text
stringlengths
0
820
decision tree update and randomization are used for DA in [57]. A set of decision trees is made robust in the face of
data set shift either by training it with the EM using den -
sity functions from the target domain (similar to [ 54]) or
by randomizing the decision trees (but, in this case, with -
out control of the adaptation objective). The authors also propose a semisupervised extension where the classifiers performing poorly on the few labeled samples in the target
domain are downweighted in the final decision. Finally, the
DA technique proposed in [ 20] for multitemporal images
addresses the challenging situation where the source and
target domains have a different set of classes. The sets of
classes of the target and the source domains are automati -
cally identified in the DA step by the joint use of unsuper -
vised change detection and the Jeffries–Matusita statistical distance measure. This process results in the detection of classes that appeared or disappeared between the domains.
The semisupervised problem has also been extensively
studied in the framework of kernel methods with SVM classifiers, which has been done especially for addressing
sample selection bias problems (see the “Transfer Learn -
ing and Domain Adaption” section). Most of the semisu -
pervised techniques proposed with SVM exploit the clus -
ter assumption, i.e., adapt the position of the hyperplane
estimated on the source domain to the target domain, assuming that it should be located in low-density regions −38.79 −29.16 6.09 26.76 39.50.20.30.40.50.60.70.80.9
Acquisition AngleKappa
Best Case (Labels from All Images)
Training with Image at 6.09° OnlyPCA KPCA
GM SSMALDA−38.79 −29.16 6.09 26.76 39.50.20.30.40.50.60.70.80.9
Acquisition AngleKappa
SVM
Figu Re 7. The classification results over the five Rio de Janeiro acquisitions (adapted from [38]). There are 100 labeled samples per class
from the nadir image used to train a classifier and then to test on the others. In the SSMA experiment, 50 labeled pixels per class are used
from the other acquisitions.
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 51
of the feature space. In [ 58], the authors employ the trans -
ductive SVM, a method that iteratively moves the deci -
sion boundary of the SVM classifier toward low-density
areas of the unlabeled target domain. Other semisuper -
vised approaches are imported in remote sensing in [ 59],
where the SVM semisupervised learning is addressed in the primal formulation of the cost function. In [ 60], the
authors regularize the SVM solution by adding a new
term in the optimization accounting for the divergence
between source and target domains (i.e., the MMD dis -
cussed in the “Adapting Data Distributions” section). By
doing so, the decision function selected depends on a ker -
nel that both projects in a discriminative space and mini -
mizes the shift between training and test data. In [ 61],
the authors cast the DA problem as a multitask learning
problem, where each source–domain pair (i.e., each task) is solved by deforming the kernel by sharing information
among the tasks.
The Laplacian SVM technique applied to the classifica -
tion of multispectral remote sensing images is presented in [62], which exploits an additional regularization term on
the geometry of both the labeled and the unlabeled sam -
ples by using the Laplacian graph. In [ 63], the authors also
use a manifold-regularized classifier in a semisupervised
setting, where the adaptation is performed by adding semi -
labeled examples from the target domain.
A specific semisupervised SVM that is defined for ad -
dressing DA problems is presented in [ 64]. The domain ad -
aptation SVM (DASVM) starts from a standard supervised
learning on the training samples of the source domain,
which is followed by an iterative procedure. At each itera -
tion, it includes in the learning cost function a subset of
unlabeled samples of the target domain adequately select -
ed while gradually removing the training samples of the
source domain. At convergence, the DASVM can accurately
classify the samples of the target domain.ADAPTATION OF THE CLASSIFIER
BY ACTIVE LEARNING In most of the previously mentioned approaches, it is as -
sumed that no label information can be obtained in the
newly acquired target domains (i.e., semisupervised DA).
This assumption may hinder the success of classification in
the case of very strong deformations or when new classes that are unseen during training appear in the test data. A
small amount of labeled data issued from the target do -
main may solve this problem efficiently. However, since
the acquisition is timely and can be costly, it becomes man -
datory to choose the samples well. AL strategies have been
proposed to tackle this challenging task and guide the DA process with the selection of the most informative target
samples [19], [27], [28], [ 65]–[67 ].
AL is the name of a set of methodologies aiming at the
interaction between a user and a prediction model, where the user provides labels based on knowledge of the task
to be solved and the model performs the prediction and highlights samples for which it has the highest uncertainty
[68]. By focusing on these samples, the user provides the
labels where they help the most and, thus, allows the clas -
sifier to migrate in a fast way toward the optimal model. Surveys on AL methods applied to remote sensing can be
found in [ 69]–[71 ]. In the case of DA, the user provides
examples coming from the target domain alone, and the
optimal classifier is the one that would have been obtained
with several examples in the target domain. The process starts with a classifier that is optimal for the source do -
main and gradually evolves to model the data distribution in the target domain. Figure 8 summarizes the AL process
for DA.
One could apply classical AL strategies in transfer-learn -
ing problems under the sample selection bias assumption with successful results, because classical AL will point out
samples close to the current decision boundaries, and the
00.511.522.53
Target Domain
xiyiUncer tain
ResponseProvides
LabelUserSource Domain
(XS, YS)
(XT)ClassifierClassification
Confidence
Classification Map
Figu Re 8. A flow chart of the AL paradigm for DA.
Authorized licensed use limited to: ASU Library. Downloaded on March 08,2024 at 03:13:37 UTC from IEEE Xplore. Restrictions apply.