Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
posed in recent remote sensing literature and discuss their
strengths and weaknesses. We also provide a series of prac -
tical examples about the use of DA in high- to very high-resolution image-processing tasks. However, we will not enter into the technical details of specific DA literature; for
interested readers, we refer to the recent surveys published
in [13], [14], [16], and [17 ].
tRanSFeR LeaRning anD Domain aDaptation
Transfer-learning problems arise when inferences have to be made on processes that are not stationary over time or
space. As previously discussed, this is the case in the anal -
ysis of remote sensing images where different acquisitions
are typically subject to different conditions (e.g., illumi -
nation, viewing angle, soil moisture, and topography). Such differences can affect the observed spectral signa -
tures of the land-cover types and, therefore, the distribu -
tion of the information classes in the feature space [ 18].
Different transfer-learning problems (and the techniques to tackle them) have been considered in the literature, including DA, multitask learning, domain generaliza -
tion, sample selection bias, and covariate shift [ 14]. In
this article, we will focus on DA, which is a particular form of transfer learning.
Let us consider two domains, called source domain and
target domain , that are associated with two images ac -
quired on different geographical areas (but with similar
land-cover characteristics) or on the same area at differ -
ent time instants. Figure 2 shows the DA problem in the
context of remote sensing image classification. The source
and target domains are associated with the joint probabil -
ity distributions
(, ) PX Ys and (, ), PX Yt respectively. The
two joint probabilities define the classification problems
on the two domains, where
X is the input (vector) vari -
able (i.e., spectral bands of
the source image with pos -
sible additional features used
to characterize the contex -
tual information of the single pixel) and
Y is the output
variable associated with a set
of classes (i.e., land-cover or
land-use information). The aim of DA methods is to adapt a classifier trained on the source domain to make predictions
on the target domain.
Supervised DA assumes that labeled samples are availa ble
for both domains. The labeled sets
xx{( ,),,(, )} Ty ys
nn 11f =
and xx{( ,),,(, )} Ty yt
mm 11f = are the source- and target-
domain training sets, respectively. Supervised DA meth -
ods focus on challenging situations where labeled target-
domain samples are less numerous than those available in
the source domain (i.e., ). mn11 In such conditions, the
proper use of source-domain information is very important
in solving the target problem. Most of the work in DA as -
sumes that source and target domains share the same set Source Image Target Image
DA
Ps(X,Y ) = Ps(X)Ps(Y|X) Pt(X,Y ) = Pt(X)Pt(Y|X) ≠
Figu Re 2. A graphical representation of the DA problem in the
context of remote sensing image classification. Source and target images can be acquired on different geographical areas (but with
similar land-cover characteristics) or on the same area at different times. The two images are associated with two different joint dis -
tributions, which characterize the two classification problems. The two distributions can differ due to different acquisition conditions (e.g., illumination, viewing angle, soil moisture, and topography).
tRanSFeR-LeaRning
pRoBLemS aRiSe wHen
inFeRence S HaVe to B e
maDe on p RoceSSeS tHat
aRe not Stationa RY o VeR
time o R Space .
Authorized licensed use limited to: ASU Library. Downloaded on March 08,2024 at 03:13:37 UTC from IEEE Xplore. Restrictions apply.
ieee Geoscience and remote sensin G ma Gazine june 201644
of classes. There are only a few papers that address DA con -
sidering differences in the set of classes between source and
target [ 19]–[22 ].
Semisupervised DA methods assume that a training set
is available only for the source domain, whereas target-do -
main information is limited to a set of unlabeled samples
xx{, ,} . Ut
m 1f = This DA setting is more challenging than
the supervised case and requires important assumptions on
the relationship between source and target domains to make
the algorithm converge to a consistent solution on the target
domain. All DA methods are based on the assumption that
(, ) PX Ys and (, ) PX Yt are different but close enough to en -
sure that the source-domain
information can be of help
for solving the target-domain learning problem. On the one
hand, if the source and target
domain are arbitrarily differ -
ent, there is no hope that the
source-domain information
will provide an advantage in solving the task in the target
domain. On the other hand,
if
(, )( ,) , PX YP XYst= no ad -
aptation is necessary, and the
model trained on the source
can be readily applied to the
target. Semisupervised DA methods are effective in situa -
tions that lie in between these two extreme cases.Unsupervised DA methods are the last family, and they
assume that two unlabeled domains have to be matched. This is the most difficult case because label information is not available for any domain. In this situation, DA methods
aim to match the marginal distributions of the two domains
()PXs and ()PXt without knowledge on the learning task
(classification or regression). Unsupervised methods can be used as preprocessing of any analysis task (e.g., clustering or
density estimation), but they imperatively need to have data sets with similar structural properties before adaptation.
Unsupervised DA models are generally feature extractors or