Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
and updating, the use of images remains static in the sense
that images are mostly used for visualization or, at best, to
compute standard indices such as the normalized differ -
ence vegetation index, which is then assumed to represent
vegetation status. Moreover, models answering the specific
needs of users are scarce and more often limited to classical
processing tasks (e.g., car detection or land cover mapping)
and cannot cover the variety of tasks in which different us -
ers could be interested. Another limitation is that end users
rarely have the technical skills necessary to design and run
ML models and would like to receive an answer to a spe -
cific question of interest asked in a natural language (e.g.,
in English).
Fortunately, many of these questions boil down to the
presence of objects, to counting, or to some kind of re -
lational attribute (e.g., whether there was an increase of
forest area or whether there are buildings in risk zones):
a model capable of pursuing some kind of reasoning
about the image content (see direction 1 in Table 1 ) but
taking into account a specific question (in English) by
a user could open the door to a new type of interaction
with remote sensing. Similar to what search engines do
on the Internet, a remote sensing VQA (RSVQA) [ 12] en-
gine could allow anyone, from scientists to laymen and
journalists, to retrieve relevant information contained in
the images.
Research into VQA is a vivid topic in computer vision
[11], where it has had considerable impact on creating sys -
tems that support vision-impaired people with everyday
tasks [ 58]. A traditional VQA system, in this context, can in -
deed be used to help people when buying groceries, cross -
ing the street, and so on.
DIALOGS AMONG USERS AND EARTH OBSERVATION
IMAGES REQUIRE BOTH REMOTE SENSING AND
NATURAL LANGUAGE PROCESSING
In remote sensing, the first VQA system was proposed by
Lobry et al. [ 12] and is summarized in Figure 3 . To become
truly general purpose, such a model needs to be trained us -
ing a large quantity of data from several areas and different
thematic objectives: in [ 12], two models were designed, one
for Sentinel-2 data and another for subdecimeter-resolution
aerial images. The models were trained with large sets of
image/answer pairs spanning tasks of classification, relative
position reasoning, and object counting. As a large quantity
of labels was necessary, OpenStreetMap (OSM) vector data
were used to automatically generate labels: following the
Compositional Language and Elementary Visual Reason -
ing (CLEVR) protocol [ 59], 100 questions per image involv -
ing objects occurring in the image (as informed by OSM)
were generated. For each image/question pair, the answer
(i.e., the label) was automatically obtained by querying
OSM directly. The data and models are openly available
Authorized licensed use limited to: ASU Library. Downloaded on March 07,2024 at 22:07:36 UTC from IEEE Xplore. Restrictions apply.
94
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2021at https://rsvqa.sylvainlobry.com/. Examples of RSVQA
model predictions are reported in Figure 4 for both resolu -
tion images. Note that, for a single image, several questions
are possible and the same model is used to answer them.PERSPECTIVES
This first work opens a wide range of possibilities for
a new line of research toward the next level of human/
image interactions. Nonetheless, all the blocks of the
FIGURE 3. An example of an RSVQA system [12]. (a) A remote sensing image and (b) a question in natural language enter two source-specific
neural networks, each one outputting a vector representing their information, and (c) both vectors are combined and become (d) the input of
a classifier that outputs all the possible answers as separate classes. RNN: recurrent neural network.
“How many cars are
there?”CNN
RNN6
7
8Yes
No
Peri Urban
Cropland
City Center
Increasing4) Question Answering 3) Feature
Fusion 1) Image Representation
2) Textual Representation(a)
(b) (c) (d)
Q: “Is there a road?”
A: “Yes (GT: yes)”
Q: “Is there a farm? ”
A: “Yes (GT: yes)”
Q: “How many small residential
buildings are there?”
A: “Between 10 and 100 (GT: 147)”
Q: “Is there a residential building?”
A: “Yes (GT: yes)”
A: “Yes (GT: yes)”
Q: “How many small residential
buildings are there?”Q: “Are there fewer commercial
buildings than water areas?”
A: “Between 10 and 100 (GT: 18)”Q: “Is there a baseball field?”
A: “Yes (GT: yes)”
A: “0 (GT: 0)”
Q: “Is there a building south of a road?”Q: “How many commercial buildings are
close to a baseball field?”
A: “No (GT: Yes)”
Q: “Is there a commercial building?”
A: “No (GT: no)”
A: “Yes (GT: yes)”
Q: “How many small buildings are there?”Q: “Are there fewer commercial buildings