Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
to align unpaired data (see
the “Unpaired” column in Table 1), which allows the
alignment of noncoregistered data (not even imaging the same location) or data with different spatial resolutions.
◗The method should be able to align data of different dimen -
sionality (see the “ D Dimensionality” column in Table 1) to
allow multisource classification.
◗The method should be able to align several domains at the same time (see the “Multisource” column in Table 1)
to enhance multitemporal adaptation instead of pair -
wise adaptation.
◗The method should be able to align in a nonlinear
way (see the “Nonlinear” column in Table 1), since the −0.65 −0.645 −0.64 −0.635 −0.630.090.0920.0940.0960.0980.1
−∆P
−0.65 −0.645 −0.64 −0.635 −0.630.090.0920.0940.0960.0980.1
−∆P
657075808590
556065707580
(a) (b)
Figu Re 5. The Pareto front estimated using a multiobjective genetic algorithm for the selection of six features. Each dot corresponds to a
feature set minimizing (1). The color indicates the OA on (a) the source test set Ts and (b) the target test set Tt, according to the reported
color scale bar (adapted from [30]).
tHe aim o F Da met HoDS
iS to a Dapt a c LaSSiFieR
tRaine D on t He SouRce
Domain to ma Ke
pReDiction S on tHe
taRget Domain .
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 201648
transformation between domains can be nonlinear be -
cause of atmospheric or illumination effects.
◗The method should be able to use labeled information
from the source domain when available (see the “Labels
in s” column in Table 1). A discriminative transform tends
to align the data sets better because it aligns the data ac -
cording to the semantic classes required by the user.
◗The method should avoid being forced to use labeled
information in all domains (see the “No labels in t”
column in Table 1), as labels might not be available in
all domains or their acquisition might have a high cost,
typically through terrestrial campaigns (see the “Adap -
tation of the Classifier by Active Learning” section).
Several methods have been proposed in recent remote sensing literature. We provide a brief review in this sec -
tion and a summary in Table 1 . Depending on the specific
situation, the analyst can use this table to select the most suitable approach.
Most of the recent literature focuses on feature-extrac -
tion strategies, where the extracted features align the data spaces with each other. In that space, the same classifier (or regressor) can be applied to all the domains. Beyond
the works dealing with traditional or multidimensional
histogram matching [ 33] or data alignment with principal
component analysis (PCA) or kernel PCA (KPCA) [ 34],
the authors of [35] propose to minimize the statistical distance between domains, which is assessed through a kernel-based dependence estimator, the maximum mean
discrepancy (MMD) [ 18]. Other studies still focus on fea -
ture extraction, but based on multiview models. In [ 36],
Nielsen aligns domains with canonical correlation analy-sis (CCA) and performs change detection therein. The ap -
proach is extended to a kernel and semisupervised version
in [37], where the authors perform change detection with
different sensors. In [ 38], the domains are matched in a la -
tent space defined through an eigenproblem aiming at pre -
serving label (dis)similarities and the geometric structure
of the single manifolds. A nonlinear (kernelized) version of the algorithm is proposed in [ 39], where the approach
is particularly appealing because it can align an arbitrary
number of domains of different dimensionality, as do CCA and kernel CCA (KCCA), but without requiring paired
examples. However, it has the disadvantage of requiring
labeled samples in all domains.
In [40], the authors relax this requirement by working
on semantic ties, i.e., samples issued from the same object
but whose class is unknown. This last method therefore requires at least a partial overlap between the images to
find the ties, either manually or by stereo matching, as in
[41]. The authors in [ 42] regularize the manifold alignment
(MA) solution with spatial information, leading to a more
stable feature representation transfer. In [ 43], they propose
a multiscale approach, considering the preservation of
both local and global geometric characteristics and relying
on clustering pairs rather than labeled correspondences.
Other recent methods rely on eigendecompositions,
such as those proposed in [ 44] and [45 ]. In [ 44], two PCA
eigensystems (i.e., one for the source domain and another
for the target domain) are aligned by minimizing their di -
vergence. In [ 45], the authors consider a sparse represen -
tation approach where they reduce the difference between domains again by minimizing the MMD. In both [ 44]
and [45 ], the authors aim to transfer category models that
are learned on landscape views to aerial views from very high-resolution remote sensing images. In [ 46], the authors
propose a set of techniques based on sample reweighing
and transformation to address different DA situations. The
study also offers a causal interpretation of the different forms of domain shift. The adaptation strategies are devel -
oped on the basis of the embedding of sample distributions in the reproducing kernel Hilbert space.
Beyond classical feature extraction, the authors in [ 47]
align multitemporal sequences based on a measure of simi -
larity between sequence barycenters, which corresponds to a global alignment of the spectra in a time series of images.
In [48], the authors consider spatial shifts in large image TABLE 1. Th E REPRESENTATION ALIgN mENT m EThODS USED IN REmOTE SENSINg.
mEThOD LABELS IN s NO LABELS IN t mULTISOURCE UNPAIRED D DImENSIONALITY NONLINEAR
Pca # { # { # #
KPca [34] # { # { # {
(ss)tca [35] # { { # { # {
cca [ 36] # { # { { # { #
Kcca [37] # { # { { # { {
ma [43] { { { { { #
ssma [ 38] { # { { { #
Kernel method for
manifold ali Gnment [ 39] { # { { { {
Gm [50] # { # { # #