Semantic-enabled Knowledge Management for GEOSS

wildlifeplaincityManagement

Nov 6, 2013 (3 years and 9 months ago)

58 views


EVALUATING TRANSFER LEARNING
APPROACHES
FOR IMAGE INFORMATION
MINING APPLICATIONS


Surya Durbha*, Roger King, Nicolas Younan,

*Indian Institute of Technology(IIT), Bombay

Center for Advanced Vehicular Systems (CAVS)

Department of Electrical and Computer Eng.

Mississippi State University, USA

Outline


Background


Image information
Mining


Transfer Learning


Knowledge Transfer in EO


Methodology


Results


Summary


Image
I
nformation Mining



Image information mining (IIM) of remote sensing
data deals with the retrieval and analysis of image
content in an image using various supervised,
semi
-
supervised,
and unsupervised classification
methods
.



In IIM
applications,
the goal is to link the semantic
conceptualization of a
phenomenon,
usually
represented by land use/ land cover
classes,
with
lower level image features.


The
“semantic gap” between the lower level
features and higher level conceptual
representation is usually reduced using a variety of
semantics
-
based
techniques
.

Understanding Related Knowledge


The challenge with human analysis of imagery is that
human
possesses an information channel that is
bandwidth limited.


The
net result is an inability to cope with the information
content in the imagery and any associated co relatable
sensors due to the breadth and the quality of the data
sources themselves.


However
, the intuitiveness of the human mind is
unparalleled and has the ability to make associations
between similar objects which is very hard to replicate
in a machine interface.


Capability to Make Associations


For example, the human mind can look at a region
of agriculture, grass land, forest patch, and shrub
land to make inferences about the similarity or
dissimilarity between various regions


The
use of prior knowledge and contextual
information about these land cover classes in
making these distinctions forms a part of the
human mind’s approach for pattern
recognition



Also, there is the ability to transfer knowledge
between entities that are similar irrespective of
them belonging to the same or interrelated
domains.


Knowledge Transfer in EO Domains



The active transfer of knowledge between
various classes (entities) is vital in many
situations in EO domain such as
:


Disaster Response


Change detection


Land Use/Land Cover


Thematic information translation from one
classification system to another




Transfer Learning for Disaster
Response


Disaster response scenarios
where the knowledge of a
particular previous disaster in a closely related domain
might be able to help in damage assessment in a new
disaster.


For example, understanding the urban classes (buildings,
streets, bridges etc) and a post earthquake identification of
these classes, should help to provide an insight into how it
would look after a closely related disaster (urban warfare,
terrorist attacks, floods etc.)


Learning
new models based on transferring prior knowledge of similar
classes between closely related tasks or domains.

Transfer Learning for Change
Detection Studies


Change detection studies use continuously updated
information to update the databases and the maps that are
produced from them.


However
, to update these databases with new information
pertaining to the current time period, it may not be always
possible to obtain a large amount of labeled samples for a
supervised classification
.



In this
situation,
transferring the already available
knowledge from a source task and using it to classify a
target task at the current time using only a few labeled
samples is useful.


Transfer Learning in a Coastal
Disaster Scenario


This
work
is focused on applying transfer learning
and studying the inter and intra domain knowledge
transfer capabilities
in
a coastal disaster scenario.


For
this
purpose,
we adapt a transfer learning
methodology based on a modified Weighted Least
Squares Support Vector Machines (WLS
-
SVM)
and investigate the recognition rate on small
sample sizes of the target classes.

Methodology



In traditional data mining the assumption is that the training
and testing data are in the same
distribution, which
is not
true in several real world situations, especially in the EO
domain where there is dynamically changing information,
although contextually it is the same entity, but in a slightly
different form
.



For example, in a coastal disaster event, an “
agriculture
land
” could
change into
a “
flooded agriculture land
”.
H
ere
the data distributions would be different, but the underlying
concept is very similar.


Semi
-
supervised methods

allow the use of
unlabeled
data
from various spatial databases to augment the paucity of
labeled testing (target) data, but they also are constrained
by the same data distribution
issues.


Methodology (Cont.)


In this
work,
we apply and evaluate the transfer
method
based on transferring instance information (proposed by
Tomassi

et al,2010,
Oraborna

et al., 2009)


The approach is based on a modified
WLS
-
SVM method to
induce transfer between instances and is based on
constraining the
hyperplanes

(LS
-
SVM) of new category to
be close to those of a subset of the previously learned
classes.


The
least squares SVM formulation enables the closed
form of Leave one out (LOO) error
and
can be used for
model selection due to its unbiased estimate of the
generalization
error.



This formulation is adapted to constrain a new model to be
close to the pre
-
trained model.

Transferring Prior Knowledge



The
transfer of prior knowledge between a set of
sources classes and a target class could be
accomplished in two ways:


Inductive transfer learning
in which the knowledge is
transferred in situations where the domains could be
different or
the same
but the source and target tasks are
different.


Transductive

transfer
refers to the situations in which
the domains are different but the tasks that need to be
learned are the same.


Approaches



A common approach in transfer learning

is the
identification of three cases before applying the process;
“what to transfer”, “when to transfer”,
and “how
to transfer




The first case refers to the kind of entities that are
transferred between tasks; these could be grouped into
transferring:


K
nowledge
of Instances


K
nowledge
of
Features


K
nowledge
of
Parameters


Relational
K
nowledge

Transfer Learning in Coastal
Disasters Applications


In a coastal disaster event such as a hurricane, floods, and
other weather events, it is important to provide rapid
response.


The
normal machine learning algorithms require a good
amount of training data to develop a reasonable good
classification model.


However
, immediately after a disaster, the data is
sparse
and only few samples are available and using which it is
necessary to train the classification algorithms.


To
provide rapid response using maps updated by remote
sensing imagery, there is a need to be able to build models
using small samples
.



Also, using prior knowledge in such a way as to enhance
the learning of a new class would prove to be very useful.

Datasets


To demonstrate this scenario, we have used the
data sets from pre and post Katrina Landsat ETM+
imagery



We selected
six land cover
classes,
i.e
.,
agriculture, fallow, flooded agriculture and
developed areas, flooded forest areas.


The
goal is to assess the ability of transfer learning
of a target class with samples as small as 8 and
also understand the interaction between the
classes for their ability to exchange prior
knowledge

Results


The
case when all the 6 classes
(related and unrelated) were used in
a LOO approach. It can be seen that
both MKT and AKT performed much
better than WLS
-
SVM and was able
to provide fairly good recognition
rate for small target sample sizes. It
is evident that prior knowledge
between the categories has helped
in better detection.

Recognition
rate for 6 classes (agriculture, fallow, flooded
agriculture and developed, flooded forest, forest, water)

Results


The
number of classes (categories)
have been reduced to 4 (randomly
removed 2 classes) to assess the
transferability of knowledge, assuming
that certain classes could be responsible
for negative transfer.


However
, the results show that the
recognition rate is slightly less than
for
6
classes
(for
a sample size of
8).


But
it is not clear which of the classes
have contributed to the reduction in the
recognition rate, more simulations and
careful selection of the classes might
help in a better assessment of these
aspects.


Recognition
rate for 4 classes (water,
forest, agriculture, flooded agriculture
developed).

Summary


There is an increasing need to develop methods that work on
interdisciplinary and interrelated data to provide a holistic
perspective for decision making. So the need for models that
are able to adapt, use and transfer knowledge across the
domains is important.


The
adapted approaches for transfer learning in this work
have demonstrated their usefulness in practical applications
such as disaster recovery
.


However, more work is needed to understand the specific
situations under which transferability of knowledge is possible
and also on negative transfer issues.


Thank you!