IEEE Geoscience and Remote Sensing Magazine - March 2023 - 112

Applying ASOS to the MapInWild dataset, we demonstrate
an example task for explainable ML. Training
a neural network on classification, we can predict the
high-resolution sensitivity maps shown in Figure 9. The
first sample shows parts of the Australian national park
Karijini. It is centrally split by a mine and a railway. The
human influence on the surroundings appears to be so
small that the model does not highlight them as such.
The second sample is located at the Sid's Mountain Wilderness
Study Area in Utah, United States. The model detects
some parts that are mainly outside the study area as
nonwild. The inner parts of the study area seem to have
wild characteristics. The third sample shows the Skadar
Lake National Park in Montenegro and Albania. At the
top, agricultural fields reach inside the national park;
this is detected by the model. The lake itself has mainly
not been predicted because of high uncertainty. This
goes with the MapInWild dataset, which has been created
by specifically looking for nonwater polygons. The
fourth sample shows the small Skwaha Lake Ecological
Reserve in Canada. This reserve has, according to the
model, mixed characteristics. The two regions detected
as nonwild are valleys. The left one includes several
streets, and the right region includes Trans-Canada
Highway 1 along a river.
We provide the preceding experiments as an initial effort
in addressing the novel task of wilderness mapping
on the large-scale benchmark dataset MapInWild. The
task has first been investigated on
a pixel scale within the frame of
semantic segmentation, where the
imperfect wilderness annotations
have been used as a proxy when
mapping the wilderness areas. On
the one hand, there are cases where
the learner extrapolates the wilderness
area to the entire scene as the
wilderness area in the annotation
is of similar appearance to the remaining
area, which is annotated
as nonwilderness (the first sample
in Figure 6). This behavior of the
learner exhibits its ability to learn
the concept of wilderness from imperfect
annotations. On the other
hand, it is seen that, when mapping
the wilderness areas, the learner
dampens the presence of wilderness
characteristics shown in the proxy.
Although there is an annotated wilderness
characteristic in the center
of the area (the third sample in Figure
6),
the existing human influWilderness
Proxy
Background
FIGURE 9. Baseline results for the test images with IDs 64113, 374681, 16385, and 18415. Rowwise,
from left to right: input image, wilderness proxy, and sensitivity map. The color scale of
the sensitivity maps is given in Figure 8. Areas with a low density in the activation space are
not mapped to sensitivities and are colored gray in the sensitivity maps.
112
ence in the surroundings might be
negatively affecting the essence and
form of the wilderness concept realized
and discovered by the learner.
The concept realized might contain
a learned pattern on 1) the size of
a wilderness area, 2) the distance
to the nearest area under human
influence, and 3) characteristics of
the site in terms of land use and
land cover. Similar to this behavior,
the learner shows strength in leaving
out the resemblance of inner
city parks and forests with a wilderness
area in the form of a forested
area (the last sample of Figure 6). In
the middle of the learner's behavior
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