Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
tection, image mosaicking,
and large-scale processing (i.e., processing involving many
image tiles) [ 7].
Remote sensing is therefore facing new opportuni -
ties. However, such opportunities cannot be seized un -
less they come with the capability to provide accurate
products in a timely manner. A bottleneck of supervised image-processing-based pipelines is the need of training
the model on reference points that are specific to every
acquisition. To be accurate, most models need to be trained on the known samples coming from the image under study.
Since obtaining new ground samples of high quality for
each image acquisition is not realistic, to retrain or adapt an existing model without such ground samples becomes mandatory. Figure 1
illustrates situations where adaptive
models might be of great use. In these situations (which correspond to those considered in this article),
only one image, i.e., shown in red
in the figure, has sufficient reference labels (e.g., obtained in an extensive
ground campaign), whereas the oth -
ers have no labeled samples or have
them only in an insufficient number.
This setting is more realistic, since,
on the one hand, extensive labeling cannot follow the pace of image ac -
quisitions, and, on the other hand,
repetitive ground campaigns are simply not often an op -
tion, mainly for economic and manpower reasons. Indeed,
gathering ground information is costly and cannot always
be performed by photointerpretation. This is particularly true when the task concerns very large areas or considers
quantities that cannot be photointerpreted by an analyst,
such as chlorophyll concentrations [ 8], plant water stress
[9], or tree species [ 10].
To address the cases described in Figure 1, a possible
solution would be to bypass the problem and assume that
the model that is already available is robust enough to
process the new images accurately. Despite the fact that
this is possible only in cases where the new image is ac -
quired by the same sensor as the previous one, it is well
known that the direct application of a pretrained model
on a new data set often provides poor results because the spectra observed in the new scene, even though rep -
resenting the same types of objects, are different from those of the scene used for training. The differences can be related to a series of deformations (or shifts) related to
a variety of effects, such as a biased sampling in the spa -
tial domain (typically if the ground sampling has been
focused on a region nonrepresentative of the new scene),
changes in the acquisition conditions (including the il -
lumination or acquisition angle), or seasonal changes.
When the new data are acquired by a different sensor, the
strategy explained previously is simply not applicable, as
most models require that all images (or domains) provide samples of the same dimensionality (and where each di -
mension has the same meaning) at test time. In this case, fusion strategies exist but generally only apply to certain combinations of sensors, and they only use the bands
that these sensors have in common or that are reason -
ably similar [ 3], which prevents reusing the models that
are already trained on the first image of the region that becomes available (which can be crucial, for example, in
postcatastrophe interventions) and exploiting sensors’ synergies in multisensorial schemes.
To process remote sensing images efficiently and accu -
rately, modern processing systems must be designed to be Model Extension on
Wide SurfacesMosaicking Model Extension on Wide
and Asynchronous Scenes
Figu Re 1. Examples of cases where DA is necessary to extend a model to new image acquisi -
tions. In all three cases, the images can be from different sensors, but only the images in red
have extensive reference labels (i.e., can be used for training an accurate supervised model).
to pRoceSS R emote
SenSing image S
eFFicient LY an D
accu RateLY, moDe Rn
pRoceSSing SYStemS
muSt Be DeSigne D to B e
RoBuSt in t He Face o F
cHange S in acQuiSition
conDition S an D
tempo RaL SHi FtS anD,
iDeaLLY, to Be aDaptiV e
to SenSoR DiFFeRence S.
Authorized licensed use limited to: ASU Library. Downloaded on March 08,2024 at 03:13:37 UTC from IEEE Xplore. Restrictions apply.
june 2016 ieee Geoscience and remote sensin G ma Gazine 43
robust in the face of changes in acquisition conditions and
temporal shifts and, ideally, to be adaptive to sensor dif -
ferences. The need for adapting existing models has been acknowledged for many years, as shown by the signature extension community [ 11], [12 ], but the change in the
amount and nature of data (as well as their resolution) cre -
ated the need for a new research direction. In this article, we advocate that the solution can be found in DA strategies, a
field that is deeply rooted in statistical and machine learn -
ing [13], [14 ].
In general, DA aims to adapt models trained to solve a
specific task to a new, yet related, task, for which the knowl -
edge of the initial model is sufficient, although not perfect.
As a traditional example in computer vision, DA methods
have been deployed to take classifiers that recognize ob -
jects in pictures from commercial websites and adapt them
to the recognition of objects photographed by simple web
cameras [ 15]. In this example, the classifier is presented a
problem with the same objective (i.e., classifying pictures
into a limited set of object classes) and the same features
but where the data relations are slightly different. For ex -
ample, at Amazon.com the pictures have no background
and the object is mostly in the center of the image, while
this is not the case in webcam images. DA is therefore used
to adapt the classifier that is accurate on Amazon.com to the new data distribution. Of course, this is just one ex -
ample. In the DA literature, models are modified to adapt to new data spaces (multimodal), related tasks (multitask), or subtle changes in probability distributions (see a recent
review in [ 16]). The connections to the multitemporal,
multisensor, and multiresolution image classification tasks
discussed previously are strong [ 7].
The aim of this review article is to provide an introduc -
tion to the DA field and to provide examples of applications of DA techniques in remote sensing. With this in mind, we
draw a taxonomy of the DA strategies that have been pro -