ArticlePDF Available


Ethics is a central and growing concern in all applications utilizing Artificial Intelligence (AI). Earth Observation (EO) or Remote Sensing (RS) research relies heavily on both Big Data and AI or Machine Learning (ML). While this reliance is not new, with increasing image resolutions and the growing number of EO/RS use-cases that have a direct impact on governance, policy, and the lives of people, ethical issues are taking center stage. In this article, we provide scientists engaged with AI4EO research (i) a practically useful overview of the key ethical issues emerging in this field with concrete examples from within EO/RS to explain these issues, and (ii) a first road-map (flowchart) that scientists can use to identify ethical issues in their ongoing research. With this, we aim to sensitize scientists about these issues and create a bridge to facilitate constructive and regular communication between scientists engaged in AI4EO research on the one hand, and ethics research on the other. The article also provides detailed illustrations from four AI4EO research fields to explain how scientists can redesign research questions to more effectively grab ethical opportunities to address real-world problems that are otherwise akin to ethical dilemmas with no win-win solution in sight. The paper concludes by providing recommendations to institutions that want to support ethically mindful AI4EO research and provides suggestions for future research in this field.
IEEE Proof
Ethics is a central and growing concern in all applications
utilizing artificial intelligence (AI). Earth observation
(EO) and remote sensing (RS) research relies heavily on
both big data and AI or machine learning (ML). While this
reliance is not new, with increasing image resolutions and
the growing number of EO/RS use cases that have a direct
impact on governance, policy, and the lives of people, ethi-
cal issues are taking center stage. In this article, we provide
scientists engaged with AI for EO (AI4EO) research, 1) a
practically useful overview of the key ethical issues emerg-
ing in this field, with concrete examples from within EO/RS
to explain these issues, and 2) a first road map (flowchart)
that scientists can use to identify ethical issues in their on-
going research. With this, we aim to sensitize scientists to
these issues and create a bridge to facilitate const ructive
and regular communication among scientists engaged in
AI4EO research, on the one hand, and ethics research, on
the other hand. The ar ticle also provides detailed illustra-
tions from four AI4EO research fields to explain how sci-
entists can redesign research questions to more effectively
grab ethical oppor tunities to address real-world problems
that are other wise akin to ethical dilemmas with no win-
win solution in sight. The ar ticle concludes by providing
recommendations to institutions that want to support ethi-
cally mindf ul A I4EO research and prov ides suggestions for
future research in this field.
Big data lies at the heart of many EO and RS researc h and
development acti vities. In fac t, applications a nd algorith ms
emerging from RS and EO research rely heavily on big data
collected via satellites, unmanned aerial vehicles, drones,
and other state-of-the-art devices. At the same time, RS and
EO research is rapidly transforming in the era of AI and
ML. AI and ML per mit the petabytes of data collected by
satellites and other EO/R S devices to be systematically or-
ganized and used to train models t hat can predict a large
diversity of events, objects, and circumstances, even in the
absence or lack of so-called ground truth.
The close relationship bet ween (big) data and A I is reflect-
ed in the German Data Ethics Commission’s understanding
of “artificial intelligence” as “a collective term for technolo-
gies and their applications which process potentially very
large and heterogeneous data sets using complex methods
modeled on human intelligence to arrive at a result which
Digital Object Identifier 10.1109/MGRS.2022.3208357
Date of current version: xxxxxx
Earth Observation and
Artificial Intelligence
Understanding emerging ethical issues and opportunities
IEEE Proof
may be used in automated applications. The most important
building blocks of A I as part of compute r science are subsym-
bolic pattern recognition, ML , computerized knowledge rep-
resentation and knowledge processing, which encompasses
heuristic search, inference and planning” [1].
In this article, we refer to the expanding field of re-
search that uses A I/ML tools for organizing, understand-
ing, and analyzing EO/RS data as “AI4EO research.” At
present, AI4EO research can be roughly divided into two
categories. On t he one hand, 1) algorit hmic capabilities are
explored to determine whether (geo)information about
the land surface and the atmosphere can be derived au-
tomatically from the available multimodal EO data with
sufficient accuracy. On the other hand, 2) fields of applica-
tion are e xplored to identify issues linked to demograph-
ics, economics, ecological, social, and other fields from
available datasets. Further, where necessary and possible,
these latter e xplorations are increasingly done in combina-
tion with other newer sources of data, such as data derived
from questionnaires [2], censuses [3], and in social media
platforms [4], among many others.
Ethics is a central and growing concern in the fields of
(big) data, AI and ML generally, and each of the preced-
ing two broad categories of AI4EO research. Ethical issues,
well beyond the now well-known issues of privacy, explain-
ability, and bias, are increasingly relevant for EO and RS
scientists for several reasons: 1) the scale and resolution of
EO/RS images are improving rapidly; 2) new use cases are
emerg ing, which go beyond t he traditiona l role of providing
information on topics such as climate and env ironmental
changes and have a direct impact on specific human popu-
lations and groups as well as governmental policies linked
to such groups [5]; and 3) AI/ML tools are increasingly used
to manage and analyze the diverse sources of EO/RS data,
and several ethical issues are commonly known to arise in
several applications that utilize such tools. Beyond these
EO/RS-specif ic reasons, ethical issues can and do arise in
the context of A I research as a result of 1) the very fact of
data collection, 2) the approaches used for data analysis,
and 3) the manner and purpose for which the results from
such analysis are used.
While a great deal of work has been done in the Europe-
an Union (EU) and beyond to compile various guidelines
and guiding motifs for data ethics and AI ethics, recent re-
search and empirical studies sug gest that these guidelines
may not be practically useful for scientists and academic
researchers [6]. Various ethics guidelines also note that
ethical issues can and should be context and technology
specific [7]. In this article, we provide scientists engaged
with A I4EO research a practically useful over view of the
key ethical issues and oppor tunities emerging in this field.
More specifically, our aim is not only to sensitize scientists
to these issues but also to help create a bridge to facilitate
constr uctive and regular communication among scientists
engaged in AI4EO research, on the one hand, and ethics
research, on the other hand.
Accordingly, we divide t his article into the following
parts. In the “Understanding Ethics: Issues, Risks, Theories,
and Opportunities” section, we provide an explanation of
basic terms, namely, ethics, ethical risks, ethical issues, eth-
ical oppor tunities, and ethical theories, in simple and prac-
tical language. In the “Categorizing and Understanding
Artif icial Intelligence and Data Ethics for Artificial Intelli-
gence for Earth Observation Researc h” section, we compile
common ethics issues, g uidelines, and guiding motifs from
five sources: 1) the EU “Ethics Guidelines for Trustwor thy
AI” [7], 2) “AI4People—A n Ethical Framework for a Good
AI Society” [8], 3) the German “Recommendations of the
Data Ethics Commission for the Federal Government’s
Strategy on A I” [9], 4) the ethical principles compiled by
Jobin et al. f rom 84 global sources [10], and 5) ethical issues
and guidelines compiled by Hagendorff [6].
We organize these issues and g uidelines under six catego-
ries, which correspond with the three fundamental values of
(business) ethics (namely, honesty, integrity, and fairness)
and three centrally relevant emerging issues (namely, privac y,
responsibility, and sustainabilit y). Note: A distinction can be
made between et hics and business et hics, with ethics being
a broader and more general term linked to all/any moral val-
ues, including personal values and decisions we make based
on our personal conscience. Business ethics, on the other
hand, relates to the values and considerations that businesses
and institutions, including teams and individuals working in
an institutional setup, must bear in mind in the per formance
of their individual and collective tasks and activities.
The ethical issues under each of these six categories are
then e xplained with the help of illustrations from cur rent
AI4EO re search. I n so doing, we aim to ident ify et hical issue s
that are specifically relevant for current and ongoing AI4EO
research, given its unique nature and scale. We also hope
that with this first-of-its-kind endeavor, AI4EO researchers
will be better equipped to identify the ethical issues they
need to be most mindf ul of at every stage of their research.
In the “Ethical Opportunities in Artificial Intelligence for
Earth Obser vation Research” section, using four prominent
fields of AI4EO research as examples, namely, 1) artisanal
and small-scale mining (ASM), 2) slum mapping , 3) the
conser vation of biodiversity, and 4) building-level popula-
tion density estimations, we explain how AI4EO scientists
can go beyond merely “avoiding” ethical issues toward ac-
tively identifying and embracing ethical opportunities to
help resolve some of the most important challenges faced
by humankind today. We conclude with some observations
and recommendations for various AI4EO stakeholders and
for possible future directions in AI4EO research.
The term ethics has been defined from numerous per-
spectives (e.g., philosophical, anthropological, historical,
IEEE Proof
economic, political, and social) in existing academic and
scientific literature [11]. At the most f undamental level, eth-
ics deals with issues of right and wrong, good and bad. As
a discipline, ethics is partially preoccupied with identif ying
and finding ways of resolving values that come into conflict
in various contexts of human life, including scientific and
academic research [12].
Ethics is also about emphasizing constructive and “ap-
propriate conduct associated with” basic human “values,”
such as honesty, responsibility, and so on, and not merely
about “discover(ing) inappropriate conduct” [13], [14], [15].
Accordingly, some understand ethics to be about identify-
ing a broad set of desirable duties and (un)desirable con-
sequences [16]. Together, these are e xpected to provide a
framework to understand and avoid negative and harmf ul
human actions, interactions, and consequences, on the one
hand, and, perhaps more impor tantly, explore, guide, and/
or facilitate positive and constr uctive human actions and
interactions, on the other hand. Within this broad under-
standing of ethics that is considered by some scholars to be
context or even culture specific (so-called “ethical relativ-
ism” [17]), several scholars are of the v iew that there exists
an objective univer sally applicable set of human va lues [18],
[19] and responsibilities [20] that are fundamental to what
makes us “human.” These values are cherished indepen-
dently of culture and conte xt, and accordingly, all human
actions (and inaction) must be based, or build on, these
fundamental values. For example, the most f undamental
principle of ethics is the so-called golden rule (“don’t do to
others what you wouldn’t want others to do to you” [21])
and three associated values, namely, honesty, integrity, and
fairness [13], [22]. Together with other basic human values,
such as responsibility and respect for human rights, these
principles lie at the heart of all (business) ethics discourses
and theories. [13].
Ethical risks are, simply put, the most likely and predictable
negative consequences of any action or inaction. Accord-
ingly, actions and inactions that are highly likely to yield
negative and undesirable consequences are labeled as un-
ethical. (Note: Et hical risks must be distinguished from the
emerging field of “risk ethics” that debates whether and in
what circumstances it may be ethically justifiable to take a
risk, e.g., a calculated risk in business. For an over view of
risk ethics, see [23] and [24].)
Unlike the term ethics, the term ethical issue has not been
widely defined in academic literature. Nonetheless, this
term, which is commonly used in literature dealing with
ethics, is arguably self-explanatory. For a fundamental un-
derstanding, it is useful to look at the basic def initions of
the ter ms ethical and issue. According to the Collins English
Dictionar y, “ethical means relating to questions of right
and wrong ,” and “an issue is an important subject that peo-
ple are arguing about or discussing.” Put together, ethical
issues can be understood as issues relating to questions of
right and wrong (or good and bad) that are being debated
and discussed by people, including scientif ic and academic
The previously mentioned idea of “ethical relativism”
comes starkly into play when we attempt to identif y and
define ethical issues. By definition, issues, whether t hey are
linked to ethics, law, economics, natural sciences, society,
or politics, are subject to debate and diverse views. This is
not least because priorities (and, therefore, considerations)
that go into determining what is right, wrong, good, and
bad, can vary from person to person, from context to con-
text, from culture to culture, and depending on the per-
spective and desired end result from whic h one views any
Indeed, even within established fields of academic re-
search, ethical issues are not uncommon. For example, in
the sphere of medical research, while the collection and
analysis of data can help gain a better understanding of
a disease and make more accurate diagnoses, the process
can also raise ethical issues linked to patient privac y: t he
ethical issue here is, should we give primary importance to
data protection to secure patient privac y (the “ethical con-
cern/risk”) or to understanding and curing diseases more
eff iciently (the “ethical opportunity”)? Here, one can see
that “ethical opportunities” can be the f lip side of “ethi-
cal concerns/risks,” which together can lead to an “ethical
Ethical issues can also arise at a more fundamental level.
For example, while stem cell research has several known
and potential (medical) benefits, it is considered by many
to be fundamentally unethical [25], [26], [27], as it violates
cultural and religious sentiments and, according to a few
thinkers, the sanctity of life. It is therefore fair to say that
ethical issues arise from a ver y fundamental and even le-
gitimate space of conflicting values. These conf licting val-
ues can take t he shape of conflicting understanding of de-
sirable and undesirable goals, approaches, and end results.
They can also arise in relation to the benef it that may result
(for an individual or for society as a whole) from the pur-
suit of these goals and approaches, weighed against their
impact and consequences on the lives of those who were
not involved in the decision to pursue these goals. It is also
possible that the same individual or communit y can, at the
same time, stand to gain from an ethical opportunity while
losing from the corresponding ethical concern/risk.
When we are faced with an ethical issue that cannot be
resolved (e.g., by finding a win-win or middle path) by the
application of basic principles of ethics and with the help
of recognized theories of ethics, we may be facing what
is termed an ethical dilemma. In an ethical dilemma, the
pursuit and accomplishment of one very important goal/
IEEE Proof
outcome inevitably leads to a substantial or irreversible
sacrifice of another, equally important, goal/outcome.
The trolley dilemma is a typical et hical dilemma that is
hotly debated in the conte xt of decision making by auton-
omous cars [28]. In the context of illegal ASM that is often
surveyed using AI and EO data, there is another ethical
dilemma that governments on the ground find hard to
tackle: permit such mining to secure the right to a live-
lihood of poor and indigenous communities or strictly
ban such mining to protect the environment and human
health [29], [30].
The term ethical opportunities also has not been defined in
existing literature. Never theless, the term has been used in
various contexts. For example, in the context of increasing
consumer awareness and demand for labels that are “eco-
friendly” and “fair trade,” “ethical business” and ethical
business models have been described as an “opportunity”
to take a g reater share of the market or to carve out a niche
market for a business’ product or ser vice [31].
Within medical and bioethics, it has been said that
“ethical insight into projects may afford oppor tunities
for initiatives t hat add to the projects’ value and the sat-
isfaction of those involved …” and that “ethics is better
applied collaboratively, equipping practitioners to maxi-
mize the ethical advantages and minimize the ethical
costs of their proposals” [32]. Therefore, ethical opportu-
nities can also be seen as means of “max imizing t he ethi-
cal advantages,” while ethical concerns/risks can be seen
as “ethical costs” that need to be minimized or avoided
altoget her.
Further, the Scientific Committee of the AI4People ini-
tiative (an Atomium–European Institute for Science, Me-
dia, and Democracy initiative) has categorized the chief
ethical opportunities offered by AI, in general, under four
broad headings corresponding to what it considers to be
“four f undamental points in the understanding of human
dignity and flourishing” [8]. As per “AI4People—An Ethical
Framework for a Good AI Society,” in each of the following
categories, when AI is “used to foster human nature and its
potentialities,” it creates ethical opportunities:
1) “who we want to become” (opportunity: “enabling hu-
man self-realization”)
2) “what we can do” (opportunity: “enhancing human
3) “what we can achieve” (opportunity: “increasing indi-
vidual and societal capabilities”)
4) “how we can interact wit h each other and the world”
(opportunity: “cultivate societal cohesion”).
The AI4People’s framework also warns that in each of the
preceding cases, if A I is overused or misused,” ethical risks
can emerge, as follows [8, pp. 690– 691]:
1) “who we want to become” (risk: “devalu ing human ski lls”)
2) “what we can do” (risk: “removing human responsibility”)
3) “what we can achieve” (risk: “reducing human control”)
4) “how we can interact wit h each other and the world”
(risk: “eroding human self-determination”).
Much of the work being done by AI4EO scientists ar-
guably falls within the categories of “enhancing human
agency” and” increasing societal capabilities.” For example,
while the satellites serv ing the Coper nicus program are now
generating petabytes of information in the form of images
of various resolutions, the processing and analysis of these
data, if under taken by human agencies alone, would most
likely take too long and might not generate the required
information in a timely and usable manner. Here, AI/ML
“enhances human agency” and “increases individual and
societal capabilities” by providing relevant information,
including predictions and patterns, that is presumably as
or more accurate and reliable than what would have been
possible by human agency alone.
In his famous piece “Tragedy of the Commons,” Gar-
rett Hardin opined that certain problems do not have a
“technical solution” but need a “fundamental e xtension
in morality” [33]. This is arg uably true even today. None-
theless, technolog y can aid in overcoming several practical
problems as well as the accomplishment of specific societal
goals that humanity has not been able to resolve or accom-
plish so far through legal, moral, and (other) nontechnical
routes. The term ethical opportunities can therefore also be
understood as including all such instances where attempts
are made either by conscious use of specif ic technologies
or by conscious exclusion of specif ic tec hnologies to con-
scientiously address e xisting real-world challenges. In other
words, embracing ethical opportunities means acting over
and above now commonplace attempts to sidestep and
avoid the most well-known ethical concer ns/risks. In this
article, we discuss such ethical opportunities, with the help
of four examples from AI4EO research.
Ethical theories and approaches are essentially methods
and tools that can help explain, diagnose, and find solu-
tions to et hical issues and dilemmas. In other words, these
theories seek to give guidance on appropriate and inappro-
priate human behavior [34]. Ethical theories are different
from ethical approaches: ethical theories are often rooted
in the historical, religious, and philosophical thought of a
specific countr y or region of the world. They are often wide-
ly known and accepted or have been subjected to signif i-
cant academic or popular debate and discussion. Accord-
ingly, ethical theories are more well-developed “systems”
for categorizing human behavior and can be used to under-
stand, analyze, and even justify actions/inactions linked to
ethical issues and ethical dilemmas.
Ethical approac hes, on t he other hand, are usually more
recent constructs, though they may be rooted in or derived
from more established ethical theories. Such approaches
may, for example, provide a more systematic or stepwise
guide for identif ying ethical issues and for determining the
correct course of action when faced with an et hical issue or
IEEE Proof
dilemma. In this article, we attempt to explain an approach
that AI4EO scientists can take to identify ethical issues in
their research. Our approach is rooted in ethical theories
that are commonly used in the Wester n world. Undoubt-
edly, utilizing ethical theories from other cultures and phi-
losophies can enrich this approach. However, this endeavor
will be undertaken in subsequent works.
Ethical theories commonly used in t he Western world are
broadly classified under the following headings:
Ethics of conduct:
Consequentialist theories: These theories of ethics focus
on consequences and argue that the “right action” is
the one that produces the maximum good, either for
oneself (ethical egotism) or for the majority of the
concerned public/population (utilitarianism). Conse-
quentialist approaches to ethics, therefore, ask ques-
tions such as, “Will the consequences of your action/
inaction be good?” and “Will the consequences of
your action/inaction result in the greatest good for the
greatest numbers?”
Deontological theories: Here, the t heories focus on the
“duties” of the people/actors who are in a position to
make a decision in relation to an ethical issue or ethi-
cal dilemma. In such theories, the ends do not justify
the means, but rather, the emphasis is on ensuring
that the means themselves are ethical/good. Deonto-
logical approaches to ethics, therefore, ask questions
such as, “Are you acting in accordance/compliance
with your duties when you act or refrain from acting
in a specific way?”
Ethics of character:
Virtue ethics: Vir tue ethics places emphasis on the mor-
al character of a person when analyzing ethical issues.
In a virtue ethics approach, questions asked can in-
clude “Would it be kind/caring/generous/benevolent
to act or not act in a specific way?” Scholars believe
that several Eastern ethical theories and associated
approaches (for example, in Confucianism) are also
closely linked to ethics of character and virtue ethics,
i.e., giving primar y importance to “internal v irtues”
of an actor/scientist, such as compassion, responsibil-
ity, and honesty [35].
Consequentialist, deontological, and virtue-based theo-
ries are all categorized under the umbrella of “normative”
ethics,” i.e., the well-established methods and steps one can
use to determine whether a course of action or inaction is
“good” or “bad.” Normative ethics is distinguished from
“metaethics” as well as from “applied ethics.” Metaethics
aims to understand the philosophical and cultural founda-
tions and origins of ethics and moral values that are domi-
nant in any society. It asks fundamental questions, such
as “What is morality?” and considers the relationship and
impact of convictions and long standing societal (and even
academic) expectations and codes that drive one to label
specific behaviors as et hical or unethical in the first place.
Applied et hics, on the other hand, caters to specif ic human
institutions, organizations, and actions and is aimed at
translating ethical pr inciples into practical applications rel-
evant for specific contexts. What we are attempting in this
article can therefore be categorized as applied ethics.
In AI4EO, the presence of a plethora of uncertainties man-
dates looking beyond consequences to duties (social re-
sponsibilities), intentions, knowledge/human values, and
concrete present action: indeed, in several, if not most,
instances, the (long-term) consequences of conducting re-
search and translating research results into concrete prod-
ucts and ser vices will remain unknown until a much later
date. A focus on consequentialist theories may not prov ide
meaningful guidance, especially at early stages of research.
Being aware of one’s broad duties, while very significant,
may also not be adequate. Indeed, guidelines for research
ethics, which provide, among other things, methodological
and procedural g uidelines on how and how not to conduct
research (e.g., obtaining informed consent from research
participants and avoiding plagiar ism), enumerate the broad
duties of researchers. Fundamental principles of ethics, e.g.,
honesty, integrity, and fairness, are also basic duties and
characteristics of any ethical researcher. However, these du-
ties are abstract and may be inadequate to g uide researchers
working with emerging technologies, such as AI4EO.
In this situation, approaches that go beyond duties and
consequences can be both relevant and useful. In Eastern
philosophy, for example, considerable emphasis is placed
on the interlinked concept s of karma (action, its causes, a nd
consequences) and dharma (personal dut y, social responsi-
bility, and religious dictates). Considerable emphasis is also
placed on “our three powers” (the three levels of shakti):
1) intention/desire/will (“iccha shakti”), 2) action (“kriya
shakti”), and 3) knowledge/human values (“gyana shak-
ti”). In Indian mythology, for example, symbols associated
with Skanda (“Lord Murugan”) vividly describe t he link
among these three powers: knowledge and human values
are meaningless if they are unaccompanied by appropriate
intention and will, i.e., a desire to act, and constructive ac-
tion. Similarly, will and action must be guided by knowl-
edge and human values for long-term benefits for oneself
and society [35], [36], [37], [38]. Neo-Confucian scholars
from China also emphasize a combination of action and
knowledge to realize morally appropriate outcomes [39].
“Knowledge,” in this context, can relate to a broad un-
derstanding of fundamental ethical principles, human val-
ues, and human rights and, in a more abstract sense, the
endeavor to develop a deeper understanding of one’s self or
true nature. Since this abstrac t understanding of knowledge
is beyond the scope of this article, we focus instead on the
concrete understanding of fundamental ethical principles
IEEE Proof
and associated human values and human rights. This foun-
dational “knowledge” and associated understanding of
ethics can then pave the way toward the identification of
ethical issues and opportunities in a systematic manner
and on a case-by-case basis.
Academic and scientific research focusing on ethical issues
emerging from AI4EO research is scant. The most promi-
nently identified issues relate to privacy, particularly in
light of the increasing resolution of EO and RS images [40].
However, as diverse sources of data as well as A I and ML
models are increasingly used with EO/RS data, the number
and breadth of ethical issues and concer ns are expected to
rise. It is therefore not possible to immediately and compre-
hensively identify all the ethical issues that may arise in the
context of AI4EO research in the (near) future.
Accordingly, in this ar ticle, following t he steps explained
in the “Methodolog y” section, we compile and categorize
the most common ethical issues and associated guidelines
extracted from the previously mentioned five sources, un-
der six categories. These six categories cor respond with the
three fundamental principles (or values) of ethics (namely,
honesty, integrity, and fairness) and the three most promi-
nent emerging values, namely, privacy, responsibility (in-
cluding the duty of care and respect for human rights), and
sustainability. We are of the view that most, if not all, ethi-
cal issues and associated g uidelines that may arise in the
future can also be categorized under one of these six broad
categories. However, the number of et hical issues that can
be categorized under one or more of the six categories may
grow with expanding tec hnological capabilities in A I4EO
research and with expanding human imagination.
After categorizing the ethical issues and associated guide-
lines, we e xplain each wit h examples from AI4EO research.
The aim is to assist AI4EO scientists acquire a better and
more practical understanding of the ethic al issues/guidelines
that are most immediately relevant for their research. This
understanding constitutes the first step to identifying ethi-
cal issues and opportunities in any ongoing AI4EO research.
The six f undamental principles, or values, of ethics and the
most prominent ethical issues and guidelines that can be
categorized thereunder are enumerated in the following
and thereaf ter e xplained in the conte xt of AI4EO research,
with concrete e xamples. These examples are der ived either
from a review of the literature or from discussions and de-
liberations with var ious AI4EO scientists over a period of
18 months (June 2020–December 2021):
1) Privacy:
autonomy and freedom
self-determination, including national sovereignty
data governance, ownership, and licensing
2) Honesty:
transparenc y
data veracity
3) Integrity:
technical robustness
safety/security, including national security
4) Fairness:
(non)bias (including nonbias in training data)
nondiscrimination and diversity
sociocultural sensitivities
democratic creation of standards
5) Responsibility:
human agency and oversight
duty of care
ensuring social securit y and cohesion
6) Sustainability:
scientific, social, and cultural (including through
education and training for the next generation of AI
sc i ent i s t s)
For the purpose of understanding and application, the
listed fundamental values/principles that form the six
main categories (i.e., honest y, integrity, fairness, responsi-
bility, sustainability, and ensuring privac y where relevant)
can be considered the major ethical duties of a researcher.
Contextual and appropriate compliance with these major
duties will usually lead to compliance with other ethical
duties and with ethical consequences listed as various sub-
categories. As a corollary, noncompliance with these major
ethical principles/duties is likely to lead to the opposite,
i.e., to unethical consequences and noncompliance with
other ethical duties. In the following, for example, compli-
ance with the duty of privacy will ensure nonstigmatiza-
tion, autonomy, freedom, and self-determination for those
affected by research and for those from whom data have
been taken for research. Noncompliance with privacy will
lead to the opposite, i.e., stigmatization, a lack of (or tak-
ing of) autonomy and freedom, and so on. For a stepwise
procedure on how to study and use the information and
explanations contained in this section, see the flowchart
in Figure 1.
Before proceeding, it is relevant to note the following.
First, needless to say, various e xpert groups have placed
ethical issues and guidelines linked to (big) data and AI
under categories that differ from the six we use here. Also,
there is a broad understanding that ethical issues (placed
under these six categories) can and do differ based on
technology and context. Second, it is useful to reiterate
that the list of et hical issues under the six broad catego-
ries is not a closed set. The ethical issues and associated
consequences are potentially ever expanding, limited
IEEE Proof
only by human imagination. Third, the ethical issues
enumerated under the six categories are not independent
and exclusive sets. Instead, in our exercise of explaining
each of the major ethical issues through examples, it is
clear that these issues overlap and can be placed under
more than one categor y. Finally, it is also necessary to
note that these issues can arise not only at the real-world
application stage of any AI4EO research and development
endeavor but also at the conceptual and/or fundamental
research stages.
Privacy has been a central value guiding the development,
implementation, and popular adoption of technologies as
old and fundamental as telephony and telegraphy as well as
more recent innovations, such as the Internet [41]. Indeed,
technologies a nd applications that fail to protect privacy may
FIGURE 1. A workflow suggestion for AI4EO researchers to identify ethical issues in their research.
Identify and List Your Ethical
Duties as a Researcher
Check Whether Yo ur Research
Goals and Steps are Aligned With
Your Ethical Duties
I Am Aware
of All My
They are
Reframe Your Research Question
to Avoid or Overcome Ethical
Issues and Dilemmas
Identify and List All Possible
Ethical Consequences
of Your Research
Read and Understand
Ethical Issues Listed
in the “Categorizing and
Understanding Artificial
Intelligence and Data Ethics
for Artificial Intelligence for
Earth Observation
Research” and “Ethical
Opportunities in Artificial
Intelligence for Earth
Observation Research”
Reread the “Categorizing and
Understand Artificial Intelligence
and Data Ethics for Artificial
Intelligence for Earth
Observation Research” and
“Ethical Opportunities in
Artificial Intelligence for
Earth Observation Research”
Sections and/or Discuss Your
Work With an Ethics Expert
Identify Research Goals and
Steps that are Likely to Lead to
any Unethical Consequences
I Am able to
I Am
Aware of
All Ethical
Ye s
Ye s
Ye s
Ye sProceed With
Your Research
IEEE Proof
fall into disrepute and lead to declining user acceptance [42]
over time. In the context of EO data, privacy concerns had
been relatively unheard of [43], [44], [45]. However, with
growing image resolutions and the emergence of drone tech-
nologies in RS, privacy concerns have become starker and
clearer [40], [46]. The problem is aggravated, as images are re-
trieved at a known point in time (when), at a specific known
location (where), and include views of potentially detectable
neighborhoods, residences, and even people (who) [40],
[47]. With this information and additional contextual infor-
mation, it may also be possible to detect specif ic activities
(what) (e.g., gardening , sunbathing, barbecuing, and so on).
In present times, privacy concerns are also commonly
flagged in most tec hnologies that rely on big data. Accord-
ingly, in AI4EO research, which uses an ever-increasing va-
riety of complementary (big) data sources, privacy concerns
may take on various dimensions based on the type of data
used and the purpose for which they are used. Today, very
high-resolution EO data are capable of characterizing indi-
vidual buildings, and sensing techniques may even be capa-
ble of assessing building quality [48]. At these detailed spa-
tial and thematic levels, the identification and analysis of
objects (e.g., houses and localities) as well as the subsequent
use to whic h an analysis is put may lead to the violation
of privac y. Undoubtedly, the nature of data (whatever their
source may be) required in various EO/RS applications is of-
ten quite different from data used in, say, the medical field.
For example, geoinformation data from social media plat-
forms, such as Twitter and Flickr [49], [50], mined with the
aim of labeling buildings and identifying areas where rescue
effor ts need to be expedited, is of a different nature than in-
dividuals’ mental health information mined from the same
platforms [51]. In this conte xt, recent research refers to Nis-
senbaum’s theory of “privacy as contextual integrity” and
suggests that this can be “a useful heuristic to guide ethical
decision-making in big data research projects” [52].
Based on contex t, the collect ion, storing, processing, and
transfer of data can be done with much greater attention to
legal and ethical requirements by (for example) adopting
supporting technologies that give data contributors greater
control over their (personal) data, on the one hand, and
support the development of the European Data Economy,
on the other hand (see the “Ethical Opportunities in Ar ti-
ficial Intelligence for Ear th Obser vation Research” section
for fur ther discussion on this). However, as we see in the
context of other ethical concerns and risks, a technical solu-
tion is not always possible. What may be necessar y in such
situations is a fundamental rethinking of the problems on
the ground and how research questions can be redesigned
to specifically address those concerns. Examples in this re-
gard are provided in the “Et hical Opportunities in Artificial
Intelligence for Eart h Observation Research” section.
Closely linked to the issue of privacy is the issue of stig-
matization. Take for example, the case of geotagged social
media data from platforms such as Twitter and Flickr. Such
data are being used with EO/RS data for a wide variety of
uses ranging from disaster relief [53] to population density
estimates in various regions [54], [55]. The use of geotagged
social media data for disaster relief may be welcomed by
the users whose lives are rescued as a consequence of such
use. However, in case of a negative assessment of housing
qualit y and in the case of slum identification, for example,
the classif ication of cer tain users (people) as slum dwellers
may not only be considered a breach of privacy but may
also lead to stigmatization of specific individuals, commu-
nities, and regions.
Here, we see that even routine tasks, such as “data label-
ing,” can result in stigmatization. Traditionally, the labeling
of large pixelated EO/RS data was done with the suppor t of
so-called ground truth data [55], [56], i.e., confirmations
and rejections received from the ground as to the accuracy
of predictions made from EO images. When these ground
truth data are not available or are expensive to collect com-
prehensively, researchers use visual interpretation based on
their subjective perceptions and experiences, or they use
social media data, census data, and various ML/AI models
to complement and confirm EO data and to make predic-
tions from EO data. While earlier EO/RS work consisted of
broad land cover/land use classifications at low resolution
[49], in the present day, these classifications become more
and more refined, increasingly supporting the identifica-
tion of regions that are more prone to crime [50], fire [57],
and economically deprived living areas, commonly called
slums [58], [59].
Labels, whet her they are based on “ground truth” or oth-
er sources/algorithms, need to be worded and placed with
an ethically mindful eye. Here, ethics requires not only ac-
curacy but also sensitivity. For example, t he understanding
of what constitutes a “slum” can var y based on countr y, cul-
ture, and geography as well as socioeconomic realities of a
region [60] and the experiences of people placing the labels
[61]. In this context, as discussed in greater detail in the
“Ethical Opportunities in A rtif icial Intelligence for Earth
Obser vation Researc h” section, translational, or applied,
ethics may require that stigmatizing labels be replaced with
constr uctive and affirmative ones. Translational ethics calls
for the actual application and implementation of theory to
practice or, in other words, ensuring that ethical t heories
and principles can be and are translated to real-world appli-
cations and practices. In medical et hics, for example, trans-
lational ethics calls for the application of science/scientific
findings to clinical practice.
While stigmatization is an ethical issue that will arg u-
ably arise only at a very advanced (research) application
stage, its avoidance is possible only if researchers are alert
to the probable reality of stigmatization (and the various
sociocultural and economic harms that can result from it)
at a very early stage of research—the stage where individ-
ual researchers have the possibilit y of applying labels re-
sponsibly and constr uctively rather than merely based on
IEEE Proof
convenience and disciplinar y custom. This is also discussed
in greater detail in the context of bias in the following and
in the “Ethical Opportunities in Artificial Intelligence for
Earth Obser vation Researc h” section.
One of the fundamental guiding ideas of data et hics is “en-
hancing protection for individual freedom, self-determi-
nation, and integrity” [62]. When data and associated data
analysis limit individual freedom and self-determination,
the principles of ethics can be breached while also violat-
ing basic human rights. Take again, for example, the issue
of slum identification using AI4EO. Under international
pressure, country governments may use research findings
to clear areas labeled as slums, dislocating thousands from
their homes and creating significant social unrest. Or gov-
ern ments may be forced to impr ison illegal a rtisan al miners
and deny them their traditional source of livelihood. In the
face of such possibilities, AI4EO researchers need to ensure
that their researc h is mindful of “ground truths” beyond
the accurac y of labels. Responsible research also needs to
go beyond avoiding ethical risks (e.g., by hiding and not
disclosing geolocations) to actively changing research ap-
proaches to address ethical concerns on the ground. We
discuss this in greater detail in the “Ethical Opportunities
in Ar tificial Intelligence for Earth Observation Research”
When applied to national contexts, the issue of self-
determination and autonomy can assume the form of
national sovereignty, which can be (adversely or appro-
priately impacted) through inter national sociopolitical
pressure. For example, consider the issue of nuclear dis-
armament: India’s preparations to make a first effort at
testing its nuclear weapons were discovered by U.S. intel-
ligence via satellite-based monitoring of the test station at
Pokhran, Rajasthan. The discovery of these plans led to
large-scale international pressure to withdraw the effor ts.
In a subsequent attempt, great pains were taken by Indian
nuclear and defense scientists to prevent the plans from
being discovered by U.S. satellites, leading to successful
testing wit hout intervention from international pressure.
Despite subsequent international condemnation for going
against the ideal of nuclear disarmament, India justified
its stand based on rising armed attac ks and threats from
Pakistan and China at its borders [63]. (The successful
nuclear tests were conducted in a desert area of Western
India. The efforts of Indian scientists to “hide” from U.S.
satellites to prepare for and conduct the tests were recently
dramatized in a popular Hindi movie, Paramanu: The Story
of Pokhran.)
Data gover nance has been defined as the “exercise of au-
thorit y and control over the management of data,” with the
purpose of increasing “the value of data and minimizing
data-related cost and risk” [64]. Data governance therefore
also includes the management of data in accordance with
the requirements of the law. In the EU, data are protected
under various laws, including t he General Data Protection
Regulation (GDPR) and the Directive on Copyright in the
Digital Single Market (DSM). Accordingly, data governance
requires the management of issues of data privacy as well as
data ownership.
In the conte xt of data ownership, European Parlia-
ment directive 2019/790 on copyright and related rights
in the DSM (the so-called DSM Directive) includes two
exceptions for “text and data mining” (see articles 3 and 4
and recitals 11, 14, 15, 17, and 18). These exceptions per-
mit text and data mining for scientific researc h pur poses
and are also impor tant for AI/ML research that relies on
(big) data [65]. Absent these e xceptions, the mining of
text and data from any source could amount to copyright
and database rights infringement. While these exceptions
were introduced to support, among other things, R&D,
including in the context of AI/ML research [66], any entity
mining and utilizing such data still needs to be mindful
of privac y considerations. Whether the subject concerns
images of people from high-resolution EO data or indi-
vidual (image) data mined from social media platforms,
the privacy of the people concerned (and of their data)
needs to be secured. In addition to being a legal require-
ment under the GDPR, privacy is a human right and must
also be secured as a matter of ethics, even in the absence
of legal regulations in most jurisdictions. Arguably, there-
fore, the ethical standard of privac y may be higher than
the legal standard. For example, the arg ument that data
are already public (e.g., since they were posted on Twitter)
may be adequate to avoid trouble with law enforcement,
but it may not help overcome ethical concerns over pri-
vacy [67], [68].
Beyond the GDPR, ethics requires that when data are
collected without the consent of people whose lives are
closely affected and associated with the data, the analysis
and dissemination of such data must not adversely impact
the fundamental human rights and welfare of such people.
This is centrally relevant, as EO data as well as AI4EO data
are being used to make polic y recommendations linked to
the United Nations (UN) Sustainable Development Goals
(SDGs). EO scientists may see this issue come into play
when their research helps identify slum regions and ille-
gal mining sites; the research may lead to policy decisions
whereby large numbers of people living in identified slums
or deriving their livelihood from small-scale artisanal min-
ing lose their homes and means of livelihood. A responsible
restructuring of research questions and research designs
may be necessary to avoid such dilemmas. Similarly, in the
context of agriculture, the EU code of conduct for ag ric ul-
tural data sharing emphasizes that “the farmer remains at
the heart of the collection, processing and management of
agricultural data” [69]. The code recognizes that farmers
who share data have the right to know and participate in
IEEE Proof
the uses to which their data are put. In AI4EO, when data
are collected remotely and without the knowledge of farm-
ers, ethical issues become even starker if the data analysis
adversely affects farmers’ (or farming communities’) right
to self-determination.
Honesty is one of the central and, therefore, most funda-
mental values of (business) ethics [13]. It highlights the
importance not only of tr uthfulness but also of trustwor-
thiness [70], [71]. Accordingly, hiding a relevant fact and
failing to disclose it raises as much of an ethical concern
as actively miscommunicating and misrepresenting a fact.
In AI4EO, the principle of honest y requires scientists to be
transparent about the (possible) shortcomings and limita-
tions of their reasoning as well as the limits to which pre-
dictions based on their research may (or may not) be ac-
curate. Therefore, issues of transparency, explicabilit y, and
data veracity are closely connected to the broader issue of
honesty. Further, as A I4EO research is replete with uncer-
tainties, honesty would require a kind of “laying bare” of
these uncer tainties so that future researchers, developers/
innovators, and policy makers could take necessar y precau-
tions while relying on the findings and products/services
based on them.
In the conte xt of performance prediction (for an AI4EO
model), honesty also comes into play when compiling
training datasets and test datasets (i.e., data used to ver-
ify the accurac y of the training data). If pixel values cor-
responding to testing sets are partly or completely seen
during the training phase (i.e., as par t of the test data), the
number of independent testing pixels gets reduced and can
lead to overoptimistic accurac y assessments. Good scien-
tific practice requires that test samples remain unseen dur-
ing training. This is important, as high-capacity models (as
with deep neural net works, with their thousands of param-
eters) are very good at over fitting training data. If the data
points in training and test subsets come from the same or
an overlapping distribution, good test accurac y/high confi-
dence in test samples does not serve as a good predictor of
performance on unseen data.
Honesty requires that AI4EO scientists e xercise caution
while selecting training and test datasets and be transparent
about the scope and overlap in suc h sets [72], [73]. Discus-
sions with AI4EO scientists revealed that overlapping test
and training datasets may result not from active dishonesty
but because scientists may not be fully aware of good scien-
tific practice in deep lear ning (or due to lack of experience/
inadequate competence, especially in the case of young sci-
entists). Under both Western and Eastern theories of ethics
(especially deontological theories), it is the duty of scien-
tists to fully educate themselves on good scientific practices
relevant to their field of work. Accordingly, any argument
based on inadequate competence and gaps in knowledge
of good scientif ic practices may not release one from ethi-
cal duties and from the consequences of “inadvertent”
dishonesty in research. At t he ver y least, a researcher will
lose credibility, affecting his/her long-term career.
In line with the requirement of honesty, transparency calls
on scient ists engaged w ith AI4EO researc h to be transpa rent
about the scope and limitations of their research. This can
include, for e xample, explanations clarifying why specific
sources of data are selected instead of or to the exclusion of
others, how scientists are e xtracting such data (e.g., which
methods are chosen and which ones are not) [74], what
parameters guide the labeling and analysis of a resulting
dataset, and what the scope and limitations of the resulting
model and analysis are. In this context, the “Et hics Guide-
lines for Trustworthy AI” [75] state that “humans need to
be … informed of the system’s capabilities and limitations.”
A challenge in t he search for scientific truth is that clas-
sified and quantif iable results are not always unambiguous
because, for example, the target object of interpretation is
conceptually ambiguous. Especially in spatial analyses,
there is no unambiguous truth. Take the terms monocen-
tric and polycentric as examples. Conceptually, monocentric
refers to a center where urban functions are concentrated,
and polycentric refers to a reduced dominance of the urban
center, giving way to many suburban local centers. While
this distinction seems conceptually clear, in spatial science,
quantified results are ambiguous in the sense that there can
be no absolute truth as to when a dominant center loses
its dominance. In such cases, therefore, honesty simply re-
quires that academic researchers declare this ambiguity so
that downstream reliance on associated research can take
this g round reality into account. As discussed in the “Ethi-
cal Opportunities in Artificial Intelligence for Earth Obser-
vation Researc h” section, such ground-level ambiguities
can also be addressed by delimiting the focus and intention
of corresponding AI4EO research.
Transparency, when taken a step further in the direction
of decisional transparency, particularly in the conte xt of
AI applications using big data [76] to reach decisions and
recommendations, becomes the requirement of explain-
ability, or explicability. Undoubtedly, not all AI and AI4EO
applications require explainability. Indeed, explainability
may not be necessary in all instances, and it may not serve
any additional purpose. In recent times, scholars have also
recommended systems that rely on verif iability to comple-
ment and substitute complex (and arguably unnecessary)
explainability requirements [77]. However, whether ex-
plainability is or is not an ethical requirement may depend
on the AI4EO use case. For example, if an EO/RS data-based
scoring method is used to determine which (rural) areas
ought to be prioritized for electrification, the specific types
of data and t he method adopted to reach that score as well
as the parameters that are considered and those that are ex-
cluded need to be made transparent and explainable [78].
IEEE Proof
This issue is also closely linked to the issue of democratic
decision making (see the following) and inclusive growth,
both of which can be either suppor ted or severely compro-
mised by A I4EO research.
Since AI/ML algorithms are used to train models using data
from one source/context to make predictions in various
other contexts, the issue of the veracity 1) of the data used
to train the models, 2) of the presumptions underly ing the
models, and 3) of the resulting predictions becomes very
relevant. Especially when using data from novel sources,
such as texts from social media data, to support building-
level population density estimates, the limitations of natu-
ral language processing, the absence of accurate geolocation
tags (e.g., due to Twitter’s decision to remove support for pre-
cise geotagging [79]), the user group of the online platform
(which may not be representative of the overall population),
and the distortions generated by specific users of the plat-
form need to be borne in mind while relying on such data
for training and analysis. The issue of data veracity can also
arise as a result of inaccurate and insensitive data labeling
(see the “Stigmatization” section). Needless to say, “truth”
can have several impor tant and justif iable subjective/local
considerations: what is truthf ully labeled as a slum from
the perspective of a European researcher may be t he average
standard of living for a South Asian middle-class family liv-
ing in a congested neighborhood. Contextual data veracity,
however, is important to avoid problems, such as the previ-
ously mentioned issue of stigma. The importance of conte xt
is also seen in archaeological geospatial data [80].
The term integrity has multiple meanings depending on
context. In a simple business ethics sense, it can be un-
derstood as “walking t he talk,” or actually doing or imple-
menting what one preaches to be “right” or “correct” [81],
[82]. In the technical sense, the word integrity is often used
in the conte xt of “system integrity” [83], [84], which es-
sentially means that the system, algorithm, AI/ML model,
and resulting application performs in the manner it is
meant to perform and that its outputs/predictions are ac-
curate to the extent the creators and owners of the system
claim it to be. System reliability, safety, secur ity, accurac y,
and overall tec hnical robustness are therefore subjects
that are closely interlinked, have important ethical impli-
cations, and can be categorized under the broad heading
of “integrity.”
In practice, integrity is also closely associated with the
previously discussed duty of “honesty” [85]. Specifically, if
a system does not perform accurately or does not do so in
specific circumstances, the ethical bar of “integrit y” requires
that one is “ honest” and transparent about the circ umstances
and extent to which t he system can/does give accurate results
[86] and/or the e xtent to which one can rely on the research
findings and applications. The implications of meeting or
not meeting the “integrity” bar also vary from application to
application, depending on the real-world end uses to which
the application is put (or intended to be put) and the impact
such an application has or is likely to have on actual human,
animal, and environmental well-being [87].
Investigations into “integr ity” can be translated into ques-
tions such as, “Does the system do what it claims to do?”
“What are the limitations?” “What is the possibility and
level of error in any result?” In EO and RS data, especially
including EO/RS data that are compiled and assessed with
the help of AI/ML models, there is a well-known problem
of uncertainty. What e xactly is the ML model learning
(from) and on what basis is the resulting image produced?
For example, for many years, it was believed, based on EO
data, that the phytoplankton concentration was higher
near the poles [88]. However, it was later discovered that
this apparent higher concentration resulted from errors in
the atmospheric correc tion software used in the application
[89]. To the extent that such measures may lead to major
national and international (R&D) investments and changes
in governmental policies, integrity requires that the scien-
tific community at least declare the probability and, where
possible, the level of error and uncertainty and identify
the limitations of its research findings. This will not only
permit more responsible research spending but also en-
sure that policy decisions are not made merely on a kind of
“blind faith” in scientific findings [88].
Needless to say, issues of (system) integrit y are closely re-
lated to safety and security. The failure of any system to
perform in the manner in which it is meant to perform can
lead to the loss of human/animal lives and to environmen-
tal and infrastructural damage. The failure and inadequa-
cy of data governance structures can also lead to security
breaches, further leading to compromised privacy and vio-
lations of ownership rights. Given the scale of AI4EO re-
search, the scale of security issues that can arise from such
research is equally high.
With the possibility of f using diverse data sources,
AI4EO research is marching boldly toward novel use cases
that were unheard of just a few decades ago. AI4EO is ex-
tensively used in defense and law enforcement to identify
enemy/unaut horized vessels (aircraft and ships), track ve-
hicles transporting illegal goods and suspected criminals,
remotely identif y car license plates, and predict possible
target areas during wars and other armed conflicts. These
use cases of AI4EO, while ethically questionable, are largely
accepted to secure nations and people.
Yet, more novel use cases of AI4EO, such as large-scale
projects aimed at predicting building-level population den-
sity, can lead to new issues of national security, especially if
resulting data and maps are publicly shared. For example,
the availability of such data can make specific buildings
IEEE Proof
more prone to terrorist attacks. In this context, therefore,
the prev iously discussed issues of data governance, includ-
ing decisions as to how, when, and to what extent to make
data public, must be given central priority. While open
source data are appreciated by the research community, is-
sues of national security may require more responsible dis-
closure of data and predictions.
Fairness is another one of the fundamental building blocks
of (business) ethics. Borrowing from Kant’s categorical
imperatives, fairness can be judged based on whether the
person creating a rule or taking an action would find the
same rule acceptable/appropriate if it were universally ap-
plied, including to him- or herself. Fairness is closely linked
with treating equals equally [83]. At the same time, it also
implies, as does the fundamental right to equality, that un-
equals not be treated equally [83]. Thus, for example, crimi-
nal law does not prescribe the same punishment for juve-
nile delinquents as it does to adult recidivists, even if the
nature of t he crime is identical. Similarly, in AI and EO, the
principle of fairness requires, again, that appropriate mea-
sures be taken, for example, while labeling data and while
making recommendations, to ensure that regions that are
in similar circumstances are treated similarly. At the same
time, the highest standards of ethics would also call for AI
and EO research to rethink research designs in ways that can
uplift marginalized segments and bring them equal status.
The principle of fairness, therefore, can be applied not only
at the time of labeling data (e.g., in the context of labeling
crime-ridden areas and slums) but also at the time of select-
ing research focus areas, allocating research funding, and
converting research findings into practical applications.
Given the scope and reach of EO/RS
research in terms of the sheer areas
it (potentially) covers, issues of fair-
ness need to be flagged at all stages of
research and all stages of data collec-
tion and analysis.
Issues of fairness also arise at
the macro level when we consider
the availability of EO/RS data and
the means to meaningfully access,
analyze, and use these data. Take
the example of climate c hange: a
study of anthropogenic inf luences
driving climate change reveals that
it is primarily industrialized na-
tions that contribute to the problem
[89]; they have significantly higher
energy consumption, carbon diox-
ide (CO2) emissions (see Figure 2),
and greenhouse gas production. Yet,
the consequences of climate change
are a burden on all nations in the
world. An important task of RS is to
monitor and collect georeferenced data and datasets on
various parameters of anthropological climate change.
These datasets are the key information source for ML mod-
els and forecast systems of climate change consequences.
Therefore, the progress of climate change monitoring de-
pends highly on the availabilit y of such datasets.
What we observe, however, is a correlation bet ween the
number of satellite missions that a country owns and the
level of the country ’s development (see Figure 3). This trend
is also visible in the availability of georeferenced datasets
related to that country. In Figure 4, it is clearly visible that
regions of high development (e.g., North America, China,
and Europe) are significantly more represented in georef-
erenced datasets, independent of the degree to which they
are affected by climate change. In Figures 2 and 3, one also
obser ves that countries with a lower Human Development
Index (HDI) score have significantly lower CO2 emissions
and are also significantly less represented in georeferenced
datasets. Yet, these countr ies are as affected by climate
change as the countr ies with higher emission rates. Fairness
requires that more industrialized nations take responsibil-
ity to assist poorer nations in combating climate change, for
example, by helping them gain access to relevant data from
their regions.
It is also necessary to recognize this situation as a lost
opportunity to learn from “best practices” for climate
change mitigation that may be adopted in developing coun-
tries but unknown in Europe and other developed regions
of the world. In India, for e xample, there is rapid farmer
adoption of agroecolog y (a system of farming) under the
name of natural farming [90]. The availability of data from
regions that have successfully migrated to these sustainable
farming systems can be hugely beneficial for Europe, where
FIGURE 2. The cumulative CO2 emissions by country [89]. Cumulative emissions are calcu-
lated as the sum of annual emissions from 1750 to a given year.
No Data 0 0.5 1 2.5
Source: Our World in Data based on the Global Carbon Project • CC BY
IEEE Proof
policy makers are actively seeking means of optimizing
farmer migration to agroecological approaches for sustain-
able farming.
Bias is an ethical issue that has frequently come into the
limelight in relation to AI applications ranging from Ama-
zon’s AI-based recruitment tool that showed bias against
women [96] to the PUL SE Photo Upsampling via Latent
Space Exploration (PULSE) algorithm that upscaled low-
resolution images of celebrities, such as Barack Obama and
Lucy Liu, into images that were distinctly “White” [97]. Bias
in AI applications frequently results from the diversity (or
lack of it) in the data used to “train” AI and ML models.
This lack of diversity and the resulting bias can arise from
using data from any source [including, in our case, visually
and automatically generated labels from RS data, from offi-
cial (cadastral) data databases [98], and from other sources,
such as social media platforms] to train (among others) ML
models [76], [99].
Based on the training data used to train A I/ML models,
various biases might creep into a system; extending beyond
a preference for white/ma le candidates for jobs , they can lead
to unfair categorizations of areas with higher immigrant
populations as “slums” and “crime ridden” [100], [101].
The ethical concern of fair ness therefore alerts the AI/ML
communit y to the high possibility (and probability) of bias
(including bias in training data, in data labels, and so on).
It also actively seeks behavior, labeling, and coding that en-
sure diversity, are nondiscriminator y, and take into account
sociocultural as well as economic and political sensitivities.
Closely connected to the concept of fairness, therefore, is
the issue of the democratic creation of standards [102] (see
the following for further discussion), whether they are stan-
dards to be followed for the selection of training data or for
labeling data used to train AI/ML models.
Fairness also requires that issues of data ownership
and licensing (e.g., of intellectual property) are clearly and
fairly addressed. Furt her, when supplementary sources of
data, such as data from social media platforms, are used
to complement EO/RS images and increase their accuracy
and resolution, it is necessary that the terms under which
such data are acquired are in keeping with the licenses of
the social media platfor m, on the one hand, and the spirit
of broader regulations, such as the GDPR, on the other.
Bias may also arise from personal and cultural prior knowl-
edge. In computer-based pretrained categories, this will
simplify, or bias, the world. A sort of “selection bias” can
also result from the convenient selection of specific sources
of data to the exclusion of others. Indeed, selection bias
may also be inevitable due to a lack of data from specific
(geographic) locations (see the “Fairness” section with re-
gard to the availability of EO/RS satellites and data. In R S,
appropriate datasets for specif ic tasks may also be neglected
due to the limited availability of researc h funds. Therefore,
the selection of any specific source (and even region) for
data collection needs to be disclosed, and the analysis re-
sulting therefrom must clarif y the scope of applicability
(e.g., geographic, cultural, and other limits) and any in-
herent limitations in the findings and the algorithms used
for analysis. This is necessary to avoid the danger of broad
policy making based on limited studies. It is also necessar y
to ensure responsible downstream/applied research.
However, the major challenge in the conte xt of bias in
training data results from a lack of
open, accessible, representative, and
generalizable data [103], [104], [105]
to permit optimal training and ap-
propriate contextual labeling and
semantic segmentation. Further, da-
tasets (e.g., in scientific publications)
usually cover only a small range of
scenes/cases in which the predictive
models and algorit hms perfor m well.
As a result, the data used to under-
stand the fundamental conte xt/rela-
tions in EO may be inadequate and
totally inappropriate to make predic-
tions in other conte xts, even where
these contexts are somewhat similar.
For example, training on crop f ields
in moderate climate regions cannot
be applied in tropical regions. Simi-
larly, data used to train A I/ML mod-
els in, say, identifying building types
are not ref lective and representative
of what is seen in diverse real-world
FIGURE 3. The number of occurrences and therefore the presence of a country in RS and geo-
graphic information systems datasets in [94] and [95].
Occurrences of the Country in GIS and Remote Sensing Data Sets
Occurrences in Data Sets
Map: Matthias Kahl • Created With Datawrapper
IEEE Proof
contexts. While data scarcity is a reality, it is the authors
responsibility to clearly and honestly state the e xtents and
limitations of their datasets, level of diversity, and overall
approach as well as resulting limitations in the scope and
applicability of the research models/findings.
It is important here to draw a distinction while acknowl-
edging the relationship between datasets and models. It may
be difficult to identify limitations of models (e.g., of feature
extractor networks trained in an unsupervised manner),
but the scope of the dataset used to train a model is always
known. The limitations of the dataset are also likely to im-
pact the output provided by the model trained using that da-
taset. Therefore, even in models trained in an unsuper vised
manner, the underlying dataset and its limitations must be
identif ied and disclosed. In the case of AI in health care, for
example, it has been discovered that models trained primar-
ily on datasets consisting of male heart patients are unable
to accurately predict heart disease in women [106]. Here, the
limitations of the dataset (and by implication, of the model)
can be known and predicted if the scientists and developers
involved clearly state that most of the training data are from
male heart patients so that doctors are more careful when
relying on this dataset when faced with a female patient.
Similarly, closer to home, i.e., in EO research, if a model for
predicting slum areas is trained on data from Brazil, it is nec-
essary to identify the scope and origin of the dataset so that
it is not (fully) relied on to predict slum areas in Nigeria and
other regions of the world where the population density, lo-
cal cultures, and average income levels may be very different.
To increase t he diversity of available data, methods to in-
centivize and ma ximize contributions of more diverse data
(in a manner that secures the privac y of contributors) from
diverse global locations and circumstances may need to be
researched to tap into the full potential (ethical opportu-
nities) offered by AI4EO research. Further, in the contex t
of diversity, the impact of technolog y on gender and vice
versa (i.e., the impact of gender on technology) is an impor-
tant area of research. The impact of technology on gender
FIGURE 4. The correlation between the relative number of satellites in relation to the number of inhabitants of a country and the country’s
ranking in the Human Development Index (HDI). A correlation between the development of the country and its number of owned satellites
can be seen. The chart is based on the analysis of the satellite catalogue in [91], the number of inhabitants for each country in [92], and the
ranking of the countries HDI in [93]. SUI: Switzerland; BEL: Belgium: LTU: Lithuania; ESP: Spain; AUT: Austria; KSA: Saudi Arabia; SVK: Slo-
vakia; MAS: Malaysia; CRC: Costa Rica; ROU: Romania: AZE: Azerbaijan; THA: Thailand; PER: Peru; ECU: Ecuador; ALG: Algeria; IRL: Ireland;
MGL: Mongolia; PAR: Paraguay; TKM: Turkmenistan; INA: Indonesia; RSA: South Africa; VIE: Vietnam; MAR: Morocco; BHU: Bhutan; USA:
United States; GBR: Great Britain; RUS: Russia; NOR: Norway; FIN: Finland; ISR: Israel; NZL: New Zealand; SGP: Singapore; DEN: Denmark;
JPN: Japan; CAN: Canada; FR: France; EST: Estonia; UAE: United Arab Emirates; SWE: Sweden; AUS: Australia; GER: Germany; NED: The
Netherlands; SLO: Slovakia; KOR: South Korea; ITA: Italy; CZE: Czech Republic; LAT: Latvia; HUN: Hungary; GRE: Greece; POL: Poland; CHI:
Chile; POR: Portugal; TUR: Turkey; ARG: Argentina; QAT: Qatar; KAZ: Kazakhstan; BUL: Bulgaria; BLR: Belarus; CHN: China; UKR: Ukraine;
BRA: Brazil; MEX: Mexico; COL: Columbia; JOR: Jordan; VEN: Venezuela; BOL: Bolivia; PHI: Philippines; IND: India; GUA: Guatemala; LAO:
Laos; PNG: Papua New Guinea; RWA: Rwanda; GHA: Ghana; NEP: Nepal; PAK: Pakistan; ANG: Angola; KEN: Kenya; NGR: Nigeria; BAN:
Bangladesh; ETH: Ethiopia; and URU: Uruguay.
1 Million
10 Million
100 Million
020406080 100 120 140 160 180
HDI Rank
Number of Satellites | Nonlinear
Exponential (Number of Satellites | Nonlinear)
People per Satellite
IEEE Proof
is also pertinent in the conte xt of EO/RS research and in
the context of geoinformation technolog y applications. In
relation to t he impact of gender on EO/RS research, a recent
study reveals that in the broad “geospatial” profession, only
19% of the practitioners are women [5]. This gender-skewed
landscape can have an impact on the manner in which re-
search is approached and research questions are prioritized
as well as the overall effect that the research has on societ y
at large.
Here, it is important to note that a “masculine” or
“feminine” research agenda is not exclusively about an
emphasis on methods/technical solutions versus em-
phasis on real-world problems but also about 1) how
methods and technical solutions would themselves dif-
fer under the scrutiny of female researchers and 2) how
the problems for which technical (and other) solutions
are sought would differ. For example, in relation to the
second point, research questions that are particularly
important for the empowerment and security of women
“on the ground” may not come to the limelight and be
taken seriously in a “masculine” research agenda that
emphasizes methods and technical solutions that seek
to address more general questions, such as the prediction
of biomass content and slum areas.
In addition to the gender component discussed in the
preceding, this important ethical consideration can, once
again, be explained in the context of data labeling. In
countries of Sout h Asia, for example, for cultural as well
as space-related reasons, joint families and denser com-
munity living are quite common [107]. Errors associated
with t he underestimation of population density and in-
accurate labeling of specific buildings and areas as “too”
densely populated and as being “slums” can be mini-
mized if researchers are aware of the likelihood of such
errors and the local sociocultural and net usage differ-
ences that can cause them. This can be done by ensuring
diversity in the sources of training data, the culture-sen-
sitive labeling of training data (that also considers inputs
from multiple stakeholders), responsible predictions, and
the declaration of any residual inaccuracies that may re-
sult from inadver tent bias.
Such standards are particularly important to prevent bias
resulting, for example, from data labeling. Indeed, ever y
“observer’s” label may differ [108]. At the same time, this
difference can be a source of rich information, and permit-
ting a diversit y of labels from researchers of various back-
grounds can yield more informed “aggregate” and, hope-
fully, less biased data labels. In the context of scales and
ranges of data labels, multistakeholder and multicultural
discussions may be necessar y to identif y culturally appro-
priate and democratically created labels and standards t hat
can avoid bias as well as stigmatization.
In the entire domain of “living” creatures, human beings
are uniquely equipped with the power to take responsibil-
ity. This human ability to take “responsibility” lies at the
heart of all laws, reg ulations, ethical principles, and moral
codes. Indeed, absent human ability to take responsibility
for actions and inactions, no legal or ethical code would
be realistic and implementable. At the interface of law and
ethics, the most impor tant responsibility of any human be-
ing is to protect human rights, fulfill a “duty of care” (e.g.,
under tort law), and remain accountable for all his or her
actions and inactions. The borders of responsibility and ac-
countability (which can also translate into issues of legal
liability), however, become blurr y in t he face of increasing
automation and reliance on AI/ML models. Never theless,
responsibility is listed as an impor tant and central ethical
value in almost all (AI) ethics guidelines. In the following,
we discuss some of the most important ethical issues linked
to responsibility in AI4EO research.
The EU “Ethics Guidelines for Trustwor thy AI” highlight
the impor tance of “human agency and oversight” in A I
systems. The active involvement of human actors (diverse
stakeholders) in every aspect of AI4EO research may not
be possible. However, such involvement is closely linked to
the prev iously discussed issue of the “democratic creation
of standards,” e.g., in the process of labeling. Further, the
quantification of uncertainty, a task undertaken by spe-
cialized AI4EO scientists, can support human agency and
oversight requirements, thereby enhancing tr ust in AI4EO
research. At the same time, it is important to bear in mind
that human agency and oversight are not necessary only
within the confines of the “system” and its nar row research
question. Human agency and oversight are also needed to
understand t he context in which t he designed system is go-
ing to be implemented and introduced and the impact that
such an introduction may have on real people and t heir
lives, livelihoods, and environment.
The human agency and oversight requirement, there-
fore, has the potential to evolve considerably in the AI4EO
research conte xt. Otherwise, while looking at Earth and its
physical elements from afar, intricate human realities, rela-
tionships, and expediencies can be overlooked and ignored,
leading to significant ethically undesirable consequences.
Take, for example, the issue of identif ying refugee ships.
While a researcher may take great pains to ensure system
integrity, honestly declare system limitations, and ensure
that the labeling of pixels is done while taking multiple con-
texts and view points into account, if the ground reality of
how refugee vehicles are treated by a country/government in
question are ignored, the sharing of data/algorithms can lead
to the loss of several lives. Similarly, in the case of research
identif ying slum areas, international pressure in the form of
HDI scores can lead local governments to tear down areas
identif ied as slums, driving thousands to homelessness.
IEEE Proof
Current AI4EO research that relies primarily or exclu-
sively on data in the form of images, texts, and numbers
(e.g., census information) can benefit greatly from also in-
tegrating qualitative real-world information into its analy ti-
cal framework and research questions. A simple approach
could, for example, be conducting a literature review to un-
derstand the cur rent sociocultural and political realities of
a region in the context of A I4EO researc h goals. The “Ethi-
cal Opportunities in Artificial Intelligence for Earth Obser-
vation Research” section discusses some concrete examples
of how such a literature review may enhance the scope and
impact of AI4EO research, on the one hand, while leading
to novel methodological approaches and ethical outcomes,
on the other hand.
EO/RS data and researc h are rapidly moving toward use
cases that have a direct and more concrete impact on gov-
ernmental policies and real human lives. In such a sit uation,
EO/RS research needs to keep a close eye on the impact it
can have on human rights. The sphere within whic h AI4EO
researchers need to exercise “care” is therefore expanding.
This is e xplained in greater detail in t he “Ethical Opportu-
nities in Ar tificial Intelligence for Earth Observation Re-
search” section, with the help of four examples from AI4EO
Sustainability is arguably not one of the traditional con-
cerns of (business) ethics [109]. However, with the expan-
sion of the concept of “corporate social responsibilit y,” the
growing emphasis on the “triple bottom line” (economic,
social, and environmental sustainability), and the empha-
sis placed on accomplishing the UN SDGs, sustainabil-
ity is now inseparable from ethics [110]. A significant and
growing body of work in AI4EO purports to support the
monitoring and, therefore, the (eventual) accomplishment
of various UN SDGs. “Sustainability,” therefore, is a major
ethical opportunity for the AI4EO (research) communit y.
AI4EO research often aims to monitor various environ-
mental conditions, including climate change [111], [112],
sea levels [112], air and water pollution [113], water qual-
ity [114], shif ts in glaciers and polar ice [115], [116], forests/
green cover [117], biodiversity [118], [119], drought predic-
tion [120], and agricultural/crop type classification [121].
These diverse uses are also closely linked to the subject mat-
ter covered by various UN SDGs, par ticularly, SDGs 6, 13,
14, and 15. Under SDG 15, Landsat and Sentinel-2 data have
been used to predict SDG indicators under SDG 15.1.1: For-
est Area as Proportion of Total Land Area, SDG 15.3.1: Pro-
portion of Land That Is Degraded Over Total Land Area,
and SDG 2.4.1: Proportion of Agricultural A rea Under Pro-
ductive and Sustainable Agriculture [122]. Together with
other datasets, such as Open Street Maps [122], scientists
predict that it would be possible to generate training datas-
ets for use at global scales.
In the Tanzanian Morogoro urban municipality, RS
data, social media data, and population data have been
used to determine urban sprawl from 2011 to 2017. The
study demonstrated the usefulness of such an approach
to understanding how the city’s expansion can impact the
whole ecosystem. It has helped to interpret how pristine
grasslands and forests, whic h provide essential ecosystem
services, such as carbon sequestration and support for
biodiversity, have been replaced by built-up land cover.
Combined with Twitter data, this project has provided
useful insights for urban planning in the region, where
this kind of data was either unavailable or of insufficient
qualit y [123]. Other work has also demonstrated that EO,
RS, and ML can contribute to sustainable development by
monitoring land degradation as well as urban expansion,
with its impact on the env ironment [124], [125], [126],
[127], [128]. At the same time, it is essential that AI and
ML models relied upon by the community do not result in
compromising and sacrificing any one or more prongs of
sustainability at the cost of another. For example, a focus
merely on environmental sustainability must not ignore
and compromise social and cultural diversity [129] and
the working/living space of local and indigenous people.
The pursuit of “sustainability” can therefore lead to sev-
eral ethical issues as one attempts to balance out (appar-
ently) divergent goals [130].
In the broader contex t of sustainability in relation to the
impact of researc h endeavors on the environment, it is im-
portant that AI4EO research itself (intrinsically) works to-
ward becoming energy efficient. For example, as accurate,
verifiable, and representative (diverse) data are not only
difficult to come by but also costly and energ y intensive
to maintain, the ethical ideal of sustainability also calls
for the fundamental research community to work toward
building algorit hms t hat support “data minimization.”
Data minimization applies not only in the GDPR sense of
the minimum accumulation and storage of personal data
but also in terms of increasing the efficiency of algorithms
so that they need fewer data points, emphasizing quality
and diversity over sheer volume.
In all f ields of A I4EO research, opportunities to do good
abound. Even a casual glace through published work pro-
vides an overview of the wide array of ethical opportuni-
ties presented by technological developments in this f ield.
For example, recent publications suggest that AI4EO re-
search can help save lives during natural disasters [131];
track endangered species [132]; monitor and control the
spread of harmful algae, such as cyanobacteria, in water
bodies [133]; track disease progression [134]; support pay-
ments for ecosystem ser vices (PES) [135], [136]; monitor
climate change [137], sea levels [133], [138], refugee sett le-
ments [51], and the process of urbanization [139]; and even
detect migrant ships and vessels to assist humanitarian aid
IEEE Proof
workers and prevent violations of human rights and inter-
national laws [140].
Yet, as discussed in the previous section, while pursu-
ing these noble intentions, there is a likelihood that ethical
risks and concerns can arise [8], justifying the exercise un-
dertaken in the previous section. However, in this ar ticle,
we aim to step beyond identifying ethical concerns, issues,
and dilemmas in AI4EO research. We do so by providing a
novel understanding of the concept of “ethical opportuni-
ties,” with the help of examples related to four dominant
fields of ongoing AI4EO research. With these examples, we
explain how a simple ground realities-informed restructur-
ing of research questions and research approaches can help
identif y et hical opportunities and find win-win outcomes
in early stages of AI4EO resea rch, without getting entangled
in a maze of ethical issues and dilemmas.
With the examples in the following, we also see how re-
structuring research questions can help accomplish various
UN SDGs, particularly, SDG 1: No Poverty, SDG 3: Good
Health and Well-Being, SDG 11: Sustainable Cities and
Communities, and SDG 15: Life on Land. Finally, as the
AI4EO community faces an uncertain future, where not all
ethical issues and concerns are immediately clear, we con-
sider this practical approach to grabbing ethical opportuni-
ties a means of minimizing ethical concerns.
It is estimated that, roughly, there are either 860 million
[141] or 1.5 billion [142] slum dwellers worldwide. These
vastly different estimates show that we do not really know
the real numbers. AI4EO research attempts to reduce e x-
isting knowledge gaps by remotely identifying so-called
slum areas and acc urately labeling them, for example, to
detect extents, morphologies, and changes over time. This,
in tur n, can help monitor various UN SDGs, such as SDG
1: No Poverty and SDG 11: Sustainable Cities and Com-
However, the term slum and related labels, as discussed
previously, have a clear negative connotation. People liv-
ing on peripheries of areas labeled as “slums” (and those
living within such areas) may object to the area being so
labeled. Indeed, there are vast discrepancies in the defini-
tion of what constitutes a “slum” in various regions of the
world [143]. It has been shown that living areas of the ur-
ban poor take on ver y different morphological forms across
the globe [134]. One manifestation of povert y is so-called
(morphological) slums, whose physical manifestations are
described with very high building densities, small and low
buildings, and complex organic structures. It is understand-
able that this proxy infor mation of building structures for
the localization of “slums” is legitimate, especially in many
areas in the Global South. It has even been shown that in
morphological slums, a large portion of the population is
among the urban poor [145], [176] .
As legitimate as the approach may sound in data-poor
regions, it also involves several ethical risks:
1) The current approach essentially means that the prox y
of morphological slum structures is also used to dub
people living in such structures as slum dwellers, who
may not be poor. In such areas, it is not uncommon
that slum dwellers ear n higher incomes over time and
improve building structures and amenities but still con-
tinue to live in the same “slum.”
2) Final map products can stigmatize these areas, and resi-
dents from these regions may have fewer opportunities,
e.g., in jobs, and they may be treated disrespectfully by
peers (e.g., other children in schools).
3) What is legitimately proved about specific build-
ing str uctures may be perceived much more positively
by residents than by researchers located in countries of
the Global North.
4) Therefore, AI4EO research may generate labels
that correspond to building structures but do not do suf-
ficient justice to the complexity of the topic of poverty
and ground-level realities.
5) Labels are generated by people who often know
about slums and slum structures only in theor y and
have neither lived in nor ever been to one. A bias in la-
beled data is the consequence, even with a consistent
conceptual approach.
6) Algorithms are able to reliably f ind only ver y
specific housing forms/structures/sizes in image data.
Slums, povert y, and housing structures of ten have local
specifics. AI/ML algorithms and models may therefore
be inadequate and inappropriate to make global and
large-scale predictions
7) There are also vir tually no datasets that clearly
spatially locate the target domain (in this case, “slums”)
so that validation is often accompanied by ver y subjec-
tive criteria.
In the case of slum identif ication using RS, we are there-
fore faced with the dilemma of either comprehensively
identif ying all congested living areas as slums to monitor
urbanization and its impacts on safety and standards of
city living over time or not identifying (all) congested living
areas as “slums” to avoid possible bias and stigmatization
resulting from culturally inappropriate labeling of specific
areas as “slums.”
To avoid or minimize these ethical issues, more con-
structive labeling can be resorted to, based on the purpose
with which the study is conducted; e.g., does the study aim
to identify areas that are more prone to fire due to popu-
lation density? Or does the study aim to identif y regions
where greater government investment is needed, say, for
upgrading sanitation facilities, toilets, and electricity cover-
age? Perhaps the aim of the study is to support the expan-
sion of green areas in economically poorer regions? More
accurate use of labels can lead to affirmative action without
stigmatizing populations. To do so, there is a need to re-
think research questions as well as research approaches. It
IEEE Proof
may also be necessar y to invest funds in acquiring ground-
level information and understanding of socioeconomic
and cultural realities.
Let us take t he example of toilets (especially public
toilets) in slum areas and in rural regions. Cur rently, in
EO/RS research, the level of development of an area is
often determined by the intensity of night light emissions
[146] and built-up str uctures [61]. Yet, in the poorest re-
gions and slums, the quality of living conditions can, for
example, be better determined based on the availabil-
ity and quality of public toilets. Slum living often may
not permit personal toilets within homes, and in such
circumstances, the availability of clean toilets, with ad-
equate distance between women’s and men’s toilets, can
be a strong indicator not only of the standard of living but
also of safety, privacy, and hygiene for women. EO proxy
information may not reflect the availabilit y (or not) of
toilets. Here, therefore, it is necessar y to collect additional
ground truth data on the availability (or not) of toilets in
specific communities so as to give more concrete direc-
tion to policy makers when faced with international pres-
sure to improve living conditions, including the safet y of
women, in “slum” areas and in economically marginal-
ized areas. This would also help bring a gender-sensitive
dimension to AI4EO research that is currently lacking
optimal attention.
Further, slum identification research can also (better)
help address real-world problems by focusing on the situa-
tion inside and outside slum areas, for example, determin-
ing whether major landfills and garbage disposal areas are
located near slums, increasing t he chances of spreading
communicable diseases and public health crises. In this
context, the story of the father of environmental justice,
Prof. Robert Bullard, is both interesting and relevant. To-
gether with his wife, Prof. Bullard sued the state of Texas,
the cit y of Houston, and Harris County, USA, when a pri-
vate company attempted to locate a landfill in the middle
of a predominantly Black middle-class suburban neigh-
borhood in northeast Houston. They conducted a study
and created a map identif ying the location of all landfills,
incinerators, and garbage dumps in Houston. The map
showed that “five out of five of the city-owned landfills,
six out of the eight city-owned incinerators and three out of
the four privately owned landfills in Houston were located
in predominantly Black neighborhoods, from the 1930s
up to 1978. Even though Blacks only made up 25% of the
population, they were getting 82% of the garbage dumped
on them” [147].
The case, Margaret Bean et al. v. Southwestern Waste Man-
agement Corp. et al., was the first in the United States to
identif y the problem of environmental racism. Similarly,
AI4EO research can help determine whether factories that
pollute air and ground water are located near slums. Such
information can, once again, uncover inequality and help
improve the quality of life of people living in congested ar-
eas, without stigmatizing them and potentially depriving
them of living space in case of extreme interventions by lo-
cal governments.
Closely linked to the issue of slum mapping is the issue of
rapid urbanization and r ural–urban migration. By 2050,
around three-quarters of the world’s population is expected
to live in cities [148]. Since a nearly perfect correlation is
verified between urbanization and the economic prosper-
ity of societies [149], this development may raise positive
expectations. However, current trends in global rural–ur-
ban migration pose fundamental challenges, altering the
physical dimensions and config urations of cities at all
scales. UN population figures, on which our current under-
standing of global urbanization trends is primarily based,
do not provide information on the distribution, pattern,
and evolution of the built environment [150]. With t he
growing rate of informal settlements, ref ugee camps, urban
corridors, and so on, city profiles are c hanging rapidly, with
local governments and city planners unable to keep track
of the changes and their short- and long-term implications.
The immense spatial knowledge gaps we face and will
continue to face will impact our societies and way of life
by changing risk levels associated with natural hazards and
infectious diseases and raising novel governance issues re-
lated to infrastructural planning and distr ibutive justice.
Despite ongoing efforts, global urban mapping approaches
are unable to deliver the necessary spatial data and geoin-
formation in the geometric, thematic, and temporal resolu-
tions needed to address these issues and adequately prepare
for the future.
Data provided by satellite-based RS, especially when
fus ed with data f rom social media platfor ms, such as Twitter
and Flickr, and with census data, are being explored to as-
sist in mapping global urbanization trends. Related to these
effor ts are research projects aimed at extracting individual
building footprints, building functions, and building-level
populations. As previously discussed in the “Categorizing
and Understanding Artificial Intelligence and Data Ethics
for Artif icial Intelligence for Ear th Obser vation Research”
section, using AI4EO for urban mapping (including the ex-
traction of building footprints and building-level popula-
tion densities) can raise several ethical concerns:
1) Data bias: Compared to other geospatial applications,
urban mapping suffers from a strong data bias: the train-
ing data used to train ML models, offered, for example,
by Open Street Map, are rich in regions of the Global
North, such as Europe and North America, where we al-
ready have enough geospatial information, and poor in
the Global South, where knowledge and geospatial in-
formation gaps need to be filled to address, for example,
UN SDG 11: Sustainable Cities and Communities. As a
result of this data bias, 1) the quality measures derived
IEEE Proof
from ML models may have representative metr ics only
for data-rich areas and 2) more severely, urban variables
predicted in the Global South may be poor in quality,
as ML models may suffer from a distributional shift—a
change in the data distribution between an algorithm’s
training dataset and the dataset when deployed.
2) Explainability and trustworthiness of the results: There is no
doubt that with the help of cut ting-edge A I4EO research,
promising results in global urban mapping have been
achieved. Yet, to effectively incor porate these results
into decision making, for example, the UN SDGs and
EU’s Green Deal, in addition to ensur ing the high accu-
racy, explainability, and traceabilit y and reproducibility
of results, it is indispensable to quantif y the uncer tainty
of these predictions from an AI algorithm. Despite in-
creasing efforts, we are far from identif ying appropriate
technical solutions.
3) Data protection: A I4EO res earch make s use of various t ypes
of data, such as satellite images and social media data.
However, improved image resolution and, in par ticular,
the unprecedented geospatial information derived from
those data raise questions under the European GDPR.
4) Security concerns: Urban variables are produced by A I4EO
research for the public good, to support further research,
and to aid constructive polic y making. Since urban en-
vironments accommodate more than half of the world’s
population on only a small fraction of Earth ’s land sur-
face, it is particularly important to draw our attention
to possible security concerns resulting from making in-
formation publicly available, for example, in the context
of conflicts. On the bright side, AI applied to satellite
imager y holds promise for the automated detection of
war-related building destruction [151]. Yet, we cannot
ignore the fact that AI4EO also affords opportunities to
identif y densely built/populated areas and critical in-
frastructures, such as power plants, hospitals, large in-
dustrial facilities, and so on, making them easy targets
for terrorist attacks and large-scale targeted destruction
by enemy forces during war. How such issues can be re-
solved remains an open question.
Efforts are ongoing to improve technical solutions to the
previously mentioned issues of e xplainability and privacy
(e.g., through privacy-preser ving ML). However, the issue
of data bias, which is also linked to a lack of availability of
diverse datasets, requires fresh thinking. Transparent and
trustworthy privacy-preser ving mechanisms can facilitate le-
gal and legitimate data shar ing by individuals (data genera-
tors) as well as by entities that host and “control” large datas-
ets/big data (such as Twitter); this, again, can improve access
to diverse datasets. To overcome issues of data scarcity and
diversity, it may be useful, once again, to rethink research
questions. Instead of merely securing privacy after collecting
data, for e xample, by hiding geolocations and the identities
of people contributing data, can systems be designed to per-
mit users to voluntarily contribute data and to incentivize
data contribution on a regular basis (in specific formats)?
Evolving research suggests that by using A I/ML models
in combination with blockchain and distributed ledger
technology (DLT), we can incentivize and encourage the
sharing of relevant georeferenced data by individuals and
corporate entities. In addition to incentivizing data shar-
ing, these technologies permit data contributors to retain
control over their data and ensure the traceability (and
transparency) of downstream data access. Such solutions
can empower data contributors (including individual con-
tributors) while also furthering the EU’s goal of fostering
the development of a “data economy” through data porta-
bility. Since these technologies can be made to maintain a
record of data access, security concer ns can also be better
Technologies such as blockchain could therefore also
be researched and deployed, including from an ethical per-
spective, for fundamental AI4EO research. In this context,
the European Space Agency has a declared research focus
on the use of EO and blockchain for tracking data flows
and adopting privacy-preser ving methods for ML [152].
Since traceability lies at the heart of blockchain/DLT-based
solutions, dedicated research in the field can also support
the selective disclosure of data (e.g., data on building func-
tions, building-level populations, and so on) to specific
people while maintaining an immutable record of who ac-
cessed the data. Such solutions can then also minimize the
security concer ns listed previously, barring unidentified us-
ers from accessing information for illegal purposes. There-
fore, we see that once we turn a research question around to
specifically accomplish ethical goals and overcome ethical
issues, the research approaches, including possible techni-
cal solutions to be considered, can also ex pand, leading to a
win-win solution for all concerned.
The UN recognizes the close relationship between land and
peoples’ livelihoods. This relationship is particularly obvi-
ous in the case of smallholder farming, herding, and ASM.
In several developing countries, notably in West Africa
countries, such as Ghana, research also reveals the close
connection between smallholder farming and ASM. Glob-
ally, an estimated 40 million people are employed direct ly
in ASM ac tivities [153], either because A SM was the way
their ancestors traditionally made a living or because lim-
ited and unviable income from small-scale farming [154]
forces them into ASM for an additional source of income. In
cur rent times, commu nities displaced by large-sc ale mining
(LSM) activities are also seen to resort to ASM as a means
of reducing their poverty [155]. ASM is also conducted by
government-authorized small and medium enterprises.
The African Commission on Human and People’s Lives
has held that “it doesn’t matter whether they (nonindig-
enous people) had legal titles to the land, t he fact that the
IEEE Proof
victims cannot derive t heir livelihood from what they pos-
sessed for generations means they have been deprived of
the use of their property under conditions which are not
permitted under Article 14.” In this decided court case, the
African Commission considered the traditional use of land
as a more important consideration to bear in mind than
the status of being “indigenous” or not [156]. Following the
rationale of this decided case, communities that have tradi-
tionally used land for ASM have the right to continue using
land for this purpose unless alternative sources of liveli-
hoods that also secure their traditional and cultural ways of
life are offered and accepted. For the remaining categories
of ASM activities, the right to a livelihood is significantly, if
not equally, implicated if the right to conduct and continue
ASM is withdrawn by government reg ulations and bans
without offering meaningful alternatives.
On the other hand, the uncontrolled continuation of
ASM can lead to the violation of other rights and obliga-
tions. For e xample, illegal ASM leads to deforestation, the
contamination of water bodies [157], and land degradation,
leaving land unfit for ag riculture [155]. ASM also nega-
tively affects the health of artisans, either due to the use of
mercury to extract gold from mines [158] or to the lack of
equipment to secure their health and safety [29]. Curre nt ly,
although ASM supports subsistence, it does not afford op-
port unities for the socioeconomic grow th and development
of artisans. Research reveals that ASM, while contributing
significantly to the national economy of various countries,
keeps artisans caught in a “poverty trap” [159].
As a result of these realities, there is considerable pres-
sure on governments to formalize and legalize ASM ac-
tivities and put adequate regulations in place to minimize
the environmental and health-related harm resulting
from (illegal) ASM. Further, there is an urgent need for
governments to create alternative economic and develop-
ment opportunities for artisans currently caught in a pov-
ert y trap, with ASM being their only source of subsistence.
This is also an urgent mandate closely linked to UN SDG
1: Zero Pover ty. Undoubtedly, these issues are not unique
to ASM but are also obser ved in the context of L SM activi-
ties [155], [157]. Nevert heless, the rapid expansion of the
area under ASM (including during periods in which na-
tional governments imposed a ban on ASM [160]) and the
long-term socioeconomic and environmental concerns
linked to ASM (see Table 1) indicate the urgent need to
find practical solutions.
ASM studies that use data and images from the Coper-
nicus/Sentinel-2 mission primarily focus on change detec-
tion over time, identifying regions of high and low mine
density [161] and seeking to improve t he accuracy of images
by using ML algorithms that reduce the problems caused by
cloud cover and other irregularities in satellite-transmitted
images [160 ], [162]. These studies have been able to moni-
tor the increase in area under ASM with increasing accuracy
[160] and are aimed at supporting regulatory endeavors that
seek to check the unauthorized expansion of illegal ASM.
Yet, these studies have themselves highlighted the inef-
ficacy of governmental bans and the rapid expansion of
areas affected by ASM, despite improved remote monitor-
ing and the increasing number of reg ulatory efforts. This
suggests that a major shift is needed in the overall politi-
cal and regulator y approach to the problem of illegal ASM
activities. While RS/EO data have helped track only the
expanding scope of the problem, changes in the overall
approach taken by AI4EO research can support and opti-
mally direct governmental as well as R&D efforts in the
direction of 1) reviving areas that were previously mined
and 2) training and employing artisans in alternative, le-
gal, and economically viable activities associated with the
revival of ASM sites.
RS/EO and AI4EO research can support the creation of
such alternative opportunities if the cur rent approach to
tracking and monitoring illegal mining activities is modi-
fied. First, as mentioned in the context of human agenc y
and oversight, A I4EO research linked to the monitoring of
ASM activities needs to take into account socioeconomic
and political “ground realities.” A literature review can,
for example, provide significant insights (such as the ones
It supports livelihoods of four categories of pe ople who
either fully or subst antially rely on it for subsistence.
It leads to significant environmental degradation, including deforestation,
contamination of water bodies, and making land unfit for agriculture.
It supports the national economy of countries where it is
widely practiced.
It has a negative impact on the health of ar tis ans.
It can be regulated to minimize environmental and
health-related harm.
It keeps artisans caught in a “pover ty trap.”
Implementing regulations seeking to legalize ASM under conditions that prevent
irreversible environmental damage and mandate revival of mining sites, is challenging.
Follow ing the mandates of such regulations is also expensive and may be out of reach
for artisans and small companies.
“Comple x issues of land tenure, social stability, mining regulation and taxation, and
environment al degradation undermine the viability and sustainabilit y of ASM as a
livelihood strategy ” [159].
IEEE Proof
discussed in this section) and support the reframing of re-
searc h questions and approac hes vis-à-vis the identif ication
and monitoring of ASM and its e xpansion.
More specifically, a multidisciplinary and sociopoliti-
cal literature review suggests that there is a need to migrate
away from the current focus on identif ying illegal sites and
changes in the extent of illegal mining over time. An al-
ternative approach could, for example, be based on using
AI4EO to predict the depth of various mining sites. This
information may then be used by governmental entities to
calculate the comparative ef ficiency of various mining sites,
on the one hand, and work with environmentalists to find
quick means of recuperating mining sites that are not too
deep, on the other hand.
Artisanal and small-scale miners are usually able to dig
only shallow mines, the environmental impacts of which
are, potentially, easier to reverse if acted on immediately
after the mining is completed. Indeed, artisans can them-
selves be trained to mine in ways that minimize environ-
mental damage and to take steps to recoup mined areas
before moving on to the ne xt site. Artisanal miners who
comply with such guidelines can then be given licenses
to mine rather than having their source of livelihood re-
moved. Legalized/licensed A SM sites can then also be con-
verted to tourist attractions, further supporting economic
To support this effort, the RS/EO community can also
fuse diverse sources of data toward the following long- and
short-term ends:
1) In the short term: Identify regions t hat are most likely
to suffer long-lasting and significant environmental
damage due to ASM and LSM. Such regions should be
marked as “highly sensitive” so as to urge local govern-
ments to immediately and effectively secure these areas
to prevent any and all ASM/LSM activities. Such regions
can, for e xample, be regions that are close to major water
bodies, wells, and agr icultural lands that are significant
sources of food.
2) In the short term: Together with ground truth data and
spatial depth analysis of EO/RS data on ASM and L SM
activities, recommendations can be made to local gov-
ernments and international suppor t agencies to invest,
with priority, in recapturing/reviving lands that are
least affected by mining. For example, as discussed,
since ar tisanal mining sites are most likely to be shal-
lower and narrower, it may be easier to revive lands
impacted by ASM than by LSM, which uses wider and
deeper excavations. Methods to identify the depth and
environmental impact of various mining sites may need
to be developed with environmental scientists who are
familiar with means of calculating land damage due to
ASM and the most cost-effective ways of reviving such
3) In the long term: A ffir mative action initiatives, such as
education, as well as concrete monetary incentives need
to be given by local governments (supported by interna-
tional f unding agencies) to ensure that artisans have the
skills, tools, and economic incentives to suppor t efforts
aimed at reviving lands affected by ASM. The environ-
mental impacts of such initiatives and incentives can
then be cost-effectively monitored by EO/RS scientists
to identify the most successful ones, thereby creating a
repositor y of best practices for the revival of lands af-
fected by ASM as well as means of gainf ully employing
artisans in jobs t hat are both sustainable and prof itable
in the long run.
One of the areas of ongoing AI4EO work involves monitor-
ing and assessing forest cover, forest density, biodiversity,
and changes therein over time. Such efforts can alert gov-
ernments to the areas where urgent governmental inter ven-
tions are necessar y to revive green cover and protect bio-
diversity. Ethical concerns and risks, however, can emerge
in this sphere of research. In the German state of Bavaria,
for example, there has been a long-standing political debate
about establishing a third national park under the Bavar-
ian nature conservation policy. The supporters of the idea
argue that such a large-scale protected area would “give the
diverse animal and plant world a habitat where it can devel-
op and adapt to the climate challenges” [163]. While recent
polls suggest that the vast majority of the German popula-
tion supports the establishment of such a park, there is a
minorit y voice that has historically disagreed. The reasons,
though not widely reported, include the understanding
that ideas to establish new parks are merely “an ecological
relic from the last millennium.” More important from an
ethical perspective was the opinion of forest owners that
their livelihoods would be threatened.
Similar to the ethical dilemma t hrown up by the moni-
toring and identification of illegal mining sites, therefore,
efforts to enrich and revive biodiversity by establishing
national parks and, more generally, restricting human ac-
tivity can result in major human rights violations of mi-
nority groups that historically reside in and derive their
livelihoods from such areas. The voices of these minorities
can also get crowded out (or outnumbered) in large-scale
surveys that seek to identify the “dominant,” or popular,
voice. In such a scenario, once again, the question to be ad-
dressed is not (only) how to mitigate the ethical risks but,
rather, how to maximize the ethical opportunity, with the
aim of finding a win-win outcome. In this case study, we
describe two approaches that are emerging from ongoing
Various UN SDGs, such as SDG 13: Climate Action and
SDG 15: Life on Land, env isage the establishment and des-
ignation of certain areas as nature parks and wilderness so
as to provide a sanctuary to wild plants and animals (i.e.,
IEEE Proof
biodiversity). Such a designation thus essentially entails
limiting as well as eliminating human activities within
such areas. In his excellent book on wilderness protection
in Europe, Kees Bastmeijer [164] discusses the genesis of
the ter m wilderness as well as the various definitions that
exist. While there is no clear mathematical delineation of
what a wilderness area is, it is reasonable to use an official
governmental definition, such as the one held by the Eu-
ropean Commission, which itself builds upon the defini-
tion of protected areas used by the International Union for
Conservation of Nature [165]: “A wilderness is an area gov-
erned by natural processes. It is composed of native habi-
tats and species, and large enough for the effective ecologi-
cal functioning of natural processes. It is unmodified or
only slightly modified and without intrusive or extractive
human activ ity, settlements, infrastructure or visual dis-
turba n ce.”
This definition assigns four central ecological aspects to
the concept of wilder ness:
1) naturalness
2) undisturbedness
3) undevelopedness
4) scale.
While scale can easily be defined in a mathematical sense,
e.g., by def ining so-called minimum mapping units, natu-
ralness refers to ecosystems functioning in a natural way
and can thus hardly be measured in a technical sense. It is,
in fact, subject to judgment by ecolog y experts based on in
situ observations. However, undisturbedness and undevel-
opedness are concepts that can potentially be well observed
from afar, namely, by RS technologies. As per the European
Commission, undevelopedness refers to the absence of
“habitation, settlements or other human artifacts such as
power lines, roads, railways, fences [that] may hinder eco-
logical processes directly or by promoting likelihood of hu-
man interference” [165].
Thus far, Earth ’s wilderness extent has largely been
analyzed by (semi)automatic approaches that rely on the
rule-based fusion of existing geospatial data sources. Most
of those approaches go back to the work of Sanderson et
al. [166], who proposed to measure the “human footprint”
by what they called the Human Influence Index (HII) to
map how much human civilization has influenced t he land
surface. To calculate the HII, they relied on four types of
geospatial input data:
1) population density, expressed by population density
2) land transformation, expressed by land use maps
3) accessibility, expressed by the proximity to traffic ways
4) electrical power infrastructure, expressed by nighttime
By rescaling all the globally available input data to a
resolution of 1 × 1 km and scaling each input source’s con-
tribution with a score between zero (for the absence of hu-
man influence) and 10 (for high human influence), they
created a map of the global human footprint. Some years
later, Venter et al. [167] used this approach to investigate
the change of human influence between 1993 and 2009.
In their study, they found that dur ing this period, the hu-
man footprint had increased by 9%, with 75% of Earth ’s
land surface e xperiencing measurable human pressure. In
a similar manner, Allan et al. [168] published temporally
comparable maps of terrestrial wildernesses. Since they de-
fined wilderness areas as pressure-free lands with a contig-
uous area of more than 10,000 km2, they found Europe to
be void of any tr uly wild areas. Ekim et al. [169] presented
the latest adaptation of the Sanderson paradigm to enable
a finer-grained calculation of the degree of naturalness at
a spatial resolution of 10 m per pixel in the resulting map
dataset. They claimed that such maps can support decision
making with respect to government-dr iven conser vation
effor ts on a regional scale.
AI4EO has potential to replace the existing rule-based
data fusion approaches by deep learning-based approach-
es to map wilderness from space, i.e., by using RS satel-
lite images. Given recent trends in explainable AI, t his can
be extended by adding interpretability and explainability
to the deep learning model to gain a better understand-
ing of what makes wildernesses wild. By answering the
question “Why would an AI algorithm map certain areas
on Earth as wilderness and others not?” A I4EO will not
only advance the automation of RS but also deepen our
understanding of what makes nature “wild” and worthy
of protection.
However, while there is obviously great opportunity in
AI4EO, there are also certain risks involved. The first risk
lies, of course, in the technical implementation of deep
learning-based AI4EO approaches: supervised deep neu-
ral networks are always as good only as the training data
they are fed with. If an approach designed to automatically
map wilderness areas from satellite imager y is provided
only with examples of currently protected landscapes, it
will probably be biased toward biomes that can mainly be
found in the Western world, e.g., the United States and Eu-
rope, as those regions have a larger share of designated con-
servation areas t han other par ts in the world (see [175]). It is
thus of crucial importance that extra care is taken to design
a model (and its training data) in a way t hat mitigates t he
bias risk and allows the identification of natural and wild
areas across the whole globe.
At a more human level, as mentioned, there is a ma-
jor ethical risk that by making the determination of po-
tential conservation areas highly automatic on a purely
mathematical basis, there is a chance that local stakehold-
ers will be overlooked. Especially in densely populated
countries, banning land exploitation by business, traffic,
agriculture, and forestry for the sake of environmental
protection does not go without opposition from region-
al communities that do not want to lose their land use
rights. Often, the argument is that there is no real nature
left anyway and that prohibiting development is an unjus-
tified expropr iation.
IEEE Proof
For several decades now, academic research has been pre-
occupied with t he issue of creating appropriate incentives
for “correct” behavior. While Hardin’s “Tragedy of the Com-
mons” recommended creating propert y rights over “com-
mons” to prevent their ot her wise cer tain ruin [33], Ronald
Coase recommended a (legal) distribution of these prop-
ert y rights in a manner that would ensure that people use
their free bargaining power to reach the most “efficient”
conclusion [171]. Order ethics, a subdiscipline of ethics,
has also evolved to emphasize that moralizing alone is not
adequate. R ather, institutional structures and incentive
mechanisms need to designed in such a way t hat ethical
(including sustainable) behavior becomes the preferred
mode of conduct [172].
There is also a stream of social science and economics re-
search that is looking into the merits of PES [173]. Building
on this line of thought, recent work recommends looking
at blockchain/DLT to incentivize and reward activities that
promote (agro)biodiversity conser vation [174]. Academic re-
searc h has also demonstrated the possibility of incent ivizing
biodiversity conser vation by using a combination of RS and
blockchain-based automated PES in Namibia [144]. Here,
the habitat integrity of an elephant corridor was assessed by
RS algorithms, and a system was demonstrated that could
trigger f ictitious blockchain smart contract payments to
surrounding communities. The surrounding communities
could therefore potentially be incentivized to contribute to
biodiversity management and maintenance, without being
dislocated from their ancestral homes and even supporting
their livelihoods through PES mechanisms.
These research trends are promising in that they point to
the pos