2. Outline
• Introduction
i. Why classification
ii. Subbottom profiler
iii. Dataset description
iv. Classification methods
• High frequency classification
• Low frequency classification
• Conclusions and recommendations
3. Why classification?
Applications:
• Offshore projects need seabed maps and information about sediment type
Examples:
• Offshore engineering
− Harbours
− Oil platforms
− Wind farms
• Dredging
(dredged material are primarily related to the type of dredging equipment)
• Morphological studies
(develop sediment transport models to predict changes of the morphology of
the riverbeds and seabeds)
4. Classification methods
• Grab samples
- Costly and time consuming surveys
• Acoustic Remote Sensing
- Cost effective
- High spatial coverage
- Non destructive analysis Box core
5. Seafloor acquisition techniques
Single beam echo sounder (SBES)
-
SBES,SBP,SEISMIC
Operating frequency : +/- 100kHz
- Observation zone: seabed surface
- Applications : fundamental sensor for
seafloor mapping
Marine seismic (impulsive, sparkers, boomers)
- Operating frequency : 5000Hz – 50 Hz
- Observation zone: deeper than 100m up to several kilometers
- Applications : geophysical mapping (oil and gas exploration)
Sub-bottom profiler
- Operating frequency : 1kHz – 30kHz
- Observation zones : near surface (1-100m)
- Applications : offshore engineering projects
6. Sub-bottom profiler
Non-linear acoustics
Working principle:
Emit 2 or more +/-100kHz (Primary)
to produce 15,10, 5 kHz (Secondary)
F = |105-100| = 5kHz
Advantages: Two sources interference
Compact dimension r ~25cm that can emit low
frequencies.
High horizontal resolution due to the very narrow aperture
angle ~ ±1.8° for all frequencies
(very narrow comparing to SBES 30 °).
High vertical resolution ~ 5-10 cm due to very short pulse
length.
Beam width
7. Dataset
- The data was acquired in the Baltic sea
near Rostock-Germany in 2007.
- The image shows a set of amplitude pressures
that corresponds to layers contrast
area3 area4
area1 area2
8. Dataset
High frequency: to determine
accurate water depth.
(High frequency can not penetrate into deep layer
due to high sediment attenuations)
Low frequency: to observe the
seabed layers.
Area1, 100kHz Area1, 5kHz
9. Signal anatomy (High freq)
Leading edge due to initial Trailing edge due to scattering
Soft sediment echo reflection from seabed and reflections from Subbottom
- Low amplitude
- Large component of volume
scatter
Rough sediment echo
- High amplitude
- Large component of surface
backscatter
The received signal is a
composite surface and volume
backscatter
10. Classification approaches
(Phenomenological, Model based)
• Phenomenological
Extract features of received echo (time spread,
energy, skewness) that allows to discriminate
between different sediment classes.
Principle
1 - Correct for echoes for depth effect.
2 - Extract features from the corrected echoes.
3 - PCA (combination of all extracted features).
4 - Clustering the sets of principle components
corresponds to number of sediments types.
11. Model based
Model based
Employ a physical model that predicts the received echo
shape.
Principle
1 – For each ping, a signal is modeled
2 – Input parameters:
Transmitted signal (pulse duration, power etc..)
Transmission loss
Directivity
Sediment backscatter
3 – Search for the mean grain size that maximize the
match between modeled and observed echo.
12. High frequency analysis
Post processing
- Filtering
- Stacking
- Alignment
Classification (Model based)
- SONAR equation
- matching process
Classification results
13. Stacking
The stacking process is essential to:
- Remove the heave variability to isolate the
seabed type.
- Decrease the envelope stochastic variability
to improve the matching process.
- Suppress the ping to ping variability by
stacking consecutive echoes allows the
sediment information in the echo shape and
spectral nature to express it self.
Waterfall plot
14. Stacking
Ensemble size has a tradeoff
between the spatial resolution
and ensemble variance.
Ensemble size is indicated
through the correlation matrix
that measures the degree
similarity of the received echoes.
Result:
Ensemble size =15 signals
Correlation matrix
15. Alignment
The alignment is done with respect to a temporal
feature.
Two alignment methods were tested:
- Minimum threshold.
(preserves signal properties)
- Maximum alignment
(preserves signal integrity)
16. Echo model
APL Backscatter model Numerical integration Modeled backscatter
(angular domain) over ∆R (time domain)
Soft
sediment
+ =
Coarse
sediment
17. Matching process
Iterative loop :
- Select the geoacoustic
parameters of sediment Mz
- Model the echo envelope of
the selected Mz
- Compare function
- Select Mz that corresponds to
the minimum E/S
18. Results (1)
The classification results using alignment
10%:
1-High model/measurement matching
degree for the soft and medium sediments.
E/S
2-Low model/matching degree for coarse
sediments
3-Good classification consistency with the Soft
general description of the area properties
Moderate
except for area4.
Rough
Classification result,
alignment 10%
19. Results (2)
The classification results using alignment
50%:
1-Moderate model/measurement matching
degree for the soft sediments.
E/S
2-Good model/matching degree for medium
and coarse sediments
Soft
3-Good classification consistency with the
Moderate
general description of area3 and 4.
Rough
4- Poor classification consistency for area1
and 2.
Classification result,
alignment 50%
20. Results (3)
Classification results using alignment 100%:
1-Very poor model/measurement matching
degree for the soft sediments.
2-Moderate model/matching degree for medium E/S
and coarse sediments
3-No classification consistency with the general Soft
description of all areas
Moderate
Rough
Classification result,
alignment 100%
24. Signal anatomy (Low freq)
Leading edge due to initial
Low frequency echo reflection from seabed
- Discrete reflections at contrast
layers. Subbottom reflections Backscatters
from low contrast layers
- The short pulse length and very
narrow aperture angle ensures to
obtain high reflected energies with
low backscatter components. Subbottom reflections
from moderate contrast
layers
High reflections from
highly contrasted layers
25. Energy model
1- Surface classification
- The aim here is to infer the sediment type
from its reflection coefficient by comparing
it to the modelled reflection coefficients.
- The received energy ERX is related to
the transmitted energy ETX via:
Assumption:
- Backscatter is negligible compared to the
coherent reflection.
15kHz echo
26. Energy model
Energy comparison
1-Reflected energies are correlated with the sediment type.
2-Energy magnitudes are different (scale factor needed).
3- Incomplete energy profile for the SBES model due to the narrow
aperture angle (we are missing the backscatter.
4- High energy fluctuations from area4 due to high surface roughness.
Energy scaling
27. Surface classification
1- Five random samples were selected
to compute the scale factor C.
2- The reflection coefficients are
estimated for the 17 traces using
one scale factor via:
3- The estimated reflection coefficients
are inverted to the corresponding
mean grain size via:
28. Extended energy model
- Energy model at water sediment interface
- Extended energy model for sub layers
29. Sample window
Reflection coefficient versus depth is only valid when the
secondary reflection is not located within the first sampling
window.
The hypothesis here is that if the spectrum of the second
sample window is the same or less than the power spectrum
of the first sample window then we are at the same
sediment layer.
Sample window = 1 pulse width:
Very short sample window has low spectral resolution
which makes the comparison difficult.
Sample window = 4 pulse width:
Long sample window captures secondary reflections. The
power spectrum of the second sample is larger than the
first sample window.
30. Sample window
Large sample window unbalances the energy Proper sample window equal to twice the
model by over or underestimating the reflection transmitted pulse.
coefficients at deep layers.
31. Model comparison
Energy model (Simons et al)
- stacking is applied on the signal envelope.
- more sensible to propagation error.
- attenuation is applied on the nominal transmitted
frequency.
- reflections are highly correlated to the observed
energies.
Energy model (freq-domain)
- stacking is applied on the pressure signal
- no propagation error due to the nature of the
computational algorithm.
- attenuation is applied on the observed frequency band.
- reflections are less correlated to the observed
energies.
32.
33. Conclusion and recommendations
High frequency
• Calibration of the source level is important to increase the
certainty of the matching process.
• Alignment threshold is a crucial factor on classification result.
• Seabed classification using the echo shape is highly influenced
by the noise.
• Matching echo features can be tested for the classification such
as (amplitude, rise slope, pulse duration) .
•These features are expected to be less influenced by the noise
level.
34. Conclusion and recommendations
Low frequency
• Calibration of source level
• The energy model can be improved by accounting additional physical
phenomena such as signal interference and backscatters.
• Error propagation limits the low frequency classification.