IEEE Geoscience and Remote Sensing Magazine - March 2023 - 106

WE USE THE KÖPPEN-
GEIGER CLIMATE
CLASSIFICATION MAP AND
THE ESA WORLDCOVER MAP
AS A PROXY TO MAINTAIN
GUIDANCE OVER THE
SAMPLING PROCESS.
wide swath mode. The Level-2A Sentinel-2 image contains
10 spectral bands (B2, B3, B4, B5, B6, B7, B8, B8A, B11,
and B12) and is subject to mosaicking to mitigate the effect
caused by clouds. We stack the
imagery of any area of interest
(AOI) for a defined period of
time in the year 2020, and for
each band in each pixel, we select
the value defining the 25%
percentile regarding reflectivities
observed in that time period.
We empirically found
that this way of compositing
results in a more informative
image than median compositing.
To further widen the scope
of the dataset, we provide four seasons of Sentinel-2 images
for every AOI. By adding a multitemporal value to the dataset
by including all seasons for every AOI, we aim to further
facilitate the investigation of wilderness mapping in a
time-series manner. Further, rather than defining a single
season period when querying the seasons, we take hemispheres
of the AOI into account and define the season periods
accordingly (i.e., Northern Hemisphere and Southern
Hemisphere). We use the remaining sources (ESA
WorldCover and the VIIRS) in their original forms and leave
the end users with more flexibility. We performed reprojection
and scaling to have the corresponding pixels across
geodata sources in the same scale and projection, which
uses the rubber-sheet algorithm as a registration method.
The curation strategy is illustrated in Figure 2.
SAMPLING STRATEGY AND EXPORTING
Considering the difficulty of adequately sampling Earth's
surface in terms of representability, we develop a guided
TABLE 1. IUCN PROTECTED AREA CATEGORIES.
WDPA CLASS
DESCRIPTION
Ia: Strict nature reserve
Ib: Wilderness area
II: National park
III: Natural monument
or feature
IV: Habitat/species
management area
V: Protected landscape/
seascape
VI: Protected area with
sustainable use of natural
resources
Strictly protected areas where human
presence is strictly limited and controlled
Slightly modified areas with little human
presence in the form of indigenous and
local communities
Functioning ecosystems subject to tourism
through zoning
Areas where preserving a particular
feature hosting a cultural value is the
dominant goal
Areas where flora species, fauna species,
or habitats are aimed to be preserved
and/or restored through informed
interventions
Areas where a distinct value is created by
human presence over time
Areas where the conservation of natural
ecosystems and ecological processes
take place
sampling approach taking the hard constraints of the WDPA
polygons into account. Prior to filtering the images, sampling
of the WDPA polygons takes place. Merely using the original
WDPA polygons is not ideal because of the uneven distribution
across continents and disparity in polygons (in terms of
both volume and size). To this end, to ensure the versatility
of the dataset and the spatial coverage of its samples, we
design a climate map and land cover type-aware semiautomated
sampling approach, where a subsampling of the
polygons is performed while improving the spatial coverage
of the WDPA polygons. The developed approach is
illustrated in Figure 3. We start the process by applying a
class-wise minimum area threshold (using the asset feature
GIS_AREA) of 5 km2 to category Ia and category Ib
and 100 km2 to category II. Then, after filtering out the
nonterritorial areas, the sampling operation takes place,
where we use a GEE function (namely, Stratified Sample())
that allows users to supervise the sampling process with
weights that can be obtained from auxiliary data. For
our case, we use the Köppen-Geiger climate classification
map and the ESA WorldCover map as a proxy to
maintain guidance over the sampling process. The sampling
weights used in the StratifiedSample() function are
formed by calculating the class-wise correlation between
the WorldCover and climate classification maps (i.e., the
frequency of each land cover class in each climate classification
map). Consequently, the calculated weights
are inversely normalized, and row-wise summations are
used as sampling weights. Hence, oversampling of underrepresented
polygons (and vice versa) is performed
by following the observation that the representability of
protected areas could be promoted by taking land cover
and climate zones into account. Further, the polygons
within a certain proximity (30 km2 for category Ia and
category Ib and 50 km2 for category II) are removed. Ultimately,
the resulting samples are exported for further
analysis after fitting a bounding box with a size of 20 km
× 20 km to the center of each polygon.
After exporting, the images are renamed after the WDPA
polygon they contain and cropped into 1,920 × 1,920 pixel
patches to remove the reprojection effect occurring near
borders. We further add 108 AOIs from populous areas that
exhibit human disturbance on Earth in various forms to
strengthen the versatility of the dataset. The final distribution
of the AOIs sampled from the WDPA polygons is given
in Figure 4. The motivation behind this manual intervention,
which increased the number of AOIs in the dataset
from 910 to the final number of 1,018, is due to the potentiality
that the model might overfit the wilderness areas
in the form of forested areas (see the ESA WorldCover
class distribution in Figure 5). A naming convention of
9000000XXX (where X is the identification number of the
patch) is adopted when including the manually selected
AOIs in the dataset. Overall, our dataset has 8,144 images
(1,018 × 8, where 1,018 is the number of AOIs in the dataset
and 8 is the number of the data sources each AOI contains)
106
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2023
IEEE Geoscience and Remote Sensing Magazine - March 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2023
