Seagrass Health
UMaine engineers develop
neural networks to help biologists monitor declining underwater meadows
Marine biologists from Maine to
Australia are working with electrical and computer engineers to monitor
an ecosystem – seagrass meadows – in retreat. Seagrass, which provides
critical habitat for commercially important fish species, occupies about
10 percent of the world's coastal seas. Unlike seaweeds, these rooted
underwater plants flower and drop their leaves like their land-based
cousins.
In the 1930s, a disease decimated a type of seagrass known as eelgrass (Zostera
marina) in Atlantic waters bordering North America and Europe. Although
the cause is uncertain, it has been linked to warming water temperatures
and microorganisms.
The beds eventually recovered, but today, eelgrass and other species of
seagrasses appear to be in decline worldwide. Some losses have been
severe. In the past 20 years, areas of the Indian River Lagoon on
Florida's Atlantic coast have lost as much as 95 percent of their
coverage. Maine's Taunton Bay has lost about 90 percent of its eelgrass
in the last six years. Additional declines are being monitored in
Australia and Europe.
A continuing decline could deal another blow to an already struggling
global fishing industry.
The job for scientists is clear: understand what's causing the decline
of these delicate habitats and develop ways to restore them. Researchers
already know that reduced light transmittance through the water is a
major factor. The problem usually starts at the deeper edges of the
beds, where the light reaching the plants is only marginal, and
progresses toward shallower regions as conditions deteriorate. Reduced
light is often related to murky water conditions that result from land
erosion.
To address this problem and even predict seagrass stress before more
beds are lost, biologist Suzanne Fyfe at the University of Wollongong in
Australia is using light reflected from seagrass leaves to develop an
early warning system. Currently, satellite or aircraft remote sensing
techniques can only detect deterioration in seagrass health after
large-scale dieback has already occurred, she says.
To turn measurements of reflected light into a predictive tool, Fyfe has
turned to University of Maine Assistant Professor Habtom Ressom, who
leads a research team in the Intelligent Systems Laboratory (INTSYS) of
the Department of Electrical and Computer Engineering. Ressom
specializes in a computer software system known as an artificial neural
network, or neural net.
In INTSYS are three faculty members, a research associate and more than
a dozen graduate and undergraduate students. The seagrass project is one
of several active studies in the lab. Others focus on DNA analysis, gene
expression and industrial process control. Working on the seagrass
project with Ressom are research associate Padma Natarajan, a 1999
UMaine graduate, and electrical engineering master's student Siva
Srirangam.
The common thread running through the lab's research is the use of
computational intelligence techniques to extract knowledge from data.
Neural nets have been around for more than 40 years and today are widely
used in industry and business. They improve voice transmission over
telephone lines, teach machines to talk, recognize patterns and analyze
financial markets. While they consist of sets of mathematical equations,
neural networks are nevertheless inspired by nature. Individual parts of
a neural net are viewed as nerve cells and the connections between them
as the junctions that link cells.
"From an engineering perspective, our brains are essentially neural
networks," says Ressom, who received his Ph.D. from the University of
Kaiserslautern in Germany in 1999. "We can learn things from what we
see. We can correct things. Initially, an artificial neural network has
no idea about the relationships between inputs and outputs. As it runs,
it will see its own errors and modify its own parameters."
To the casual observer, the neural net seems to perform statistical
magic. It doesn't depend on knowledge of a specific system, but it does
require quality data. Moreover, its ability to learn and adjust gives
the neural net an advantage over conventional modeling approaches,
especially in dealing with complex systems.
A seagrass ecosystem fits that model. A case in point is Fyfe's effort
to predict seagrass stress on the basis of reflected light. Fyfe uses a
device known as a spectroradiometer to identify the changes in the light
reflected from seagrass leaves.
Ressom's neural network transforms the database of information into a
mathematical tool. That tool can then be applied to remote sensing data
to predict stress levels in sea-grass meadows before dieback occurs.
Running mathematical models is a bit like playing a game of darts in the
dark. Scientists may know that their results are accurate within a
certain range, but they don't know exactly how close they are to the
bull's-eye. By improving model accuracy, neural nets turn the lights up
a bit, letting scientists know that their results are closer to the
mark.
Poor water quality, including murkiness or algae that keep out sunlight,
is a threat to seagrass. In collaboration with scientists from St. Johns
River Water Management District in Florida, Ressom and his team are
studying how to use water quality data to achieve a better estimate of
light reduction and, thus, seagrass health.
"The district monitors seagrass in the Indian River Lagoon in
correlation with water quality parameters, such as nutrients, turbidity
and clarity. They call seagrass ‘the barometer of the ecosystem,'" says
Ressom.
In 2002, Ressom met with scientists from the district in a workshop on
biodiversity and ecosystems for Indian River Lagoon. He co-organized it
with Natarajan and other UMaine faculty members: Mohamad Musavi, George
Markowsky, Thomas Wheeler, Anthony Stefanidis and Cristian Domnisoru.
The district already uses a model to correlate water quality parameters
with light attenuation, says Ressom. Using the district's water quality
data, the UMaine neural network came closer than the district's own
model in predicting the relationship between water quality and light
traveling through the water and the impact on seagrass health. "Neural
networks try to correlate difficult-to-measure variables with
easy-to-measure variables," says Ressom. "The advantage is that no prior
information is necessary. That's why we are able to jump into these
subjects. Our background is in electrical and computer engineering. I
personally have no knowledge of the biological relationships."
In addition to their seagrass work, Ressom and his team are working with
the National Aeronautics and Space Administration to apply a neural net
to ocean data from satellites. Their goal is to estimate chlorophyll
concentrations, an indication of algal growth and ocean vitality.
by Nick Houtman
July-August, 2003
Click Here
for more stories from this issue of UMaine Today Magazine.
