Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
However, there remain significant issues to address:
◗The data to be included : Each source must add value, being
correlated to the task and independent from each other.
In an extremely multimodal setting, the volume and
velocity of data acquired from each source cannot be
directly controlled because they depend on completely
independent acquisition systems. Given that, it becomes
important to understand how the spatial coverage and
quality of each source vary. For example, clouds have
a significant impact on optical imagery but do not af -
fect synthetic aperture radar (SAR). Similarly, the cover -
age and quality of various social media sources depend
on a variety of conditions, including their proximity to tourist landmarks, population density, and differences
in the culture of a given social network (interested read -
ers are referred to discussions about social media data
quality in [ 36] and [ 37]). Therefore, we must choose
among not only satellite sources, but also social media
platforms, and must often apply further filtering, e.g.,
including only the data collected from cellphone cam -
eras [ 38] or only certain types of scenes [ 34].
◗The quality of the matching between sources : Each source
must be georegistered to reduce spatial uncertainty dur -
ing training. The satellite/aerial-to-ground matching at
the scene level is highly challenging due to the large se -
mantic gap between the ground and overhead scene, but
it is still resolvable. For instance, Lin et al. [ 39] proposed
a dual-adversarial solution for an unsupervised satellite/
aerial-to-ground scene-adaptation solution. However,
it becomes very crucial when object-level matching is
concerned, e.g., when approaching automatic geolocal -
ization [ 40], [41], and, in particular, when considering
image synthesis [ 42]–[44], where strong geometric mod -
els of the various modalities, with the ability to model
uncertainty, are strongly needed.
PERSPECTIVES
Considering social media data as extra sources for remote sens -
ing analysis is gaining momentum. Aside from classification
FusionQuer y Location
FIGURE 2. An example of a system that provides the likelihood of the presence of an object with multiple modalities. The ground and
satellite images provide hard and weak (picture-based) labels, respectively, to create a heatmap that shows the presence of objects in urban
areas. The location information is used to perform the fusion. (Sources: Google Maps; geoawesomeness.com; Wikipedia; Wikimedia.)
Authorized licensed use limited to: ASU Library. Downloaded on March 07,2024 at 22:07:36 UTC from IEEE Xplore. Restrictions apply.
93
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINEand automatic geolocalization, using these additional data
could unlock new applications, such as modeling sound -
scapes [ 45], landscape scenicness [ 46], or place-perception
analysis [ 47]–[49]. Further, one could study phenomena
closer to the consumer providing the data; for example, one
could think of a system finding the most visually appealing
driving route.
This will inevitably raise the question of data set biases
because social media data are personalized views of space:
photographers tend to take pictures from places that are
easy to reach; they are biased in the types of subjects they
prefer and usually take more pictures in the daytime and
in good weather conditions (see, for instance, the overs -
ampling of particular photographic forms and scenes in
the Instagram account insta_repeat (https://www.instagram
.com/insta\_repeat/). This problem was recently considered
when observations collected by citizen scientists were
used for species distribution mapping [ 50]. In general,
biases in learning models is a growing topic of study in
both the ML (see, for example, [ 51] and [ 52]) and social
media (see the reviews in [ 53] and [ 54]) communities.
There is ample room for such studies in remote sens -
ing and biases issued by fusion in multimodal settings
or in hallucinations when using generative adversar -
ial networks.
Using social media also implies the development
of models that are robust to differences in the appear -
ance of classes, which becomes critical when predict -
ing in new geographies or in time moments. As acquir -
ing new labeled data is not always an option, one could
envision using these alternative sources as a form of
weekly supervised training data or even as an unsu -
pervised supervisory signal for knowledge discovery,
as Law and Neira have proposed [ 55] to explore the ur -
ban latent space of London’s streetscapes. Considering
that the data acquired by autonomous vehicles and In -
ternet of Things devices will push this need even fur -
ther, it will also unlock the potential of mapping on
demand with extremely multimodal remote sensing.
Finally, further integration could make it possible to
use social media as an early detection system for events,
such as natural disasters [ 56]. For instance, social me -
dia imagery could be used to detect damaged structures
or people in distress [ 57]. Such a system could even be
used to cue satellite image acquisition over areas of in -
terest based on image content, location-data densities,
or tweets.
DIRECTION 3: INTERACTIVE AND SEMANTIC ML
With their massive increase in availability, remote sensing
images are now used beyond scientific research. First, im -
ages are available worldwide and with a high update rate.
But they are also much more accepted by the general public:
no one is surprised anymore when they are shown a sat -
ellite view from Google Maps; consumer-level drones can
be used by virtually anyone for all kinds of tasks, such as farmers monitoring crops, ecologists surveying animals, or
architects keeping track of construction sites.
But, despite the massive potential for image acquisition