariates.
Prioritizes sampling within those quantile to improve representation.
To use this function, user must first specify the number of quantiles
(nQuant
) followed by either the nSamp
(total
number of desired sample units to be added) or the
threshold
(sampling ratio vs. covariate coverage ratio for
quantiles - default is 0.9) parameters.
#--- remove `type` variable from existing - causes plotting issues ---#
existing <- existing %>% select(-type)
sample_ahels(
mraster = mraster,
existing = existing, # existing sample
plot = TRUE
) # plot
#> Simple feature collection with 360 features and 7 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431110 ymin: 5337730 xmax: 438530 ymax: 5343170
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> type.x zq90 pzabove2 zsd strata type.y rule geometry
#> 1 existing 3.41 25.5 0.82 1 new rule1 POINT (437570 5342770)
#> 2 existing 3.05 17.9 0.56 1 new rule1 POINT (433510 5340890)
#> 3 existing 2.43 9.1 0.35 1 new rule1 POINT (437630 5338990)
#> 4 existing 7.06 9.1 1.85 1 new rule1 POINT (433510 5341070)
#> 5 existing 2.15 1.8 0.26 1 new rule1 POINT (437790 5342710)
#> 6 existing 7.21 40.4 1.87 1 new rule1 POINT (435450 5342070)
#> 7 existing 7.47 35.5 1.97 1 new rule1 POINT (438270 5339550)
#> 8 existing 5.15 17.6 1.24 1 new rule1 POINT (435450 5342850)
#> 9 existing 7.08 78.7 1.52 1 new rule1 POINT (437110 5338070)
#> 10 existing 9.58 35.9 2.62 1 new rule1 POINT (433390 5341410)
TIP!
Notice that no threshold
, nSamp
, or
nQuant
were defined. That is because the default setting
for threshold = 0.9
and nQuant = 10
.
The first matrix output shows the quantile ratios between the sample and the covariates. A value of 1.0 indicates that the sample is representative of quantile coverage. Values > 1.0 indicate over representation of sample units, while < 1.0 indicate under representation.
sample_ahels(
mraster = mraster,
existing = existing, # existing sample
nQuant = 20, # define 20 quantiles
nSamp = 300
) # desired sample size
#> Simple feature collection with 500 features and 7 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431110 ymin: 5337710 xmax: 438530 ymax: 5343190
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> type.x zq90 pzabove2 zsd strata type.y rule geometry
#> 1 existing 3.41 25.5 0.82 1 new rule1 POINT (437570 5342770)
#> 2 existing 3.05 17.9 0.56 1 new rule1 POINT (433510 5340890)
#> 3 existing 2.43 9.1 0.35 1 new rule1 POINT (437630 5338990)
#> 4 existing 7.06 9.1 1.85 1 new rule1 POINT (433510 5341070)
#> 5 existing 2.15 1.8 0.26 1 new rule1 POINT (437790 5342710)
#> 6 existing 7.21 40.4 1.87 1 new rule1 POINT (435450 5342070)
#> 7 existing 7.47 35.5 1.97 1 new rule1 POINT (438270 5339550)
#> 8 existing 5.15 17.6 1.24 1 new rule1 POINT (435450 5342850)
#> 9 existing 7.08 78.7 1.52 1 new rule1 POINT (437110 5338070)
#> 10 existing 9.58 35.9 2.62 1 new rule1 POINT (433390 5341410)
Notice that the total number of samples is 500. This value is the sum
of existing units (200) and number of sample units defined by
nSamp = 300
.
sample_existing
Acknowledging that existing
sample networks are common
is important. There is significant investment into these samples, and in
order to keep inventories up-to-date, we often need to collect new data
for sample units. The sample_existing
algorithm provides
the user with methods for sub-sampling an existing
sample
network should the financial / logistical resources not be available to
collect data at all sample units. The functions allows users to choose
between algorithm types using (type = "clhs"
- default,
type = "balanced"
, type = "srs"
,
type = "strat"
). Differences in type result in calling
internal sample_existing_*()
functions
(sample_existing_clhs()
(default),
sample_existing_balanced()
,
sample_existing_srs()
,
sample_existing_strat()
). These functions are not exported
to be used stand-alone, however they employ the same functionality as
their sample_clhs()
etc counterparts.
While using sample_existing()
, should the user wish to
specify algorithm specific parameters
(e.g. algorithm = "lcube"
in sample_balanced()
or allocation = "equal"
in sample_strat()
),
they can specify within sample_existing()
as if calling the
function directly.
I give applied examples for all methods below that are based on the following scenario:
We have a systematic sample where sample units are 200m apart.
We know we only have resources to sample 300 of them.
We have some ALS data available (mraster
), which we
can use to improve knowledge of the metric populations.
See our existing
sample for the scenario below.
#--- generate existing samples and extract metrics ---#
existing <- sample_systematic(raster = mraster, cellsize = 200, plot = TRUE)
sample_existing(type = "clhs")
The algorithm is unique in that it has two fundamental approaches:
existing
and the attributes it
contains.#--- sub sample using ---#
sample_existing(existing = e, nSamp = 300, type = "clhs")
#> Simple feature collection with 300 features and 3 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431127.7 ymin: 5337731 xmax: 438522.7 ymax: 5343201
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> zq90 pzabove2 zsd geometry
#> 624 18.10 89.1 6.05 POINT (435306.4 5341335)
#> 826 12.90 81.7 3.29 POINT (438163.9 5341522)
#> 68 11.40 66.5 3.04 POINT (434574.5 5337759)
#> 310 10.00 57.4 2.62 POINT (435593.8 5339296)
#> 517 11.00 4.5 3.42 POINT (435277.8 5340703)
#> 505 10.70 54.4 2.84 POINT (432664.3 5341708)
#> 218 15.60 93.4 4.04 POINT (437618.6 5337874)
#> 837 4.02 33.5 0.80 POINT (435995.5 5342570)
#> 238 6.47 17.0 1.68 POINT (435076.9 5339066)
#> 220 20.10 96.0 6.38 POINT (437992 5337731)
raster
distributionsOur systematic sample of ~900 plots is fairly comprehensive, however
we can generate a true population distribution through the inclusion of
the ALS metrics in the sampling process. The metrics will be included in
internal latin hypercube sampling to help guide sub-sampling of
existing
.
#--- sub sample using ---#
sample_existing(
existing = existing, # our existing sample
nSamp = 300, # desired sample size
raster = mraster, # include mraster metrics to guide sampling of existing
plot = TRUE
) # plot
#> Simple feature collection with 300 features and 3 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431113.9 ymin: 5337716 xmax: 438551.4 ymax: 5343230
#> Projected CRS: +proj=utm +zone=17 +ellps=GRS80 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs
#> First 10 features:
#> zq90 pzabove2 zsd geometry
#> 91798 18.50 84.9 5.72 POINT (438408.1 5339928)
#> 91759 13.00 78.6 3.83 POINT (437963 5339885)
#> 91622 14.90 81.9 3.53 POINT (432147.4 5341478)
#> 91753 17.10 86.2 4.26 POINT (436656.3 5340387)
#> 91578 3.07 9.0 0.52 POINT (437977.5 5338808)
#> 91986 15.90 90.5 3.30 POINT (436412 5341981)
#> 91293 12.80 49.8 3.64 POINT (434904.7 5338061)
#> 91526 16.30 90.4 3.46 POINT (433238.8 5340415)
#> 91428 16.50 90.8 4.23 POINT (434143.5 5339425)
#> 91977 11.90 64.1 3.31 POINT (434545.2 5342699)
The sample distribution again mimics the population distribution quite well! Now lets try using a cost variable to constrain the sub-sample.
#--- create distance from roads metric ---#
dist <- calculate_distance(raster = mraster, access = access)
#> |---------|---------|---------|---------|=========================================
#--- sub sample using ---#
sample_existing(
existing = existing, # our existing sample
nSamp = 300, # desired sample size
raster = dist, # include mraster metrics to guide sampling of existing
cost = 4, # either provide the index (band number) or the name of the cost layer
plot = TRUE
) # plot
#> Simple feature collection with 300 features and 4 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431127.7 ymin: 5337716 xmax: 438523 ymax: 5343230
#> Projected CRS: +proj=utm +zone=17 +ellps=GRS80 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs
#> First 10 features:
#> zq90 pzabove2 zsd dist2access geometry
#> 91493 12.70 84.9 2.73 462.34228 POINT (433167 5340229)
#> 91781 3.52 27.0 0.67 34.10512 POINT (434674.6 5341363)
#> 92020 26.20 90.3 8.16 44.12010 POINT (437977.2 5341593)
#> 91789 18.20 72.8 3.61 260.31160 POINT (436541.4 5340645)
#> 91699 16.60 92.7 3.87 80.57519 POINT (432477.6 5341779)
#> 91349 23.10 95.9 4.89 92.21993 POINT (436541.7 5337860)
#> 91507 12.30 85.6 3.32 60.25314 POINT (435967.2 5339152)
#> 91440 8.45 52.1 2.02 31.96237 POINT (437130.4 5338276)
#> 91762 15.90 92.5 3.90 378.10387 POINT (438523 5339669)
#> 91899 16.40 73.4 3.33 48.68666 POINT (436383.3 5341349)
Finally, should the user wish to further constrain the sample based
on access
like other sampling approaches in
sgsR
that is also possible.
#--- ensure access and existing are in the same CRS ---#
sf::st_crs(existing) <- sf::st_crs(access)
#--- sub sample using ---#
sample_existing(
existing = existing, # our existing sample
nSamp = 300, # desired sample size
raster = dist, # include mraster metrics to guide sampling of existing
cost = 4, # either provide the index (band number) or the name of the cost layer
access = access, # roads layer
buff_inner = 50, # inner buffer - no sample units within this distance from road
buff_outer = 300, # outer buffer - no sample units further than this distance from road
plot = TRUE
) # plot
#> Simple feature collection with 300 features and 4 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431127.7 ymin: 5337716 xmax: 438537.5 ymax: 5343230
#> Projected CRS: +proj=utm +zone=17 +ellps=GRS80 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs
#> First 10 features:
#> zq90 pzabove2 zsd dist2access geometry
#> 91319 5.98 50.0 1.44 73.00914 POINT (431415.1 5340688)
#> 91539 4.80 41.1 1.27 71.19559 POINT (434631.5 5341808)
#> 91653 13.90 91.3 2.82 165.10225 POINT (437632.5 5342369)
#> 91504 19.00 92.0 3.59 287.19337 POINT (435981.3 5340861)
#> 91467 21.30 97.2 6.49 72.76406 POINT (436211.1 5340344)
#> 91477 6.31 12.5 1.65 176.07778 POINT (433109.4 5341751)
#> 91592 18.10 83.5 4.60 76.81471 POINT (436713.6 5341651)
#> 91232 15.30 93.5 3.20 122.48509 POINT (435278.1 5337917)
#> 91414 3.46 16.6 1.53 59.06977 POINT (433267.4 5341047)
#> 91305 16.50 90.8 4.23 184.24413 POINT (434143.5 5339425)
TIP!
The greater constraints we add to sampling, the less likely we will have strong correlations between the population and sample, so its always important to understand these limitations and plan accordingly.
sample_existing(type = "balanced")
When type = "balanced"
users can define all parameters
that are found within sample_balanced()
. This means that
one can change the algorithm
, p
etc.
sample_existing(existing = e, nSamp = 300, type = "balanced")
#> Simple feature collection with 300 features and 3 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431128 ymin: 5337701 xmax: 438537.5 ymax: 5343201
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> zq90 pzabove2 zsd geometry
#> 1 20.3 91.5 5.56 POINT (431157 5337787)
#> 3 20.6 95.7 5.96 POINT (431228.8 5337974)
#> 6 15.0 95.8 2.87 POINT (431788.8 5337759)
#> 10 18.2 89.9 5.34 POINT (431673.9 5338017)
#> 19 21.1 88.5 6.60 POINT (431932.4 5338132)
#> 20 13.1 84.4 3.60 POINT (432119 5338060)
#> 25 29.0 92.3 8.92 POINT (433052.4 5337701)
#> 26 19.7 96.5 3.65 POINT (431257.4 5338606)
#> 27 17.5 94.0 3.27 POINT (431444.1 5338534)
#> 29 18.8 43.4 5.13 POINT (431817.5 5338391)
sample_existing(existing = e, nSamp = 300, type = "balanced", algorithm = "lcube")
#> Simple feature collection with 300 features and 3 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431171.2 ymin: 5337730 xmax: 438508.9 ymax: 5343216
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> zq90 pzabove2 zsd geometry
#> 3 20.6 95.7 5.96 POINT (431228.8 5337974)
#> 5 13.2 94.2 2.99 POINT (431602.1 5337831)
#> 9 15.8 90.9 4.00 POINT (431487.2 5338089)
#> 10 18.2 89.9 5.34 POINT (431673.9 5338017)
#> 14 14.1 77.1 3.99 POINT (432420.6 5337730)
#> 17 17.0 88.0 4.03 POINT (431559 5338276)
#> 18 19.7 90.2 6.39 POINT (431745.7 5338204)
#> 21 14.9 97.6 2.96 POINT (432305.7 5337989)
#> 24 27.1 95.1 7.11 POINT (432865.7 5337773)
#> 28 18.4 94.0 3.25 POINT (431630.8 5338462)
sample_existing(type = "srs")
The simplest, type = srs
, randomly selects sample
units.
sample_existing(existing = e, nSamp = 300, type = "srs")
#> Simple feature collection with 300 features and 3 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431128 ymin: 5337701 xmax: 438551.7 ymax: 5343230
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> zq90 pzabove2 zsd geometry
#> 1 13.50 58.5 3.83 POINT (435148.7 5339252)
#> 2 6.39 8.7 1.55 POINT (434445.1 5339094)
#> 3 8.56 56.9 2.22 POINT (435981.6 5338075)
#> 4 17.20 86.1 3.98 POINT (432319.9 5339697)
#> 5 17.00 95.2 4.01 POINT (431759.9 5339913)
#> 6 11.70 96.7 2.45 POINT (436240.1 5338190)
#> 7 5.14 11.8 1.17 POINT (438551.7 5340301)
#> 8 17.40 89.1 4.81 POINT (433454.4 5338190)
#> 9 16.10 86.4 4.31 POINT (432822.3 5341004)
#> 10 11.90 94.4 2.70 POINT (432003.8 5341104)
sample_existing(type = "strat")
When type = "strat"
, existing
must have an
attribute named strata
(just like how
sample_strat()
requires a strata
layer). If it
doesnt exist you will get an error. Lets define an sraster
so that we are compliant.
sraster <- strat_kmeans(mraster = mraster, nStrata = 4)
e_strata <- extract_strata(sraster = sraster, existing = e)
When we do have a strata attribute, the function works very much the
same as sample_strat()
in that is allows the user to define
the allocation
method ("prop"
- defaults,
"optim"
, "manual"
, "equal"
).
#--- proportional stratified sampling of existing ---#
sample_existing(existing = e_strata, nSamp = 300, type = "strat", allocation = "prop")
#> Simple feature collection with 299 features and 4 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431113.9 ymin: 5337716 xmax: 438522.7 ymax: 5343230
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata zq90 pzabove2 zsd geometry
#> 1 4 17.4 72.2 4.99 POINT (433353.7 5340157)
#> 2 4 21.7 80.1 4.38 POINT (436728.1 5340574)
#> 3 4 21.2 70.4 6.34 POINT (437474.8 5340287)
#> 4 4 19.7 88.5 5.17 POINT (438235.9 5338923)
#> 5 4 18.2 89.9 5.34 POINT (431673.9 5338017)
#> 6 4 20.4 93.3 4.69 POINT (433511.7 5339453)
#> 7 4 19.4 88.5 4.79 POINT (434976.5 5338247)
#> 8 4 22.2 87.2 5.86 POINT (436857.1 5342024)
#> 9 4 16.6 82.9 5.20 POINT (435679.8 5341191)
#> 10 4 21.1 94.3 4.48 POINT (434990.7 5339956)
TIP!
Remember that when allocation = "equal"
, the
nSamp
value will be allocated for each strata.
We get 400 sample units in our output below because we have 4 strata
and nSamp = 100
.
#--- equal stratified sampling of existing ---#
sample_existing(existing = e_strata, nSamp = 100, type = "strat", allocation = "equal")
#> Simple feature collection with 400 features and 4 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431128 ymin: 5337702 xmax: 438551.7 ymax: 5343215
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata zq90 pzabove2 zsd geometry
#> 1 4 21.1 88.5 6.60 POINT (431932.4 5338132)
#> 2 4 18.5 84.9 5.72 POINT (438408.1 5339928)
#> 3 4 19.9 96.3 4.82 POINT (431616.3 5339539)
#> 4 4 21.7 80.1 4.38 POINT (436728.1 5340574)
#> 5 4 19.4 96.1 4.84 POINT (431645 5340171)
#> 6 4 21.2 81.1 6.45 POINT (438120.7 5341967)
#> 7 4 20.2 95.0 5.73 POINT (431788.5 5340544)
#> 8 4 26.1 90.2 6.12 POINT (435651.1 5340559)
#> 9 4 21.6 78.1 6.77 POINT (431774 5341621)
#> 10 4 26.3 88.8 8.52 POINT (434071.8 5339238)
#--- manual stratified sampling of existing with user defined weights ---#
s <- sample_existing(existing = e_strata, nSamp = 100, type = "strat", allocation = "manual", weights = c(0.2, 0.6, 0.1, 0.1))
We can check the proportion of samples from each strata with:
Finally, type = "optim
allows for the user to define a
raster
metric to be used to optimize within strata
variances.
#--- manual stratified sampling of existing with user defined weights ---#
sample_existing(existing = e_strata, nSamp = 100, type = "strat", allocation = "optim", raster = mraster, metric = "zq90")
#> Simple feature collection with 100 features and 4 fields
#> Geometry type: POINT
#> Dimension: XY
#> Bounding box: xmin: 431128 ymin: 5337787 xmax: 438551.7 ymax: 5343143
#> Projected CRS: UTM Zone 17, Northern Hemisphere
#> First 10 features:
#> strata zq90 pzabove2 zsd geometry
#> 1 4 22.0 90.1 5.49 POINT (437043.8 5341952)
#> 2 4 18.4 86.4 6.01 POINT (438221.5 5339999)
#> 3 4 26.1 90.2 6.12 POINT (435651.1 5340559)
#> 4 4 21.0 57.7 6.87 POINT (436124.9 5341234)
#> 5 4 22.2 94.1 7.23 POINT (433483.1 5338822)
#> 6 4 19.4 96.1 4.84 POINT (431645 5340171)
#> 7 4 20.4 93.3 4.69 POINT (433511.7 5339453)
#> 8 4 22.4 91.2 4.78 POINT (432635.6 5341076)
#> 9 4 28.3 92.6 8.71 POINT (434789.5 5341105)
#> 10 4 20.3 91.5 5.56 POINT (431157 5337787)
We see from the output that we get 300 sample units that are a
sub-sample of existing
. The plotted output shows cumulative
frequency distributions of the population (all existing
samples) and the sub-sample (the 300 samples we requested).