Understanding Ocean Surface Temperature Features

Abstract

The aim of this project was to develop a prediction system that uses Artificial Intelligence, machine learning using training data and Image Processing (AI) to extract training data from Sea Surface temperature (SST) images to predict the ocean surface, temperature features around the coast of the Southern African region. Region growing and histographic algorithms were used in the image processing section to extract thermal fronts as training data from the available SST images. A Temporal Bayesian Network was developed as the prediction model which used approximate stochastic learning and inference algorithms based on the Maximum Likelihood Algorithm (MLE). User-Centered Design (UCD) and Human-Computer Interaction (HCI) methods were used to develop user-friendly and easy to understand Graphical User Interfaces (GUI). Results and evaluations of the project revealed that a generally successful prototype implementation of a prediction system that used AI, machine learning and image processing was developed.

Full text

Understanding Ocean Surface Temperature Features
Technical Paper Number:
Submitted as part of the University of Cape Town,
Department of Computer Science Honours Project 2005
Nemanja Spasic
Jar
e
d Tilanus
Anet Potgieter
University Of Cape Town
University Of Cape Town
University Of Cape Town
Abstract
The aim of this project was to develop a prediction system
that uses Artificial Intelligence, machine learning using
training data and Image Processing (AI) to extract training
data from Sea Surface temperature (SST) images to predict
the ocean surface, temperature features around the coast of
the Southern African region.
Region growing and histographic algorithms were used in
the image processing section to extract thermal fronts as
training data from the available SST images. A Temporal
Bayesian Network was developed as the prediction model
which used approximate stochastic learning and inference
algorithms based on the Maximum Likelihood Algorithm
(MLE). User-Centered Design (UCD) and Human-
Computer Interaction (HCI) methods were used to develop
user-friendly and easy to understand Graphical User
Interfaces (GUI).
Results and evaluations of the project revealed that a
generally successful prototype implementation of a
prediction system that used AI, machine learning and
image processing was developed.
Categor
ies and Subject Descriptors:
G.3
[Mathematics of Computing
]:
Probability Statistics
Probabilistic
Algorithms, Stochastic Processes, Time
Series Analysis
;
H2.4 [Database Mana
gement
] System
Relational Database; H5.2 [Information Interfaces and
Presentation
]: User Interfaces
Graphical User
Interfaces (HCI), U
ser
-centered design
;
I 2.4 [
Artificial
Intelligence
]: Knowledge Representation Formalism and
Methods
Temporal logic; I 2.3 [Artificial Intelligence
]:
Deduction and Theorem Proving
Inference Engine; I 2.6
[Artificial Intelligence]: Learning
Parameter Learning;
I 4.6 [Image Processing and Computer Vision
]
Segmentation
Region Growing, Pixel Classification;
I
5.1 [
Pat
tern Recognition]: Models
Statistical;
I 5.3
[
Pattern Recognition
]: Clustering
Similarity Measure
General Terms:
Design, Theory, Human
-
Factors
Additional Key Words and Phrases:
Bayesian Network, Temporal Bayesian Network,
Approximate Stochastic Learning, Approximate Stochastic
Inference, Poisson Distribut
ion,
Human
-
Computer
Interaction
, Region Growing, Histographic, Segmentation.
1.
I
NTRODUCTION
"Prediction is very difficult, especially if it's about the
future."
Nils Bohr, Nobel laureate in Physics
Although the above quote, from Nils Bohr
[10]
, perfectly
describes the reality of prediction, it is one of the most
common activities that is done in all aspects of today s
world. Some examples of prediction include predicting the
weather, predicting sports results, predicting people s
behavior, predicting the age of matter using carbon dating,
predicting experimental results, etc.
The uncertain or unforeseen factors that are involved in
prediction make prediction so difficult.
There have been several attempts at minimizing the effect
of such uncertain and unforeseen factors to make
prediction more accurate. Some of the attempts include
using experience, using scientific reasoning and using
previous observations or data, known as training data , as
a guideline to the prediction. Lately, the use of statistics to
model future predictions based on probabilities has gained
interest within the researching and Computer Artificial
Intelligence (AI) communities. An example of using
statistics to aid with prediction can be found in Nielsen s
work
[11]
, which uses statistics to predict the risk in a wall
condensation technique.
Although prediction can be narrowed down to a small
prediction error rate using such techniques, for example
prediction of weather, one c
an never be 100 percent certain
that the prediction that was made is an accurate prediction.
That is the main driving forcing behind finding better
prediction techniques that will give better predictions about
desired events.
This project focuses on detection and prediction of ocean
surface, temperature features that are present in the
Atlantic and Indian oceans around the Southern African
Region using image processing, Artificial Intelligence
(AI), machine learning and stochastic algorithms.
Furthermore,
the project presents the use of User-
Centered
Design (UCD) and Human Computer Interaction (HCI) to
design user-friendly and easy to use interfaces for the
prediction system.

The project was taken in collaboration with the Zoology
Department at the University Of Cape Town and a French
company named de l Institut de Recherche pour le
Dévelop
ement
.
2.
MOTIVATION AND
B
ACKGROUND
2.1 Motivation
The collaborating companies expressed the following
problem areas, which created the project domain:
A visualiza
tion
and prediction tool for
SST
images
(including daily, 5 day and monthly)
.
Identify
persistent
and re
-
occurring
features
Develop a prediction system that will be user-
friendly and easy to use and interpret by non-
expert users.
In
addition to the above p
roblems,
the fishermen and
a
nglers in the Southern African r
egion
could benefit form a
prediction system that could monitor and predict ocean
surface, temperature feature. The reason being, as Dan
Rudnick suggests, Oceanic weather influences the
distribut
ion of phytoplankton, which other marine life
feeds upon.
[13]
Hence, by monitoring and predicting the
ocean surface, temperature features the fishermen and
anglers could predict where the most likely accumulation
of phytoplankton would be found, thus leading to the most
likely accumulation of marine life. Furthermore, biologists
who study the migration of fish to monitor whether or not
the ocean surface temperature features influence the
migration of fish could also benefit from the project
prediction sys
tem.
Thus,
the project would add to the value
of Artificial Intelligence and computer science in today s
world.
Finally, as Semtner A.
[15]
suggests that only the most
advanced parallel computers are fast enough to produce
high
-quality ocean simulations and accurate global climate
predictions of temperature and
precipitation. Furthermore,
Wang P. et al
[17]
support the views of Semtner in terms
of computationally expensive ocean models. Thus, there is
a need for a prediction system that is not computatio
nally
expensive and can be utilize
d
on a personal computer (PC)
,
which what the targeted users of the system will have
access to
.
2.1.1 High
-
Level Project Requirements
From the
stated
problem areas, the following overall
project requirements were establis
hed:
Design software to detect the actual positions of
temperature features from SST images
Design a prediction model for prediction of ocean
surface features over a time series period
the feature
bei
n
g: fronts, filaments and gyres
Design a prediction model that will be able to identify
ocean surface, temperature features that are persistent
features and features
that are re
-
occurring features.
Design a prediction
system
, which will be able to use
existing data
as training data
Design a prediction model that uses minimum system
requirements in its predi
ction of ocean surface features
Design a prediction model that will be user-
friendly
and easy to use for non
-
expert users
2.1.2Division of project work
The project was divided into two sections, which are
detailed below: Image Processing and Understanding of
the available ocean surface temperature data and C
reation
of a prediction tool, using AI and machine learning, which
can predict ocean surface temperature features over time,
using available data as tr
aining data.
2.2 Background
2.2.1. Threshold Algorithms
Thresholding is used to distinguish classes of pixels in an
image based on their intensity value
[16]
. In the simplest
(binary) case an image is segmented by assigning pixels to
one class if they ar
e below the threshold and to the other
class if they are above or equal to it.
One can extend this to more than two classes of pixels,
where a class is defined by an upper and a lower threshold.
Where k is the number of classes, there are k
-
1 thresholds,
and a class is defined as all pixels with an intensity value
of between k
i
and k
i-1.
Histographic methods are designed to automatically
identify the optimal threshold to use in an area. A
histogram of the pixel values in the area is created and
examined
for the existence of 2 peaks.
[12]
describes a method where one segments the image
using a threshold and evaluates the goodness of
segmentation for each of the possible threshold levels for
the image. If the maximum goodness value (the value
correspo
nding to the best threshold level) exceeds a given
limit, the area is said to have 2 populations.
The goodness measure is defined to be the ratio
of the total
variance to the within class variance.
[3]
describe a very similar approach for applied to th
e
identification of SST fronts. They apply this method on
local, overlapping windows within the image. The size
of the windows is a critical tuning parameter in this
approach, as is the amount of overlap between these
windows.
If a window is identified
as having two populations the
special compactness of the populations is evaluated. This
compactness is measured as the ratio of edges between

pixels of the same class to the total number of edges
between pixels.
A more statistically correct method for hi
stographic
threshold identification is to fit 2 Gaussian distributions to
the histogram and take the intersection between these two
distributions as the threshold value.
2.2.2. Region Growing
Region growing as the name suggests is the process of
merging n
eighboring areas into
larger
regions to segment
an image.
A region can be defined as follows: let R be and image and
R, R
2, Rn
such that:
Where P(R) is a logical predicate over the points in set R
i
and Ø is the null set.
One can group them these met
hods into two main classes,
seeded and unseeded.
Seeded region growing
[1]
starts from some set of points
selected either manually or automatically and grows
regions around these points. Pixels bordering on these
regions are iteratively added to the seed
ed regions until all
the pixels in the image are included in one of the regions.
Thus in the final segmentation each region will contain
exactly one of the seed points.
Unseeded region growing starts from many small regions
and merges them based on some
measure of similarity to
form larger regions. These methods usually begin by
dividing the image up as blocks. Split and merge
[6]
methods make these initial blocks large and then split them
recursively until the blocks meet some homogeneity
criteria. Other
approaches include making each pixel in the
original image a region, or making regular blocks of a
small number of pixels. These small regions are then
merged until some criteria such as the number of regions
or a threshold in the difference between regio
ns is reached.
2.2.
3
Bayes
Rule
Due to the rule of symmetry in probability the conditional
probability can be expressed without using the joint
probability
P (A,B). Hence, an alternative way to calculate
P(A|B) is possible, and is known as Bayes Rule. The rule
can be modeled with the following equation, using e as an
evidence variable and H as the hypothesis:
The use of Bayes
Rule underlies all modern AI systems for
probabilistic
inference
[5]
2
.2.
4
Bayesian Network
s
(BN)
A Bayesian Network (BN) is an AI technique that uses
probabilistic reasoning to calculate the conditional
probability of an event occurring, based on related,
existing data about the event, which is known as training
data. The calculations are based on a cause-
effect
relationsh
ip between the events in the network. However, a
BN can not model temporal relations between and within
the random variables.
A BN can be graphically represented using a directed
acyclic graph (DAG) structure
which has to conform to the
following rules:
1.
A
set of random variables, either discrete or
continuous, or uncertain quantities make up the
nodes of the BN.
2.
A set of directed links or arrows connects pairs of
nodes. This is a representation of a cause and
effect relationship.
3.
Each node N has a set of
conditional probability
distributions P(Ni | Cause(Ni)) , which can be
stored in a conditional probability table (CPT),
that quantify the effect of the cause(s) or parent(s)
on the node.
4.
The graph has no directed cycles.
5.
The root(s) of the network should h
ave a prior
probability P(root) in stead of the CPT
The following
figure
, adopted from
[14
], gives an example
of a Bayesian network.
It represents a situation where an
alarm responds to a burglary, but also goes off when an
earthquake occurs, notifying the
police and the fire
department
.
Figure
1: Showing a Bayesian Network
The tables next to each node represent the CPT tables for
that particular node. P(X) denotes the probability of X
being true. This could be extended to include the
conditional probabilities of each node being false, by the
simple calculation: 1
-
p.
2.2.
5
Bayesian Network Learning/Training
Before one can start using a BN, the network structure and
parameters have to be specified from the available training
data. Murphy K.
[8]
suggests that in order to train a BN,
two operations have to be performed. Firstly, the structure

of the BN has to be established and then the parameters of
the CPT tables have to be evaluated. Leaning the structure
of the BN can prove to be an NP-hard problem and b
e
much more difficult than parameter learning, especially
when the BN has hidden nodes that are not directly
observed from the training data.
There are four cases that have to be examined for learning
a BN: when the network structure is known and data
var
iables are observable, when the network structure is
known and the data variables are hidden / missing, w
hen
the network structure not known but the data variables are
observable
when the network structure not known and the
data variables are hidden/missin
g.
In the first case where the structure is known and the data
variables are fully observable the leaning required would
be parameter learning of the CPT.
[8]
and
[14]
suggest the
use of a maximum-likelihood estimation (MLE) algorithm
to do parameter lear
ning in this case.
To illustrate this type of learning, the following example
can be used, which was adopted from
[8]
:
where
a
produces supplies bags of balls to a sport facility, which
can contain only be soccer balls or volley balls. The
probability distribution of the bags is not known. Hence
the probability of getting a bag with only volley balls is ,
and the probability of getting a bag with only soccer balls
is 1- . The fraction of volley balls in a particular bag can
be denoted a hypothesis h. Suppose that N bags were
received, of which there were (s) soccer balls and (v)
volley balls. The following mathematical model can be
used to demonstrate this
example
:
The MLE can be used to do parameter leaning in this case
and is given by the value of that maximizes the above
expression.
Using the log-
li
kelihood the above equation
can be model as : vlog + slog(1- )
.
Differentiating the
equation with respect to one gets:
Hence, the proportion of volley balls in a given bag is the
proportion of observed volley balls from the existing data.
This means that this type of learning is a simple count of
occurrences over the entire data set. When the data is
continuous, a Gaussian distribution can be used to perform
MLE instead. However, the MLE algorithm has
disadvantages with small data sets. A more detailed
elaboration on the M
LE algorithm can be found in
[14]
,
[8]
and
[9]
The second case, where the structure of the network is
known
, but some data variables are hidden or data is
missing, according to
[14]
and
[8]
, would be best suited to
the Expectation Maximization (EM) algorithm which uses
inference. The EM algorithm can be used to find a
(locally) optimal Maximum Likelihood Estimate of the
parameters.
[8]
An exhaustive explanation of the EM
algorithm can be found in
[14]
,
[8]
and
[9]
.
In the third case, were the network structure is not known,
but all the data variables are observable a search through
all possible network structures needs to be performed to
determine which network structure fits the data best. This
problem is an NP-hard problem.
Possibl
e solutions to this
case are to use greedy algorithms (e.g. scoring metric or
hill
-climb algorithm) and clustering technique which find
the most likely number of clusters. This case is exhaustive
explained in
[14]
,
[8]
and
[9]
According to
[8]
, the last case, where both the network
structure is not known and the data variables are hidden or
data is missing can be solved by using a combination of the
EM algorithm and a network structure search. Murphy K.
[8]
also suggests that introducing hidden variables c
an
make the BN model more compact and can sometimes lead
to easier structure learning and better results.
2.2.
6
Bayesian Network Inference/Querying
The basic task of inference in a BN is to establish the
posterior probability of a set of query variables/
nodes
given the state of some observed event / variable. In other
words, this means that we wish to know the probability of
and effect variable given the current state of the cause
variables. Inference in a BN can be either exact or
approximate. The choice of inference methods to use is
dependent on the network structure. Furthermore, the
choice of exact inference or approximate inference
depends on the quantity and availability of data. If the data
that is available is sparse and only a small quantity
of
d
ata
is
available,
then approximate inference would be more
suitable because to perform exact inference, a large data
set is required. Finally, exact inference tends to be more
computationally expensive than approximate inference,
hence is one is looking for a BN that is less
computationally expensive and exact results are not a strict
requirement, then approximate inference would be more
suited.
Possible BN Learning Algorithms include, for exact
learning
:
Enumeration
[14]
,
junction tree
[14]
[9]
[2]
,
varia
ble elimination
[14]
[9]
, clustering algorithms
[14]
,
linear algebra (for Gaussian nets)
,
Pearl's message passi
ng
algorithm (polytrees)
[14]
and for approximate inference:
likelihood weighting
[14]
,
[9]
and sampling algorithms,
e.g. Mar
kov chain Monte Carl
o (MCMC)
[14]
,
[9]
2
.2.7
Dynamic/Temporal Bayesian Network
A Dynamic Bayesian Network (DBN) is an extension of a
BN to facilitate the modeling of temporal observations of
stochastic random variables over discrete time steps.

Murphy
[8]
suggests that a "temporal Bayesian network"
(TBN) would be a better name than "dynamic Bayesian
network", since it is assumed that only the parameters of
the model change and the model structure over time.
Hence, a DBN will be referred to as a TBN in the rest of
the
report
.
A
TBN has to have the following
CPTs
for each variable:
a
prior probability, a transitional model CPT from time slice
t to t+1 and a sensory model CPT for the attributes of the
nodes/variables.
These can be specified for the first time
slice only as a TBN is assumed to be time-
invariant
, hence
the structure of the TBN is assumed to be constant in each
time slice. The first time slice is simply un
-wrapped for the
required number of time slices when an inference
operation is performed. In such a way, a TBN can be
converted into a BN. An example of a TBN would be the
following example which relates to the project where a
n
ocean temperature feature s (e.g. a front) attributes change
is being predicted over time. Hence, the attributes can
either change or not. The attributes that the feature has are
temperature and
position.
The following diagram could be
used to model the TBN (unwrapped for 2 time slices):
Figure
2
: Showing a Temporal Bayesian Network (TBN)
2.2.
8 Hidden Markov Mod
els
A Hidden Markov Model (HMM) is a temporal
probabilistic model which has only one discrete random
state variable per time slice and models how that state
variable behaves over time. The state variable is linked to a
single discrete hidden node that forms the HMM. Hence,
an HMM is the simplest kind of TBN. One has to note that
all HMMs are TBN, but not all TBN are HMMs.
[14]
The
example below taken from
[8]
demonstrates a possible
HMM
, where Q is the hidden node and Y is the observed
node. The model is un
-
wrapped for four time slice
s:
Figure
3
: Showing a Hidden Markov Model
2
.2.
9
Temporal Bayesian Network Learning
Murphy
[9]
suggests that the learning techniques for TBN
are mostly straightforward extensions of the learning
algorithms applied to BN. The only difference between
TBN and BN is that online, as well as offline, parameter
learning can be applied. This is a great advantage when the
training data is continuously being increased or updated.
As in a BN, a TBN has structural learning and parameter
learning. Hence, most of the previously mentioned learning
algorithms for BN are applicable to TBN. A discussion of
all the learning algorithms that can be applies to TBN can
be f
ound in Murphy
[9]
.
2
.2.
10
Temporal Bayesian Network Inference
The most common types of
inference performs using a
TBN include: filtering, prediction and smoothing and these
are shown in the diagram below.
Figure
4
: Showing various kinds of inference in a Temporal Bayesian
Network
[9
]
Several algorithms exist to perform inference in
a
TBN. As
in the case of BN, inference can be exact or approximate.
The same ideology for choosing between exact inference
and approximate inference in a BN applies to a TBN
because the TBN is an extension of a BN. Some of the
inference algorithms include for exact inference: f
orward
-
Backward algorithm Dugad
[4]
, Murphy
[9]
,
[14]
,
J
unctio
n tree algorithm Barber
[2]
, Frontier algorithm
Murphy
[9]
and for approximate Inference the Boyen-
Koller (BK) algorithm Murphy
[9]
,t
he factored fr
ontier
(FF) algorithm Murphy
[9]
, loopy belief propagation
(LBP) Murp
hy
[9]
, s
tochastic algorithms Murphy
[9
]
2
.2.
11
Poisson Distribution
The Poisson distribution (PD) is a discrete probability
distribution
[18]
that expresses the probability of a
number of events occurring in a fixed time if these events
occur with a known average rate, and are independent of
the
time since the last event.
[18]
There
is no limit to the
number of values that a Poisson random variable can
assume which give it several advantages over other
probability distributions such as the binomial distribution.
As
adopted from
[18]
, the Poisson distribution of k, a
poison random variable, is given as:
Where: k is a non-
neg
ative integer
k = 0,1,2..
n,
e is the base of the natural logarithm (e =
2.71828...), k! is the factorial of k, is a positive r
eal
number that is, according to
[18]
, equal to the average
number of successes occurring in a specified interval. The
cumulative Poisson distribution, which is modeled as:
And
can be used to get the cumulative effect all Poisson
random
events k0..kn, where kn is the desired Poisson
random event.

2
.2.1
2
User
-
Center Design
User
-Centered Design (UCD) is a methodology used to
design interactive computer programs for targeted users.
UCD places the targeted users at the center of the design
stage and builds the final product based on the reaction of
the user to the continually evolving prototype. The
international standard ISO 13407
[5]
is a standard that
provides guideline for UCD that should be followed when
de
signing an interactive comput
er
based systems. The
standard
specified
UCD principles that have to be met to
ensure a user centered design is achieved
and
also specifies
the
activities that should be carried out to ensure that the
principles of UCD are met in designing a product.
3
.
METHOD
3
.1
.
Histographic Algorithm
This algorithm follows the idea of the algorithms described
in
[3]
and
[12]
. It segments the image by looking at local
25 * 25 pixel windows within the image. For each
possible threshold level in the window
(256 for the
1bit
grey scale images used in this project)
the optimal
threshold is determined. Based on this threshold, it
determines whether the window contains 2 distinct
populations of pixels and whether these 2 populations are
spatially distinct. If both of these
criteria are true, a front is
determined to exist.
For each pixel intensity level in the image it is assumed
that two populations exist, separated by this threshold. This
segmentation is then assessed by calculating the variance
within the whole populati
on V
T
and the variance within
each class V
1
and V
2
and calculating the ratio R = (V
1
+
V2
) / V
T
. The threshold that produces the highest value for
this ratio is considered to be the optimum threshold.
The variance is defined as
Where x
i
is the value
of pixel i and µ is the average
intensity of all the pixels in that window.
If the ratio value R corresponding to the optimum
threshold is above a given threshold the window is
considered to contain two classes. These classes are then
checked for their sp
atial compactness by checking the ratio
of pixels that occur on the inter
-
class boundary. The
number of pixels that neighbor on pixels of a different
class is calculated by looking at each pixel and its
neighbors below and to the left. In this way all the
boundaries between pixels are checked. If the ratio of this
number to the number of pixels in the smaller class is
above a certain ratio, the classes can be said to be in
spatially distinct.
If the window passes both these tests, a front exists
between t
he two classes identified. The pixels on the inter
-
class boundary are marked as edge pixels and
3
.2
.
Region Growing Algorithm
This algorithm is an unseeded region growing algorithm
that segments an image by growing homogeneous regions.
Regions are define
d here as a group of spatially coherent
pixels. The image is initially divided into tiny regions, each
containing one pixel. At each iteration the two most similar
neighboring regions are merged to form a bigger region.
Several similarity measures were i
mplemented in this
project. They include one based on the difference between
the average pixel intensities of the regions, one based on
the edge strength and one based on a combination of these
two, with deferent weightings.
On initialization each border
between regions is located
and an object representing this edge is created and placed
in a priority queue. This object stores the difference
between the two regions it connects, and thus the most
similar regions can be found by taking the edge off the
fro
nt of the priority queue.
When regions are merged, their point sets are joined. Since
the similarity measure is based on the pixels in the region,
all the edges which pointed to the old regions have to be
updated. This requires removing them from the pr
iority
queue and re
-
inserting them with their new values, a very
resource intensive job.
Edges that, after a merge, end up
joining the same two regions also have to be merged.
The regions are iteratively merged in this way until a
stopping criterion is
reached. This could be a threshold in
the similarity measure, or a number of regions. The latter
was implemented in this project.
3.
3.
Prediction Tool
Figure 5: Showing the Prediction Tool
3.
3.
1 The Prediction Tool
The prediction tool was designed using AI agents for each
feature prediction and was modularized to improve
performance
, database access
and testing capability.

3.
3.2
Temporal Bayesian Model
The TBN network chosen to represent each ocean surface,
temperature f
eature was the following:
Figure 6: Showing the TBN model
3.
3.
3 Learning algorithm
An approximate
,
stochastic
, online parameter learning
algorithm was applied, which was based on the Poisson
distribution.
The learning algorithm was adopted from the
MLE a
lgorithm
and was a by-product of the inference
algorithm.
3.
3.
4 Inference algorithm
Due to the nature of the training data, an approximate,
stochastic inference algorithm based on the MLE algorithm
was used to perform inference. The current feature
attrib
ute was matched to the closest matching training set
data to perform the MLE algorithm. In the case where a
feature s attribute was a complex one, the Euclidean
Distance was used to evaluate the closest training data
match.
3.
3.
5 Prediction Tool Interface
A Prediction tool interface was designed to allow easy
access to the prediction tool and to display the prediction
results. The visualization of the prediction results was in
the form of graphs and tables, which were found to be
user
-friendly and easy to
use.
The prediction engine
interface was designed to hide the internal workings of the
TBN from the end
-
user.
Figure 7: Showing the Prediction engine results
3.
3.6
Database selection
A MySQL database server was chosen to store the training
data.
3.
3.7
Update Module
An update module was designed to allow the users to enter
training data into the database and to hide all the internal
workings of the database from the end
-
user.
3.
3.8
User
-
Centered Design (UCD)
UCD was used in the form of participatory and PICTIVE
sessions to ensure that users were at the center of the
design stages of all interfaces that were created. Iterative
prototypes were designed based on these sessions and were
continually evaluated by users, then modified until the
final user-
fr
iendly and easy to use interfaces were
established.
Human Computer Interaction (
HCI
) methods
were also
employed
to ensure that a user-friendly interface
was developed. In addition, HCI methodologies were used
while evaluating the interface to ensure the
participants
of
the sessions were treated with respect.
4
. RESULTS
4.1 Image Processing
Due to the nature of the problem, one can not perform a
rigorous evaluation of the segmentation. To do that one
would require a ground truth set of images to compare to.
From a casual observation one can see that the histograhpic
algorithm seems to identify fronts more reliably. It does
not, however form continuous lines, as can be seen here
Figure 8
: Showing
the discontinuous lines
produced by
the histographuc algorithm
These
are usually due to the window level boundaries,
where different windows have different optimal thresholds.
The region merging algorithm is particularly sensitive to
tuning parameters. This can be seen by the drastic
difference between the images produced with different
parameters.

Fig
ure 9: two different sets of parameters for the region merging
algorithm applied to the same image. The first image here is produced
using a merging criterion of the difference between average pixel
intensity and
the strength of the edge, weighted equally. The edge strength
is calculated as the maximum strength along the edge. The second image
uses the criterion as the edge strength only, defined as the maximum
strength of along the edge. The third uses a merging criterion of edge
strength only, taken as the average over the whole edge.
4.2
Prediction
Tool
4.2
.1 Learning Algorithm
The learnt Poisson rates were compared to Poisson tables
and were found to match, hence this result showed that a
successful learning
algorithm
Comparison Of Poisson Rates
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0 2 4 6
Lambda
Possion Rate
Actual Poison
Rates
Learning
Algorithm
Poisson Rates
Fig
ure
10: Showing resul
ts of multiple
-
cross validation
4.2
.2 Inference Algorithm
The inference algorithm was tested using multiple-
cross
validation using a sub-set of the training data. Results
showed that the error rates for the inference algorithm were
varied and can be attributed to tuning of i
nference
parameters using sampled training data. However, after
further evaluation of the inference algorithm, results
showed that a reasonably acceptable inference algorithm
was developed.
Multiple-Cross Validation Testing
0
0.1
0.2
0.3
0.4
0.5
0 5
10 15
Iteration
Error Rate
Error Rates
Fig
ure 11
: Showing resul
ts of multiple
-
cross validation
4.2.3
Evaluation of GUIs
The evaluations of both the update module and the
prediction engine were done by means of questionnaires,
and in general confirmed that the interfaces were user-
friendly and easy to use by non
-
experts.
5
. CONCLUSION
Comparing the delivered project to the project
requirements, one can conclude that almost all the required
goals were accomplished, except for the fact that the Image
Processing part of the project could only detect thermal
fronts, and could not detect filaments and gyres. Besides
this requirement, all the other project requirements
were
met. Although the prediction engine was not tested using
real
-time data for filaments and gyres, it was tested using
thermal front training data and was found to predict
reasonably acceptable results. Hence, because the same
algorithms were used in all three feature predictions, one
can state that the prediction would also work for filaments
and gyres, although this would have to be tested further.
Although it was not possible to
rigorously
evaluate
the
dete
ction
algorithms, they appear to
recognize
most fronts.
The region growing algorithm is very sensitive to tuning
parameters, and does not perform as well as the
histographic algorithm.
According to the evaluation done, the interfaces designed
were user-
friendly
and easy to use by non-
exp
erts, which
met the initial project requirement.
H
ence
, to conclude, through a global view, the project was
an
implementation
of
a prediction system which could
predict ocean surface, temperature features which used AI
techniques, machine learning and Image P
rocessin
g to
extract training data successfully.
6. FUTURE WORK
6.1 Prediction model
The prediction model that was
implemented
was
a
naïve
model in which all attributes and all feature were assumed
to be
independent
. Hence, a good are for future work
would be to extend the model to incorporate feature and

attribute
interactions to design a more realistic ocean
prediction model.
6.2 Learning algorithm
The learning algorithm that was implemented was a
parameter learning algorithm that assumed that the
structure of the
network
was known, As future work, this
could be extended to a structure learning algorithm, which
attempts to find the best possible network
structure
for the
training data set.
6.3 Inference Model
The inference algorithm that was implemented was a
naïve
maximum likelihood inference which assumed
attribute
independence
. This could be extended to an
inference
algorithm that takes into
account
the interactions with all
other attributes and features for a more realistic inference
prediction.
7
. REFERENCES
[1]
Adams, R. and Bischof, L. 1994. Seeded region
growing.
IEEE Transactions on Pattern Analysis and
Machine Intelligence
, 16:641
647
[2]
Barber, D. 2003. Probabilistic Modelling and
Reasoni
ng Dynamic BayesianNetworks: Discrete Hidden
Variables
Available
at :
http://anc.ed.ac.uk/~dbarber/pmr/pmr.html
[3]
Cayula, J. F., and Cornillon, P. 1992. Edge detection
algorithm for
SST images.
Journal of Atmospheric and Oceanic
Technology
, 9, 67
80.
[4]
Dugad, R et al. 1996. A tutorial on Hidden Markov
Models.
Indian Institute of Technology, Bombay, Signal
Processing
and Artificial Neural Networks Laboratory,
Technical Report SPANN
-
96.1
[5]
I
SO 13407:1999
Human
-
centred design
processes for
interactive systems
.
[6]
Horowitz, S.L. and Pavlidis, T. 1974, Picture
segmentation by a directed
split
-
and
-
merge procedure
. In Proc. 2nd. Int. Joint Conf.
on Pattern Recognition
, pages 424
433.
[7]
Lin,
Z., Jin, J. and Talbot, H. 2000. Unseeded region
growing for 3D image segmentation. Selected
papers from
the Pan
-
Sydney workshop on Visualization
, p.31
-
37,
December 01, Sydney, Australia
.
[8]
Murphy K., (1998) A Brief Introduction to Graphica
l
Models and Bayesian Networks [Online]
Available
at:
http://www.cs.ubc.ca/~murphyk/Bayes/bnintro.html
Date
accessed
: 20/9/2005
[9]
Murphy, P. 2002 .Dynamic Bayesian Network:
Representation
, Inference
and
Learning.
Available
at:
www.ai.mit.edu/~murphyk/Thesis/thesis.pdf
[10]
Niels Bohr, Education on the Internet &
Teaching
History[
Online
]
Date
Accessed:6/9/2005
Available
at:
http://www.spartacus
.
schoolnet.co.uk/USAbohr.htm
[11]
Nielsen, A., (1995), Use of statistics for prediction of
risk for condensation in a wall construction cold walls
radiation technique:
the case of quality of air
,
Int
ernational
Symposium On Moisture Problems In Building Walls
,
Porto
-
Portugal, 11
-
13 September, pp. 1.
[12]
Otsu, N. 1979. A threshold selection method from
gray level histograms,
Pattern Recognition
, Vol 9, no. 1,
pp. 62
66
.
[13]
Robert Monroe (Ma
y/June
2002)
Underwater
Weather: From liquid hurricanes to aquatic fronts,
scientists seek to unlock the mysteries of oceanic
weather
.
Weatherwise magazine
[14]
Russell S., Norvig P. (2003) Artificial Intelligence A
Modern Approach 2 ed New Jersey: Pearson Education,
Inc.
[15]
Semtner A.(April 2000) Ocean and climate
modeling
ACM Press New York, NY, USA
[16]
Sahoo, P. K. and Soltani, S. and Wong, A. K.C. and
Chen, Y. C. 1988. A survey of thresholding techniques.
Computer Vision, Graph
ics, and Image Processing
,
volume 41, number 2, pp 233
260
.
[17]
Wang P., Katz D. S., Chao Y (November 1997)
Optimization of a parallel ocean general circulation model
Proceedings of the 1997 ACM/IEEE conference on
Supercomputing (CDROM)
[18]
Wiki
pedia, the free encyclopedia : Poisson
d
istribution [Online] Available at:
http://en.wikipedia.org/wiki/
Poisson_distribution
Date accessed: 25/9/2005

This document was created with Win2PDF available at http://www.daneprairie.com.
The unregistered version of Win2PDF is for evaluation or non-commercial use only.