IEEE Geoscience and Remote Sensing Magazine - December 2022 - 170

for example, with some segments of the population turning
against governmental safety regulations, protests from
people may be visible in diverse forms on social media
platforms. Recent governmental bans of weekend protests
against COVID-19 measures [201] (which were later overturned
by a court ruling [202]) and alleged inaccuracies in
reporting of the number of participants in previous protests
[201] also highlight the need to use geoinformation from
social media data carefully.
BIAS IN TRAINING DATA
Published work has also highlighted the ethical issues that
can or do arise from using data from any source (including
from social media platforms) to train (inter alia) ML
models [203], [204]. Several recent incidents have highlighted
the (negative) unintended impact that biased training
data can have on ML-based tools. Amazon's AI-based
recruitment tool that showed a bias against women is one
example [203]. More recently, the PULSE algorithm, which
uses a process called upscaling to convert low-resolution
pixelated images to high-resolution ones, was also found
to be biased. When used to enhance facial images of real
human beings, the resulting faces were " distinctly white, "
even when the input low-resolution images were of Barack
Obama or Lucy Liu [205].
Right from the start of any research endeavor, bias can
result from selection (inclusion or exclusion) of specific
data sources. For example, in different countries, different
social media platforms are popular, and the selection of
any one (at the exclusion of others) can lead to bias (e.g.,
Flickr in the United States, VKontakte in Russia, and Weibo
in China [206], [207]).
It is therefore necessary for researchers to be aware of
and disclose the reasons for selecting specific data sources
over others. Often, the main reason may be the access rights
given by some platform providers based on their existing
end-user licenses. Not all social media platforms grant such
access. This convenience-based platform selection may itself
lead to biases. In Flickr, for example, the type of details
(e.g., camera specifications) that can be entered by users
suggests that it is used only by a very specialized, artistic
audience (see the " Geoinformation Extraction From Social
Media Images " section). Accordingly, if determinations of
building or region aesthetics are based on Flickr data, there
is a high likelihood of bias due to individual " artistic " depictions
of specific regions or locations. Similarly, in Twitter,
a significant percentage of the text and data shared may
specifically relate to places Twitter users are proud to be associated
with or found at, e.g., famous buildings or popular
tourist destinations.
In 2020, an estimated 3.6 billion people used various
social media platforms [208]. Yet, in the context of geoinformation
mining from social media data, where one
of the aims is to train models to make predictions for
regions from where data may not be available, researchers
are well aware of the digital divide [209], [210], [211].
170
In developing countries, residents of rural and slum areas
may increasingly use social media platforms in the
future (efforts to create social media platforms for use
in local languages are underway). However, only a small
minority of the global population uses " popular " social
media platforms [209]. The elderly and the disabled may
also be severely underrepresented in datasets taken from
social media platforms. The platforms that, by their nature,
permit or encourage public sharing of information
or images (e.g., Twitter), are used by an even more limited
number and " category " of people. When data harvested
from such sources are used to train ML models to,
for example, identify " slum areas " or estimate buildinglevel
population density, several precautions must be exercised,
and several limitations of the dataset accurately
disclosed so as to alert users to potential gaps or inaccuracies
in the predictions.
Let us say, for example, that social media text data (see
the " Geoinformation Extraction From Social Media Text
Messages " section) from Twitter users are used to train ML
models, and these models are used to predict building-level
population density in various regions of the world. Twitter
demographics reveals that 63% of Twitter users are over 35
years of age [211]. Will models trained on Twitter data then
be able to accurately predict building-level population density
in countries like India, where an estimated 65% of the
population is under 35 years of age [212]?
Following (platform) selection bias, bias can also creep
into training data at the time of labeling. Known or unknown
personal concepts, prejudices, and cultural beliefs
of the " observer, " i.e., the one placing the labels, needs to
be identified and comprehensively declared at the outset
[209], [213].
TRANSPARENCY AND EXPLAINABILITY
Other issues that are frequently flagged, particularly in the
context of AI/ML applications that utilize big data, relate
to ensuring fairness, (decisional) transparency [204], and
explainability [214]. Given issues such as stigmatization
(discussed previously) and security and sovereignty (discussed
next) that may arise from labeling of data, it is important
that AI algorithms are able to trace back their steps
from the end result they deliver to the specific steps taken
to reach that result. The extent to which this is necessary
and possible in the context of various EO/RS applications
utilizing geoinformation data from social media needs to
be examined.
SECURITY (INCLUDING IN THE CONTEXT
OF DATA STORAGE)
It is necessary to distinguish between the immediate use to
which the harvested social media data are put, and other
uses to which the data, once stored, may be put in the future.
To the extent that the data collected from social media
are used, for example, for semantic labeling of buildings,
or labels that link to the beauty and qualities of a specific
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2022
IEEE Geoscience and Remote Sensing Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2022
