IEEE Geoscience and Remote Sensing Magazine - March 2023 - 104

analyzes the activation maps and utilizes the concept of occlusion
sensitivity in the activation space.
The task of wilderness mapping is interesting in a twofold
way: First of all, wilderness areas are an essential element
of the natural environment and provide native habitat for
many species, which are often endangered. Thus, the very
task itself is important in the context of conservation and
environmental protection. Second, though, the term wilderness
is of a philosophical nature and is comparably ill defined
from a mathematical/technical standpoint [20], [21].
Explainable ML methods
GIVEN THE VAGUE
DEFINITION OF THE TERM,
WE ARGUE THAT
WILDERNESS IS NOT AN
INSTANCE TO MAP, BUT A
CONCEPT TO DISCOVER.
can be used to address the
task of wilderness mapping.
The objective of these methods
is to uncover patterns in
the decision-making process
of the mapping task with the
goal of improving the understanding
of what makes nature
wild. Explainable ML approaches,
which are used for
the discovery of new scientific
knowledge, such as the exploration of wildlife characteristics,
combine interpretation tools that present complex processes
as in neural networks in a human-understandable
space with domain knowledge to derive explanations [22].
Taking the ML perspective further, we claim that training
models for wilderness mapping are not only an ideal
testbed for methodical developments in explainable ML
but also in other subfields, such as weakly supervised
learning. Thus, the dataset will not only cater to the environmental
RS community but also address technical audiences
beyond that.
WILDERNESS MAPPING
Anthropogenic pressure (i.e., human influence) on the environment
is the largest single cause of loss of biological
diversity [23], and the conservation of ecological dynamics
and biodiversity is now more important than ever before.
With the ever-increasing prominence of the need for environmental
conservation, the world is in desperate need of
environmental monitoring and quantifying of the degree
of human influence on the environment.
While there is a lot of work being done in fields such as
unsupervised, semisupervised, and weakly supervised ML,
most of the established methods still rely on the supervised
learning paradigm, rendering the need for carefully annotated
training data. In the context of environmental RS, this
can be achieved for rather well-defined target classes, such
as " water " or " forests, " but it becomes a significant challenge
if a less technical interpretation of the environment
is the desired goal. As an example, many researchers in the
past have presented rule-based approaches to study concepts
such as naturalness [24], [25], human influence [26],
or wilderness [27]. All of these approaches are carefully designed
by domain experts and cannot be validated against
104
actual, biophysically measurable ground truth. The term
" wilderness " is philosophical rather than technical in nature
and subject to many different definitions [20], with the
definition of the European Commission being [29], p. 10]:
A wilderness is an area governed by natural processes. It is
composed of native habitats and species, and large enough for
the effective ecological functioning of natural processes. It is
unmodified or only slightly modified and without intrusive or
extractive human activity, settlements, infrastructure or visual
disturbance.
This definition assigns four central ecological aspects
to the concept of wilderness: 1) naturalness, 2) undisturbedness,
3) undevelopedness, and 4) scale. While scale can
easily be defined in a mathematical sense, e.g., by defining
so-called minimum mapping units, naturalness refers
to ecosystems functioning in a natural way and can thus
hardly be measured in a technical sense, but is subject to
judgment by ecology experts based on in situ observations.
Undisturbedness and undevelopedness, however,
are concepts that can potentially be observed with the use
of RS technologies. As per the European Commission, undevelopedness
refers to the absence of " habitation, settlements
or other human artefacts such as power lines, roads,
railways, fences [that] may hinder ecological processes directly
or by promoting the likelihood of human interference "
[29], p. 12].
While the task of wilderness mapping could be considered
as a special case of land cover mapping, there is a
subtle difference to be noted that appear from the ill-posed
definition of the term wilderness. Since there are nomenclatures
with precise class definitions to follow for conventional
land cover mapping, it is relatively easier to decide
what class to assign to the instance of interest, especially
in the presence of domain knowledge. However, given the
vague definition of the term, we argue that wilderness is
not an instance to map, but a concept to discover.
It is undeniable that wilderness areas pose an utmost significance
given their role in harboring species and habitats,
which are indisputable elements of the environment. This
feature of the term wilderness gives a rise to incorporating
explainable ML methods with the purpose of interpreting
the decisions made by the DL.
Wilderness has no biophysical basis and thus cannot
be easily delineated or categorized. Instead, wilderness is
a cultural concept [30], and the associated definition of
wilderness changes over time. Moreover, changing policy
and management strategies make the identification and
protection of wilderness compelling. There exists a diverse
set of explanations of the word " wilderness, " most
of which are engaged with describing it in a philosophical
way, which might seem vague from a technical point of
view. Besides, understanding wilderness-a key element
of nature-poses great importance since the protection
of nature can be fostered by better defining its ill-defined
elements. Consequently, we adopt a bottom-up approach
and address the ambiguity in nature's vital elements (e.g.,
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
