IEEE Geoscience and Remote Sensing Magazine - December 2022 - 102

adoption of agroecology (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
policy makers are actively seeking means of optimizing
farmer migration to agroecological approaches for sustainable
farming.
(NON)BIAS, INCLUDING NONBIAS IN TRAINING DATA
Bias is an ethical issue that has frequently come into the
limelight in relation to AI applications ranging from Amazon's
AI-based recruitment tool that showed bias against
women [96] to the PULSE Photo Upsampling via Latent
Space Exploration (PULSE) algorithm that upscaled lowresolution
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 official
(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 AI/ML models,
various biases might creep into a system; extending beyond
a preference for white/male 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 fairness therefore alerts the
AI/ML community 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 ensure diversity, are nondiscriminatory, 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
standards 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. Further, 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 platform, on the one hand, and the spirit
of broader regulations, such as the GDPR, on the other.
Occurrences of the Country in GIS and Remote Sensing Data Sets
Occurrences in Data Sets
121
NONDISCRIMINATION AND DIVERSITY
Bias may also arise from personal and cultural prior knowledge.
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
regard to the availability of EO/RS satellites and data. In
RS, appropriate datasets for specific tasks may also be neglected
due to the limited availability of research funds.
Therefore, the selection of any specific source (and even
region) for data collection needs to be disclosed, and the
analysis resulting therefrom must clarify the scope of applicability
(e.g., geographic, cultural, and other limits) and
any inherent 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 necessary to ensure responsible
downstream/applied research.
However, the major challenge in
Sources:
https://freegisdata.rtwilson.com/
https://github.com/Seyed-Ali-Ahmadi/Awesome_Satellite_Benchmark_Datasets
Map: Matthias Kahl * Created With Datawrapper
FIGURE 3. The number of occurrences and therefore the presence of a country in RS and geographic
information systems datasets in [94] and [95].
102
the context of bias in training data
results from a lack of open, accessible,
representative, and generalizable
data [103], [104], [105] to permit
optimal training and appropriate
contextual labeling and semantic
segmentation. Further, datasets (e.g.,
in scientific publications) usually
cover only a small range of scenes/
cases in which the predictive models
and algorithms perform well. As a
result, the data used to understand
the fundamental context/relations
in EO may be inadequate and totally
inappropriate to make predictions
in other contexts, even where these
contexts are somewhat similar. For
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2022
https://freegisdata.rtwilson.com/ https://github.com/Seyed-Ali-Ahmadi/Awesome_Satellite_Benchmark_Datasets
IEEE Geoscience and Remote Sensing Magazine - December 2022

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