Seabed classification using SBES data

mudlickfarctateAI and Robotics

Nov 25, 2013 (3 years and 8 months ago)

49 views

Seabed

type

clustering

Identification

of

seabed

type

(such

as

mud,

sand

and

rock)

is

of

value

in

many

applications

including

seabed

mapping,

coastal

management

and

seabed

conservation
.

There

are

two

main

types

of

echo

sounders

for

such

purposes,

multi
-
beam

(MBES)

and

single
-
beam

(SBES)
.

The

difference

between

the

two

types

of

echo

sounders

is

that

while

SBES

has

one

transceiver

that

emits

and

detects

echo

time

series

at

normal

incidence

to

the

seabed,

MBES

has

multiple

transceivers

sending

sound

waves

at

various

angles

towards

the

seabed
.

Usually,

seabed

clustering

is

performed

from

data

collected

by

MBES

during

sea

survey

trips

while

single
-
beam

echo

sounders

(SBES)

are

usually

employed

for

the

measurement

of

bathymetry

(depth)
.

However,

it

is

believed

that

SBES

time

series

data

also

contains

useful

information

for

seabed

type

analysis

due

to

its

ability

to

achieve

deeper

penetration

of

the

seabed

substrate
.



Challenges

of

SBES

clustering


Influence

of

environmental

factors

(salinity,

sea

water

temperature,

bathymetry,

slope)
;


Characteristics

of

SBES

echo

which

are

frequency
-
dependent
;


Operational

limitations

(automatic

gain

control,

ship

movement)
.



Ground

truthing

All

existing

clustering

approaches

require

extensive

ground

truthing

(such

as

grabbing,

coring

and

visual

inspection)

which

are

not

practical

in

deep

waters
.

By

adopting

a

statistical

approach,

the

objective

is

to

perform

multi
-
frequency

clustering

as

an

alternative

to

extensive

ground

truthing

prior

to

label

validation
.

In

this

study,

features

from

three

frequencies

(
12
,

38
,

200

kHz)

are

collected,

spatially

filtered

and

normalised

before

clustering
.


Seabed classification using SBES data

Peter Hung (NUIM), Seán McLoone (NUIM), Xavier Monteys (GSI)





Background

Research

presented

in

this

poster

was

funded

by

a

Strategic

Research

Cluster

Grant

(
07
/SRC/I
1168
)

by

Science

Foundation

Ireland

under

the

National

Development

Plan
.

The

authors

gratefully

acknowledge

this

support
.

Methodology

Results and future work

Fig
.

1

Single
-
beam

echo

sounder

Fig
.

2

Transceiver

and

data




capture

equipment

0

50

100

150

200

250

-
100

-
80

-
60

-
40

-
20

0

20

Echo return (dB)

Distance from sea level (m)

Peak

Amplitude

Water

column

Seabed

substrate

A

B

Region of

Interest

Fig
.

3

Data

processing

flowchart

Fig
.

4

SBES

time

series

Feature

extraction


Due

to

the

high

dimensionality

of

time

series

data,

the

large

number

of

samples

collected

in

sea

trips

and

the

heteroscedastic

noise

contained

within,

a

total

of

four

features

are

extracted

from

the

raw

echo

data
.

The

temporal

mean

and

associated

standard

deviation

are

standard

features

that

has

direct

relationships

with

seabed

geology
.

To

convey

information

about

the

relationship

between

adjacent

time

series,

measures

of

spatial

randomness

and

spatial

correlation

are

proposed
.



Optimal

depth

selection

A

novel

adjacent

mean
-
square
-
error

metric

is

used

to

estimate

the

optimum

time

series

interval

for

feature

extraction
.

This

is

based

on

the

assumption

that

seabed

substrate

should

produce

gradually

changing

features

unless

the

echo

captures

additional

information,

such

as

sidelobe

backscatter

and

background

noise
.

Determined

separately

for

the

above

and

below

time

series

segments,

optimal

depth

is

defined

as

the

point

of

inflexion

or

minimum

in

the

following

plots
.


Semi
-
automatic

statistical

approach


To

achieve

consistent

and

reliable

classification,

each

processing

step

needs

to

be

carefully

assessed
.

Since

SBES

data

contains

less

geological

information

and

redundancy

for

quality

assurance

compared

to

MBES

data,

additional

measures

need

to

be

taken

to

ensure

the

quality

of

the

raw

data
.

The

pre
-
processing

stage

involves

further

expert

inspection

for

data

integrity,

initial

‘cleanup’

to

mitigate

the

effects

of

systematic

deviations

of

sonar

measurements

and

spatial

sampling

to

improve

the

signal
-
to
-
noise

ratio
.


Seabed

detection

Two

approaches

are

employed
.

The

first

assumes

the

seabed

is

located

at

the

peak

amplitude

of

each

echo

time

series

return
.

The

second

method

estimates

the

location

of

the

seabed

by

spatially

smoothing

the

bathymetry

using

a

second
-
order

Butterworth

filter
.

Utilising

both

bathymetric

approaches

allows

the

detection

of

‘bad’

data

samples
.




0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
0
20
40
60
80
100
120
140
160
180
Above seabed (m)
Mean Adj. MSE


12 kHz
38 kHz
200 kHz
0
5
10
15
20
25
0
5
10
15
20
25
30
Below seabed (m)
Mean Adj. MSE


12 kHz
38 kHz
200 kHz
Above seabed (m)

Below seabed (m)

5 m

2.2 m

1.3 m

9.2 m

2.1 m

Fig
.

5

Optimal

depth

selection

from

mean

adjacent

mean
-
square
-
err

Clustering

on

features

derived

from

optimally

selected

intervals

yields

more

distinctive

and

continuous

(and

hence

geologically

plausible)

clusters

than

if

the

full

time
-
series

intervals

are

employed
.



Preliminary

results

suggest

that

the

best

performance

is

obtained

by

employing

QT

local

clustering

with

the

optimally

selected

intervals
.


Overall,

data

fusion

using

multi
-
frequency

SBES

data

improves

information

richness

and

data

quality

control

improves

the

reliability

of

clustering

results

and

provides

a

more

geologically

realistic

interpretation

of

the

survey

area
.



Future

work
:


Improving

the

computation

and

memory

efficiency

of

the

QT

local

and

MSC

clustering

algorithms



Determination

of

an

optimum

feature

set

from

SBES

time
-
series

for

seabed

classification

Fig
.

6

Clustering

results

from

optimal

and

maximal

depths

Optimal depth

Maximal depth

Optimal depth

Maximal depth

PCA +
k
-
means

QT local

Best
result

Clustering

PCA

with

k
-
means

is

the

industry

standard

for

clustering

seabed

data
.

However,

it

is

less

effective

at

dealing

with

non
-
Gaussian

clusters

and

requires

the

number

of

clusters

to

be

specified

a

priori
.

In

this

work

extensions

of

quality

thresholding

approaches,

including

local

quality

thresholding

(
QT

local
)

and

max
-
separation

clustering

(
MSC
)

are

developed
.

These

have

several

advantages

over

k
-
means,

including

automatic

cluster

number

determination

and

robustness

to

outliers
.