IEEE Geoscience and Remote Sensing Magazine - June 2016 - 17

...

View 2 (v2)

View k (vk)

f1
f2

Target Label (y )

...

View 1 (v1)

Consistency Constraint:
f1(v1) = f2(v2) = . . . fk (vk )

FiGure 6. A schematic illustration of multiview learning where

the data is made up of multiple views, v1-vk, and model fi learns
the mapping from vi to the target class (i varies from 1 to k). The
models are constrained to agree with each other over unlabeled
instances in accordance with the consistency constraint.

the data. Cotraining then makes use of the fact that because
every model represents the same target class, even though
each model was learned using a different view of the data,
the predictions of the multiple models should mutually
agree with each other on any given unlabeled instance. Cotraining leverages this concept to constrain the learning of
every predictive model to be consistent across all unlabeled
instances, which are plentifully available in a number of
applications. This ensures that the learning of a predictive
model at every view of the data doesn't suffer from poor
performance due to the phenomena of overfitting, which
commonly arises in the presence of limited training data.
Instead, by maximizing the
the multivieW learNiNG
mutual agreement among the
paradiGm provides aN
model predictions over all uneNtirely diFFereNt Way
labeled instances, multiview
oF leveraGiNG iNFormalearning is able to regularize
tioN From various vieWs
the learning at every view, resulting in the learning of a rooF the data While makiNG
bust representation of the taruse oF a huGe Number oF
get class. For example, in the
uNlabeled iNstaNces that
context of using regressionare easily available iN
based models for every view,
aNy laNd-cover chaNGe
the regularization among the
applicatioN.
views would correspond to
the minimization of the disagreements between multiple
model predictions on unlabeled instances along with minimizing the prediction errors on training instances. Multiview learning is particularly relevant when the multiple views
of the target class are highly heterogeneous and made up of a
diverse set of features.
Figure 6 provides a schematic illustration of the underlying concept behind multiview learning, where information
from multiple views of data is being combined to predict a
common target class. Furthermore, the predicted outputs
of models over different views are constrained to be similar
to each other over all unlabeled instances, with the theme
of cotraining-based multiview learning. This makes use of
the assumption that the different views of the target class
18

contain adequate information for discrimination but suffer
from improper learning due to the paucity of labeled data.
However, the learned model of every view is desired to provide consistent predictions with the learned model of other
views on unlabeled instances since each model attempts to
learn the same target class.
In the context of land-cover change detection, the different views of the data could be the multiple sources of
information about the same land-cover change phenomena. As described in the "Use of Multisource Data" section,
multisource data arises in a number of land-cover changedetection problems where different sources of data show
different data types and characteristics, making it difficult
to combine them in a joint analysis. Furthermore, groundtruth information is often scarce, resulting in a paucity of
training data. However, it should be noted that even though
training labels are scarce in a number of change-detection
scenarios, there is often an abundance of unlabeled data
instances during the testing stage because remote sensing
observations are easily available across large geographic
regions and over long temporal periods, which presents
ample opportunities for multiview learning to make use of
multiple sources of remote sensing observations to detect
land-cover changes with a high accuracy, even in the presence of limited training data. As an illustrative application,
consider the problem of predicting whether a particular
pixel inside an area of farmland has corn or soybean plantations in a given harvesting season using remote sensing
data. Due to the presence of noise in the remote sensing
signals and the complexity in distinguishing corn pixels
from soybean pixels, the performance of any classifier
on a single date in the harvesting season might be poor.
However, by considering remote sensing observations on
different dates in the same harvesting season as different
views of the same target class, i.e., whether a pixel is corn
or soybeans, we can learn a different classification model
for every date in the harvesting season. By constraining the
predicted class labels to be consistent across multiple dates
in the same harvesting season over unlabeled pixels, we
can leverage information across the different dates (i.e.,
views) for obtaining a more superior performance of corn
and soybean classification than what could be obtained by
a single classifier trained on a specific date.
As a promising direction of research, designing multiview techniques that are able to handle a high degree of
noise and missing values in each of the views will be very
useful for land-cover change detection because remote
sensing data sets often suffer from poor data quality due
to cloud and aerosol obstructions. However, the impact of
noise and missing values is different for different remote
sensing sources of data, obtained from different sensors
at different time steps. Hence, constraining consistency
among the predictions of different views can help overcome noise and missing values and therefore improve robustness by simultaneously using information from all of
the views. Multiview learning approaches can also be used
ieee Geoscience and remote sensing magazine

june 2016

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2016