Title of Invention

A SYSTEM FOR CLASSIFYING SEAFLOOR ROUGHNESS USING ARTIFICIAL NEURAL NETWORK(ANN)

Abstract A system for classifying seafloor roughness using artificial neural network (ANN) hybrid layout from unprocessed multi-beam backscatter data, charactersied in that a means for generating unprocessed multi-beam backscatter r. m. s. data attached to the input of a self-organizing map (SOM) preprocessor (20), the said preprocessor being attached through one or more Learning Vector Quantization (LVQ) variants (21 and 23) to memory/display module (22).
Full Text The present invention relates to a system for classifying seafloor roughness using artificial neural network (ANN) hybrid layout from unprocessed multi-beam backscatter data. More particularly, the present invention relates to a system for online seafloor roughness classification from unprocessed multi-beam angular backscatter data using unsupervised learning as a pre-processor and supervised learning as the concluding block for improved classification, resulting in a highly efficient hybrid neural network layout to classify an unclassified dataset.
DETAILED DESCRIPTION OF THE INVENTION
Accordingly, the present invention provides a system for classifying seafloor roughness using artificial neural network (ANN) hybrid layout from unprocessed multi-beam backscatter data, charactersied in that a means for generating unprocessed multi-beam backscatter r. m. s. data attached to the input of a self-organizing map (SOM) preprocessor (20), the said preprocessor being attached through one or more Learning Vector Quantization (LVQ) variants (21 and 23) to memory/display module (22)..
In an embodiment of the present invention, the means for generating unprocessed multi-beam backscatter r.m.s. data comprises a multi-beam acoustic device mounted beneath a ship's hull and attached to an r.m.s. estimator module through a beam former module.
In another embodiment of the present invention, the multi-beam acoustic device comprises a linear array of transducers connected to a roll-pitch-heave sensor through cable connection boxes and an array of transmit-receive systems.
In yet another embodiment of the present invention, the multi-beam acoustic device comprises of two identical arrays of acoustic transducers mounted at right angles to each other.
In still another embodiment of the present invention, each array of the acoustic transducer is a combination of several sub-arrays and each sub array consists of multitude of elements.
In a further embodiment of the present invention, each element form a set of channels. In one more embodiment of the present invention, the arrays can be used either for transmission or for reception of signals.

The present invention relates to a system for classifying seafloor roughness using artificial neural network (ANN) hybrid layout from unprocessed multi-beam backscatter data. More particularly, the present invention relates to a system for online seafloor roughness classification from unprocessed multi-beam angular backscatter data using unsupervised learning as a pre-processor and supervised learning as the concluding block for improved classification, resulting in a highly efficient hybrid neural network layout to classify an unclassified dataset.
BACKGROUND ART
Hitherto known neural classifier for seafloor classification [Z. Michalopoulou, D. Alexandrou, and C. de Moustier, "Application of Neural and Statistical Classifiers to the Problem of Seafloor Characterization", IEEE Journal of Oceanic Engineering, Vol. 20, pp. 190-197 (1994)] describes a self-organizing map (SOM) network that is applied to multi-beam backscatter dataset. The drawback of this system is that it can use only processed data. Another drawback is its unsuitability for on-line application.
An alternate system [B. Chakraborty, R. Kaustubha, A. Hegde, A. Pereira, "Acoustic Seafloor Sediment Classification Using Self Organizing Feature Maps", IEEE Transactions Geoscience and Remote Sensing, Vol. 39, No. 12, pp. 2722-2725 (2001)] describes a SOM network wherein single-beam dataset is used for seafloor classification, and this system is more suited to online use. However, a limitation of this system is that it requires pre-processing of the time-series dataset prior to classification.
In US Patent Application No. 09/814,104 the Applicants have described a system which is incorporated in seafloor classification. This system described in this application estimates the seafloor acoustic backscattering strength with recorded root-mean-square (r.m.s) echo-voltage and the signal duration for each beam. In this system, multi-beam angular backscatter data have been acquired from the various seafloor areas around the Indian Ocean using a multi-beam acoustic system (Hydrosweep) installed onboard the Ocean Research Vessel Sagar Kanya. A drawback of the aforesaid system is that it requires large time-overhead to correct the raw data

data using unsupervised learning as a pre-processor and supervised learning as the concluding block for improved classification, resulting in a highly efficient hybrid neural network layout to classify an unclassified dataset.
DETAILED DESCRIPTION OF THE INVENTION
Accordingly, the present invention provides a system for classifying seafloor roughness using artificial neural network (ANN) hybrid layout from unprocessed multi-beam backscatter data, charactersied in that a means for generating unprocessed multi-beam backscatter r. m. s. data attached to the input of a self-organizing map (SOM) preprocessor (20), the said preprocessor being attached through one or more Learning Vector Quantization (LVQ) variants (21 and 23) to memory/display module (22)..
In an embodiment of the present invention, the means for generating unprocessed multi-beam backscatter r.m.s. data comprises a multi-beam acoustic device mounted beneath a ship's hull and attached to an r.m.s. estimator module through a beam former module.
In another embodiment of the present invention, the multi-beam acoustic device comprises a linear array of transducers connected to a roll-pitch-heave sensor through cable connection boxes and an array of transmit-receive systems.
In yet another embodiment of the present invention, the multi-beam acoustic device comprises of two identical arrays of acoustic transducers mounted at right angles to each other.
In still another embodiment of the present invention, each array of the acoustic transducer is a combination of several sub-arrays and each sub array consists of multitude of elements.
In a further embodiment of the present invention, each element form a set of channels. In one more embodiment of the present invention, the arrays can be used either for transmission or for reception of signals.
Fig. 4 presents a block schematic of the signal processing hardware used for beamforming
of the backscattered signal stream.
Fig. 5 represents the schematic block diagram of the hybrid ANN layout using SOM as the
preprocessor and L VQ I for improved classification.
Fig. 6 indicates the block schematic of an alternate hybrid ANN layout using SOM
network followed by LVQ2 for fine-tuning of cluster boundaries.
Fig. 7 shows the schematic block diagram of the optimum hybrid network layout of SOM,
LVQ1, and LVQ2 to achieve the best results of seafloor classification.
Fig. 8 illustrates the clustering of the unprocessed data vectors presented to the SOM
network to achieve improved classification results using the supervised learning features
ofLVQI and LVQ2 both independently and combined.
Table of seafloor classification results based on SOM, L VQ I, and LVQ2 layouts.
(Table Removed)
The present invention will now be described in detail with reference to the accompanying
drawings which are given for explaining the present invention in a more clear manner and
therefore should not be construed to limit the scope of the present invention in any
manner.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
Fig. I represents the schematic of a single-beam echo-sounder, wherein an acoustic
transducer [1] mounted beneath a survey ship [2] transmits acoustic pulses to the seafloor
in a direction perpendicular to the sea surface and receives the echo that is backscattered
from the seafloor. As the ship moves along a given preplanned track, a stream of
successive echoes backscattered from the consecutive segments on the seafloor are
received by an onboard receiver hardware and are recorded in its data memory module as
a time-series dataset. A drawback of this single-beam echosounder is that the seafloor area
coverage achievable in this manner is limited to the footprint of a single acoustic beam
that impinges on the seafloor.
Fig. 2 shows the schematic of a multi-beam echo-sounder [3] that is used for simultaneous
seafloor mapping over large areas. In this geometry, the transducers are spaced in such a
way that overlapping seafloor coverage is obtained. The fan-beam acoustic system
consists of a linear array of transducers mounted beneath the ship's

for range-related gain, seafloor slope correction, and insonification-depth
normalization.
Yet another system [B. Chakraborty, H.W. Schenke, V. Kodagali, and R. Hagen,
"Seabottom Characterization Using Multi-beam Echo-sounder: An Application of the
Composite Roughness Theory", IEEE Transactions Geoscience and Remote Sensing,
Vol. 38, pp.2419-2422 (2000)] describes a system for seafloor classification, wherein
it has been observed that the seafloor roughness parameters (power-law parameters)
are the ideal parameters for classification. The drawback of this system is that seafloor
classification can be implemented only after carrying out physical modeling of
composite roughness parameters.
OBJECTS OF THE INVENTION
The main object of the present invention is to provide a novel system for seafloor
classification using artificial neural network (ANN) hybrid layout with the use of
unprocessed multi-beam backscatter data.
Another object of the present invention is to provide a system for on-line (i.e., realtime)
seafloor classification using backscatter data after training the self-organized
mapping (SOM) network and learning vector quantization (LVQ) network.
Yet another object of the present invention is to provides a system that incorporates a
hybrid network using unsupervised SOM as the first block for coarse classification of
the seafloor backscatter data and supervised LVQ for highly improved performance in
the said classification.
Still another object of the present invention is to provide a system which incorporates
a combination of two variations of the LVQ layout to work together to achieve the
best classification results.
SUMMARY OF THE INVENTION
The present invention relates to a system for classifying seafloor roughness using
artificial neural network (ANN) hybrid layout from unprocessed multi-beam
backscatter data. More particularly, the present invention relates to a system for online
seafloor roughness classification from unprocessed multi-beam angular backscatter
data using unsupervised learning as a pre-processor and supervised learning as the
concluding block for improved classification, resulting in a highly efficient hybrid
neural network layout to classify an unclassified dataset.
DETAILED DESCRIPTION OF THE INVENTION
Accordingly, the present invention provides a system for classifying seafloor
roughness using artificial neural network (ANN) hybrid layout from unprocessed
multi-beam backscatter data, said system comprising a means for generating
unprocessed multi-beam backscatter r.m.s. data attached to the input of a selforganizing
map (SOM) preprocessor (20), said SOM preprocessor being attached
through one or more Learning Vector Quantization (LVQ) variants (21 and 23) to a
memory/display module (22).
In an embodiment of the present invention, the means for generating unprocessed
multi-beam backscatter r.m.s. data comprises a multi-beam acoustic device mounted
beneath a ship's hull and attached to an r.m.s. estimator module through a beam
former module.
In another embodiment of the present invention, the multi-beam acoustic device
comprises a linear array of transducers connected to a roll-pitch-heave sensor through
cable connection boxes and an array of transmit-receive systems.
In yet another embodiment of the present invention, the multi-beam acoustic device
comprises of two identical arrays of acoustic transducers mounted at right angles to
each other.
In still another embodiment of the present invention, each array of the acoustic
transducer is a combination of several sub-arrays and each sub array consists of
multitude of elements.
In a further embodiment of the present invention, each element form a set of channels.
In one more embodiment of the present invention, the arrays can be used either for
transmission or for reception of signals.
In one another embodiment of the present invention, the multi-beam acoustic device is
connected to the beam former module through a preamplifier and a time varying gain
adjustment circuit.
In an embodiment of the present invention, beam forming is accomplished using
appropriate delays.
In another embodiment of the present invention, the beam former module is connected
to the r.m.s. estimator module through a digital to analog converter, a filter, and a
analog to digital converter.
In still another embodiment of the present invention, a display means is optionally
connected to the analog to digital converter.
In yet another embodiment of the present invention, the display means is connected to
the analog to digital converter through a bottom-tracking gate.
In a further embodiment of the present invention, the output pattern of the r.m.s.
estimator module is the envelope of the r.m.s. signal amplitude Vs. beam number in
cross-track direction.
In one more embodiment of the present invention, self-organizing map (SOM)
preprocessor classifies the seafloor data into various roughness types and clusters
them.
In one another embodiment of the present invention, the roughness parameters are
distinguished based on the ship's cross-track angular multi-beam signal backscatter
shape parameter.
In an embodiment of the present invention, each cluster formed represents an unique
pattern of the input data.
In another embodiment of the present invention, the number of clusters thus formed is
equal to the number of differing patterns of received seafloor data set.
In yet another embodiment of the present invention, the clusters are formed with the
inherent unsupervised learning feature of the SOM preprocessor.
In still another embodiment of the present invention, the clusters are formed without
any prior knowledge of the number of the different types of input patterns.
In a further embodiment of the present invention, the Learning Vector Quantization
(LVQ) variant overcomes the imperfection in classification arising from the process of
unsupervised classification done by SOM preprocessor.
In one more embodiment of the present invention, the improvements seafloor
classification is achieved by supervised learning.
In one another embodiment of the present invention, the supervised learning is
imparted to the LVQ by a human interpreter based on the ground truth data set.
In an embodiment of the present invention, said system incorporates a LVQ designed
to avoid misclassification at the central portion of the weight-distribution of each
cluster or a LVQ designed to distinguish the overlapping tails of the weight
distribution of adjacent clusters or a combination of both.
In another embodiment of the present invention, the LVQ designed to avoid
misclassification at the central portion is based on "reward-punishment" criterion.
In yet another embodiment of the present invention, the LVQ moves the weight-vector
from the input if it is wrongly represented and the weight-vector is made to match the
input more closely if its correctly represented.
In still another embodiment of the present invention, the LVQ designed to distinguish
the overlapping of tails employs the technique of redistribution of the weights of the
overlapping portion of the adjacent clusters to the respective parent clusters.
In one more embodiment of the present invention, human interpreter can make use of
the results displayed on the display device to make further judgements on the quality
of classification.
The present invention more preferably provides a novel system for seafloor
classification using artificial neural network (ANN) hybrid layout with the use of
unprocessed multi-beam backscatter data, which comprises of an artificial neural
network system that consists of a self-organizing map (SOM) preprocessor [20],
learning vector quantization variants LVQ1 [21] and LVQ2 [23], and the
memory/display module [22], wherein the input to the said SOM network [20] is
derived from the output of an r.m.s. estimator module [19], which received its input
from an A/D converter [16], which in turn acquired its input from a beam-former [13]
attached to two identical perpendicularly oriented arrays [7] and [8] of a multi-beam
acoustic device of Fig. 3, the said signal after having been rendered into analog format
and filtered with the use of appropriate electronic hardware circuitry [14] and [15]
respectively; the said SOM network [20] receiving the unprocessed multi-beam
backscatter r.m.s data derived from the different pre-formed beams, and classifying
the seafloor data into various roughness types that are distinguished based on the
ship's cross-track angular multi-beam signal backscatter shape parameter, wherein a
unique cluster is formed to represent a specific pattern of the input data; the number of
clusters thus formed being equal to the number of differing patterns of the received
seafloor data set; the said clusters having been formed without any prior knowledge of
the number of different types of input patterns; the said clusters of data set being
subsequently input to the LVQ1 [21]; the imperfection in classification arising from
the process of unsupervised learning without any background knowledge being
partially overcome by the LVQ1 network [21] as in Fig. 5; the said improvement in
seafloor classification achievable by supervised learning based on the feedback
provided to the said LVQ1 network [21] by the human interpreter based on the ground
truth data set using an appropriate weight-updating criterion, with a correctlyrepresenting
weight-vector having been made to match the input more strongly, while
a wrongly-representing weight-vector having been moved away from the input, so as
to avoid misclassification at the central portion of the weight-distribution of each
cluster; the outputs of the said LVQ1 [21] being subsequently input to the memory
and display module [22], wherein the results thereby displayed having the utility for
human interpretation to be made to enhance the quality of classification; the system of
the present invention having the capability to incorporate an alternate hybrid layout
wherein the LVQ1 [21] module could be replaced by another supervisable module
LVQ2 [23] as in Fig. 6 that performs the alternate function of distinguishing the
overlapping tails of the weight distribution of adjacent clusters to implement seafloor
classification with minimal error; said system having additional capability to
incorporate an improved hybrid layout, the said layout having the advantage of
implementing seafloor classification based on the differing abilities of LVQ1 [21] and
LVQ2 [23] as in Fig. 7, thereby providing the best possible classification, taking into
account the central as well as tail portions of the weight distributions in the process of
classification.
In an embodiment of the present invention, the neural network layout is a modelindependent
system that would provides the capability to the use of unprocessed
backscatter data from the seafloor for the purpose of classification.
In another embodiment of the present invention, online seafloor classification is
possible subsequent to the training phase of the network, thereby providing a costeffective
system having the capability to circumvent the need for pre-processing of the
raw data.
In yet another embodiment of the present invention, a one-dimensional (i.e., bar-plot)
presentation of a multitude of self-organized clusters of unprocessed (i.e., raw) input
dataset and subsequent classification are provided to the human interpreter for further
judgment of the quality of classification, and additional capability of visualization of
the received input vectors in real time.
In still another embodiment of the present invention, said capabilities are extendable
to the processed backscatter data as well.
In a further embodiment of the present invention, said system is capable of real-time
redirecting cross-track multi-beam angular backscatter data received from the echosounder
installed onboard the ship/AUV, to a remote databank.
In one more embodiment of the present invention, said redirection is carried out by
representation of an extensive dataset by a few clustering units as formed in the said
system layout on a ping-by-ping basis.
NOVELTY AND INVENTIVE STEP
The novelty and inventive step of the present invention provides for an ingenious
system for seafloor classification using artificial neural network (ANN) hybrid layout
with the use of unprocessed multi-beam backscatter data, thereby circumventing the
need for the conventional laborious and time-consuming preprocessing task that
would have been required otherwise. The system of the present invention allows online
(i.e., real-time) seafloor classification using backscatter data after training the
self-organized mapping (SOM) network and learning vector quantization (LVQ)
network. Further, the system of the present invention has the unique capability for the
combined application of unsupervised SOM followed by supervised LVQ to achieve a
highly improved performance in the said classification, which is hitherto non-existent.
The system of the present invention has the additional capability for the use of a
combination of the two variations of the LVQ layout to work together to achieve the
best results in seafloor classification.
The novel system for seafloor classification using artificial neural network (ANN)
hybrid layout provides:
1. The capability of using unprocessed multi-beam backscatter data, thereby
circumventing the need for the conventional laborious and time-consuming
preprocessing task that would have been required otherwise.
2. The ability for on-line (i.e., real-time) seafloor classification after training the
self-organized mapping (SOM) network and learning vector quantization
(LVQ) network using a large time-series dataset in the absence of background
information.
3. The unique capacity for the combined use of unsupervised SOM followed by
supervised LVQ to achieve a highly improved performance in the said seafloor
classification, which is hitherto non-existent.
4. The means for the use of a combination of the two variants of the LVQ layout,
namely LVQ1 (that minimizes misclassification) and LVQ2 (that adjusts
overlapping weights of adjacent clusters), to operate together to achieve the
best results in seafloor classification.
BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS
In the drawings accompanying this specification:
Fig. 1 represents the schematic of a single-beam echo-sounder conventionally used for
seafloor mapping.
Fig. 2 shows the schematic of a multi-beam echo-sounder used for seafloor mapping
over a larger area with the use of an array of acoustic beams.
Fig. 3 indicates the geometry used in multi-beam echo-sounder for collection of
acoustic backscatter data from the seafloor.
Fig. 4 presents a block schematic of the signal processing hardware used for beamforming
of the backscattered signal stream.
Fig. 5 represents the schematic block diagram of the hybrid ANN layout using SOM
as the preprocessor and LVQ1 for improved classification.
Fig. 6 indicates the block schematic of an alternate hybrid ANN layout using SOM
network followed by LVQ2 for fine-tuning of cluster boundaries.
Fig. 7 shows the schematic block diagram of the optimum hybrid network layout of
SOM, LVQ1, and LVQ2 to achieve the best results of seafloor classification.
Fig, 8 illustrates the clustering of the unprocessed data vectors presented to the SOM
network to achieve improved classification results using the supervised learning
features of LVQ1 and LVQ2 both independently and combined.
Fig. 9 shows the Table of seafloor classification results based on SOM, LVQ1, and
LVQ2 layouts.
The present invention will now be described in detail with reference to the
accompanying drawings which are given for explaining the present invention in a
more clear manner and therefore should not be construed to limit the scope of the
present invention in any manner.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
Fig. 1 represents the schematic of a single-beam echo-sounder, wherein an acoustic
transducer [1] mounted beneath a survey ship [2] transmits acoustic pulses to the
seafloor in a direction perpendicular to the sea surface and receives the echo that is
backscattered from the seafloor. As the ship moves along a given preplanned track, a
stream of successive echoes backscattered from the consecutive segments on the
seafloor are received by an onboard receiver hardware and are recorded in its data
memory module as a time-series dataset. A drawback of this single-beam echosounder
is that the seafloor area coverage achievable in this manner is limited to the
footprint of a single acoustic beam that impinges on the seafloor.
Fig. 2 shows the schematic of a multi-beam echo-sounder [3] that is used for
simultaneous seafloor mapping over large areas. In this geometry, the transducers are
spaced in such a way that overlapping seafloor coverage is obtained. The fan-beam
acoustic system consists of a linear array of transducers mounted beneath the ship's
hull [4]. This fan-beam echo-sounder system consists essentially of two main subsystems,
namely, acoustic sub-system [5] and data processing sub-system [6]. The
acoustic sub-system comprises of a linear array of transducers installed at a certain
suitable angular separation, cable-connection boxes, an array of transmit-receive
systems for the respective transducer array, and usually a roll-pitch-heave sensor for
input to the online processing system. The data-processing control system employs a
shipboard computer, data-storage devices, and so forth. The multi-beam echo-sounder
is a device that provides higher resolution seabed mapping. This system is capable of
achieving a greater coverage as well as generating a high-precision map. The use of
multi-beam echo-sounding system results in the generation of seafloor profiles, which
are correlated with respect to the single central track, and allows more reliable
correlation of intersecting tracks.
Fig.3 represents a schematic of the beam configuration of the multi-beam echosounder
used as part of the novel system of the present invention for seafloor
classification using artificial neural network (ANN) hybrid layout with the use of
unprocessed multi-beam backscatter data, which consists primarily of two identical
arrays [7] and [8] of acoustic transducers mounted at right angles to each other. Each
array is a combination of several sub-arrays [7.1],...[7.n] and [8.1],....[8.n]
respectively, each consisting of a multitude of elements. These elements form a set of
channels [9.1], [9.2], ....[9.n], each channel consisting of a group of elements in
series. Similarly, for other counterpart transducer array, the receiving channels are
[10.1], [10.2],....[lO.n]. Because both the arrays are identical, they can be used for
either transmission or reception. In the receiving mode, signals from the different
channels are pre-amplified, corrected for attenuation, and thereafter beam-forming is
performed using appropriate delays. Altogether, a fixed number of pre-formed beams
(p.f.b) are formed within the 90° swath. The beams are equally spaced (-1.5°), average
beam-widths varying from 1.5° to 2.3° for deep-water surveys. The half-power beamwidth
becomes approximately double when the system is operated in shallow water
mode. The beam-formed outputs are tapped at the bottom-echo module to estimate the
depth corresponding to the beams from different directions. In order to achieve
uniform insonification over a swath of 90°, the system is designed to operate in a
swath mode.
Fig. 4 illustrates the block schematic of the signal processing hardware used for beamforming
of the backscattered signal stream, which consists primarily of a pre-amplifier
[11], time-varying gain adjustment circuitry [12], beam-former [13], digital-to-analog
converter [14], filter [15], analog-to-digital (A/D) converter [16], bottom tracking gate
[17], and display [18]. The digital data used for seafloor classification are derived
from the output channel of the A/D converter [16]. This signal is input to an r.m.s.
estimator module [19], whose outputs are used as the input vectors for further
processing.
Fig. 5 represents the schematic block diagram of the hybrid ANN layout used in the
system of the present invention. This layout consists of a self-organizing map (SOM)
network [20], learning vector quantization variant 1 (LVQ1) [21], and the
memory/display module [22]. The input to the said SOM network [20] is derived from
the output of the r.m.s. estimator module [19]. In the system of the present invention,
the said SOM network [20] receives the unprocessed multi-beam backscatter r.m.s.
data (whose pattern is the envelope of the r.m.s. signal amplitude versus beam number
in the cross-track direction) obtained from the different pre-formed beams, and
classifies the seafloor data into various roughness types. The different roughness types
are distinguished based on the ship's cross-track angular multi-beam signal
backscatter shape parameter. Each cluster so formed represents a unique pattern of the
input data. Accordingly, the number of clusters formed is equal to the number of
differing patterns of the received seafloor data set. These clusters are formed with the
inherent unsupervised learning feature of the SOM network without any prior
knowledge of the number of different types of input patterns. The clusters of the data
set so formed by the SOM [20] are input to the LVQ1 [21]. The imperfection in
classification, which is inherent in the said SOM network [20] due to the nature of
unsupervised learning without any background knowledge (of the input data types), is
partially overcome by the LVQ1 network [21]. This improvement in seafloor
classification results from the supervised learning based on the feedback provided to
the said LVQ1 network [21] by the human interpreter, judging from the ground truth
data set. The said improvement is based on the "reward-punishment" criterion,
wherein a correctly representing weight-vector is made to match the input more
strongly, while a wrongly representing weight-vector is moved away from the input,
so that it will avoid misclassification in future comparisons. The said LVQ1 [21]
works more efficiently at the central portion of the weight distribution of each cluster
formed in the SOM [20] than at the tail portion of the said distribution because of the
inability of the said LVQ1 [21] to distinguish the overlapping tails of adjacent
clusters. The outputs of the said LVQ1 [21] are subsequently input to the memory and
display module [22], wherein the human interpreter can make use of the results that
are thereby displayed to make further judgments on the quality of classification.
Fig. 6 shows the schematic block diagram of an alternate hybrid ANN layout used in
the system of the present invention together with the layout shown in Fig. 5. This
layout is similar to the one given in Fig. 5, the only difference being that while the
LVQ1 [21] avoids misclassification to a greater extent at the central portion of the
weight distribution of each cluster formed in SOM [20] than at the tail portion of the
said distribution, the LVQ2 network focuses on minimizing misclassification (arising
in SOM) as a result of its ability to distinguish the overlapping tails of the weight
distribution of adjacent clusters. The said LVQ2 [23] employs a technique for
redistribution of the weights of the overlapping portions of the adjacent clusters to the
respective parent clusters. The outputs of the said LVQ2 [23] are subsequently input
to the memory and display module [22], wherein the results are displayed and stored
to enable the human interpreter to make further judgments on the quality of
classification.
Fig. 7 illustrates the schematic block diagram of an improved hybrid ANN layout used
in the system of the present invention together with the layout shown in Figs. 5 & 6.
This layout has the advantage of implementing seafloor classification based on the
differing abilities of LVQ1 [21] and LVQ2 [23], thereby providing the best possible
classification, taking into account the central as well as the tail portions of the weight
distributions in the process of classification.
Fig. 8 illustrates the clustering of the unprocessed data vectors presented to the
aforesaid SOM network [20] and the supervised learning features of LVQ1 [21] and
LVQ2 [23], both independently and combined. In this illustration, two weight-vectors
located on either side of the boundary (shown as thick vertical line) have been
designated as lying within the overlapping region of adjacent clusters. In the
illustration of Fig. 8, the segments [24.1], [24.2], and [24.3] represent three
consecutive weight-distribution-patterns [24] formed by the unsupervised SOM [20],
wherein the symbols, (|), and (/) indicate three different classes of seafloor
roughness parameters. As a representative example, the weight-distribution-pattern
[25] at the output of the supervised architecture LVQ1 [21] clearly indicates the
removal of the central misclassified portions of [24.1] and [24.2], and their
redeployment into the parent cluster represented by the segment [24.3], thereby
providing a better redistribution in segments [25.1], [25.2], and [25.3] of the full
weight-distribution-pattern [25]. The weight-distribution-pattern [26] illustrates the
effectiveness of the supervised layout LVQ2 [23] wherein the segments [26.1], [26.2],
and [26.3] indicate the removal of the misclassified data vector (/) from the tail
portions of [24.1] and [24.3] of the SOM-output [24] and their redeployment into its
parent segment [24.2] as shown in [26], while ignoring the classification error in the
central portions of the weight-distribution-pattern in [26.1] indicated by (•). The
weight-distribution-pattern [27] illustrates the effectiveness of classification with the
combined use of LVQ1 [21] and LVQ2 [23] wherein the misclassifications found in
both the central and tail portions of [24] are corrected to provide a fully corrected
weight-distribution-pattern [27].
Fig.9 shows a Table of roughness parameters and classification results (in percentage)
for three seafloor areas examined using LVQ1, LVQ2 and a combination of both the
above supervised layouts, using SOM as the preprocessor in each case.
Generally, in the Hydrosweep multibeam echosounding system, the two outer beam
values on the port side and the starboard side are erroneous and are neglected.
Consequently, the original 59 beam input vector becomes an array of 55 beam values
from angle -45° to +45°. In this embodiment, each of the 55 beam values for 5 input
vectors at a time is averaged for input vector smoothening. By reducing variations in
input patterns from the same seafloor region, we obtain marginally improved
classification performance albeit at the cost of computational time.
The preprocessor block uses unsupervised learning to form a self- organised mapping
of 100 input vectors from various seafloors. The input data vectors are converted from
the decibel scale to their natural values. Each vector is then normalised to match the
range of the weights. The data is still unprocessed in the sense that no preprocessing
algorithm (PROBASI) has been applied to it. From each seafloor area, a representative
backscatter data set is chosen for training the SOM network. In this architecture,
neighbouring entities called neurons compete with each other by the mutual lateral
interaction of their weights. The weights in the neighbourhood of the closest matching
neuron to the input vector presented are updated from an initial random distribution,
so as to bring them closer to this input vector. This leads to the formation of local
neighbourhoods, depicting a particular class. The learning rate is an exponentially
decaying function 0.5/(t02) where / is the number of iterations, thus causing the rate of
weight updating to be initially high and subsequently decreases gradually for finer
tuning in successive iterations, until ultimately only one neuron excites in response to
an input presented. At the end of execution of the SOM algorithm, the input space
arranges into coarse clusters, each representing a unique input pattern, without any a
priori information. Future input vectors presented in the testing phase are assigned to
a cluster in the output space that best matches the input vector. The power law
parameters (gamma and beta) for the different seafloor regions, obtained by physical
modeling, are used to correlate the obtained results.
The trained weight matrix after employing the SOM network are used as the inputs for
a supervised learning network, namely LVQ. Here, training and testing is carried out
simultaneously for a few iterations until the network weights belonging to each cluster
are truly representative of that cluster. The learning rate is still exponential but with a
much smaller initial value to allow fine tuning of neuron weights. In LVQ1, with user
control for 3-4 input vectors from each class, the erroneously classifying weight vector
is moved away from the input. If classification is correct, the weight vector is
strengthened towards the input and eventually misclassification is avoided in further
testing. LVQ2 goes a step further in overcoming the overlapping weight biases in
regions of crossover of adjacent clusters by applying the window function at their
boundary for correction to 1-2 neurons. The window is divided into half and
corrections are made only when the excited neuron falls on the wrong side of the midplane
by moving that vector away from the input and simultaneously bringing the
neuron on the correct half of the window closer to it. A plot is made of probability
density function (pdf) of excited neuron versus the frequency of excitation.
In particular, we used three seafloor regions A, B and C for our comparative study.
The classification results are tabulated in Fig. 9 for SOM, LVQ1 and LVQ2 for
unprocessed backscatter data.
On comparing the tabulated power law parameters (gamma and beta), acting as an
indicator of seafloor interface roughness, with self-organized clusters after training by
SOM, we have notice that areas A and C are of the same seafloor type. Between the
two, classification results have indicated enhanced performance with LVQ for area C,
which is rougher as indicated by its higher Chra value (rms roughness between two
points separated by a distance of 100 m). The most significant contribution of this
invention is the ability of the network to give superior response to unprocessed data,
thereby eliminating the need for tedious preprocessing of the multi-beam backscatter
data for seafloor roughness theory, thereby providing online indication of roughness
even using raw data.
ADVANTAGES OF THE PRESENT INVENTION
The main advantages of the present invention are:
1. It provides the capability for combined use of the two variants of the learning vector
quantization (LVQ) network to achieve the best classification of the seafloor
characteristics, which is a capability that is hitherto non-existent.
2. It provides a self-organization of multi-beam input data vectors into coarse
clusters in the output space without any a priori information.
3. It provides the capability to the use of raw dataset as input vectors to the
classification network.
4. It reduces computational time overhead, which is a capability that is not
available with hitherto known physical models.
5. It provides a means for online seafloor classification using raw backscatter
data subsequent to suitable training of the network.
6. It provides for high network performance by its ability for the successive
application of a duality of means for avoidance of misclassification and the
capability for recovery of overlapped information.
7. It allows for the capability for easy and real-time transmission of remote data
received by the echo-sounder onboard the ship/autonomous underwater
vehicle (AUV), to the remote databank; the said transmission having been
carried out by the representation of an entire dataset of cross-track multibeam
angular backscatter information by a few clustering units formed in the
said system on a ping-by-ping basis.


WE CLAIMS:
1 .A system for classifying seafloor roughness using artificial neural network (ANN) hybrid layout from unprocessed multi-beam backscatter data, charactersied in that a means for generating unprocessed multi-beam backscatter r. m. s. data attached to the input of a self-organizing map (SOM) preprocessor (20), the said preprocessor being attached through one or more Learning Vector Quantization (L VQ) variants (21 and 23) to memory/display module (22).
2. A system as claimed in claim 1, wherein the means for generating unprocessed multi-beam backscatter r. m. s. data comprises a multi-beam acoustic device mounted beneath a ship's hull(4) and attached to an r. m. s. estimator module (19) through a beam former module(13).
3. A system as claimed in claim 2, wherein the multi-beam acoustic device comprises a linear array of transducers(l) connected to a roll-pitch-heave sensor through cable connection boxes and an array of transmit-receive systems.
4. A system as claimed in claim 3, wherein the multi-beam acoustic device comprises of two identical arrays of acoustic transducers(l) mounted at right angles to each other.
5. A system as claimed in claim 4, wherein each array of the acoustic transducer( 1) is a combination of several sub-arrays and each sub array consists of multitude of elements.
6. A system as claimed in claim 5, wherein each element form a set of channels.
7. A system as claimed in claim 4, wherein the arrays used is either for transmission or for reception of signals.
8. A system as claimed in claim 2, wherein the multi-beam acoustic device is connected to the beam former module through a preamplifier and a time varying gain adjustment circuit.
9. A system as claimed in claim 2, wherein beam forming is accomplished using appropriate delays.

10. A system as claimed in claim 2, wherein the beam former module is connected to the r. m. s. estimator module through a digital to analog converter, a filter, and a analog to digital converter.
11. A system as claimed in claim 10, wherein a display means is optionally connected to the analog to digital converter.
12. A system as claimed in claim 11, wherein the display means is connected to the analog to digital converter through a bottom-tracking gate.

13. A system as claimed in claim 1, wherein the output pattern of the r. m. s. estimator module is the envelope of the r. m. s. signal amplitude Vs. beam number in cross- track direction.
14. A system as claimed in claim 1, wherein self-organizing map (SOM) preprocessor classifies the seafloor data into various roughness types and clusters them.
15. A system as claimed in claim 14, wherein the roughness parameters are distinguished based on the ship's cross-track angular multi-beam signal backscatter shape parameter.
16. A system as claimed in claim 14, wherein each cluster formed represents an unique pattern of the input data.
17. A system as claimed in claim 14, wherein the number of clusters thus formed is equal to the number of differing patterns of received seafloor data set.
18. A system as claimed in claim 14, wherein the clusters are formed with the inherent unsupervised learning feature of the SOM preprocessor.
19. A system as claimed in claim 14, wherein the clusters are formed without any prior knowledge of the number of the different types of input patterns.
20. A system as claimed in claiml, wherein the Learning Vector Quantization (LVQ) variant overcomes the imperfection in classification arising from the process of unsupervised classification done by SOM preprocessor.
21. A system as claimed in claim 20, wherein the improvements seafloor classification is achieved by supervised learning.
22. A system as claimed in claim 21, wherein the supervised learning is imparted to the LVQ by a human interpreter based on the ground truth data set.
23. A system as claimed in claim 1, wherein said system incorporates a LVQ designed to avoid misclassification at the central portion of the weight-distribution of each cluster or a LVQ designed to distinguish the overlapping tails of the weight distribution of adjacent clusters or a combination of both.
24. A system as claimed in claim 23, wherein the LVQ designed to avoid misclassification at the central portion is basedon"reward-punishment"criterion.
25. A system as claimed in claim 24, wherein the LVQ moves the weight-vector from the input if it is wrongly represented and the weight-vector is made to match the input more closely if its correctly represented.

26. A system as claimed in claim 23, wherein the LVQ designed to distinguish the overlapping of tails employs the technique of redistribution of the weights of the overlapping portion of the adjacent clusters to the respective parent clusters.
27. A stem as claimed in claim 1, wherein a human interpreter may use of the results-displayed on the display device to make further judgements on the quality of classification.
28. A system as claimed in claim 1, wherein the neural network layout is a model-independent system that would provide the capability to the use of unprocessed backscatter data from the seafloor for the purpose of classification.
29A system as claimed in claim 1, wherein online seafloor classification is possible subsequent to the training phase of the network, thereby providing a cost-effective system having the capability to circumvent the need for pre-processing of the raw data.
30 A system as claimed in claim 1, wherein a one-dimensional (i. e. bar-plot) presentation of a multitude of self-organized clusters of unprocessed (i. e. raw) input dataset and subsequent classification are provided to the human interpreter for further judgment of the quality of classification, and additional capability of visualization of the received input vectors in real time.
31 A system as claimed in claim 1, wherein said capabilities are extendable to the processed backscatter data as well.
32 A system as claimed in claim 1, wherein said system is capable of real-time redirecting cross-track multi-beam angular backscatter data received from the echo-sounder installed onboard the ship/AUV, to a remote databank.
33 A system as claimed in claim 32, wherein said redirection is carried out by representation of an extensive data set by a few clustering units as formed in the said system layout on a ping-by-ping basis.
34. A system for classifying seafloor roughness using artificial neural network (ANN) hybrid layout from unprocessed multi-beam backscatter data, substantially as herein described with reference to the examples and figs 1 to 8 of the drawings accompanying this specification.

Documents:

2132-DELNP-2004-Abstract-(11-08-2008).pdf

2132-delnp-2004-abstract.pdf

2132-DELNP-2004-Claims-(11-08-2008).pdf

2132-delnp-2004-claims.pdf

2132-delnp-2004-complete specification (granded)-(11-08-2008).pdf

2132-DELNP-2004-Correspondence-Others-(11-08-2008).pdf

2132-delnp-2004-correspondence-others.pdf

2132-delnp-2004-description (complete)-11-08-2008.pdf

2132-delnp-2004-description (complete).pdf

2132-DELNP-2004-Drawings-(11-08-2008).pdf

2132-delnp-2004-drawings.pdf

2132-delnp-2004-form-1.pdf

2132-delnp-2004-form-13.pdf

2132-delnp-2004-form-18.pdf

2132-delnp-2004-form-2-(12-09-2008).pdf

2132-delnp-2004-form-2.pdf

2132-delnp-2004-form-3.pdf

2132-delnp-2004-form-5.pdf


Patent Number 223606
Indian Patent Application Number 2132/DELNP/2004
PG Journal Number 40/2008
Publication Date 03-Oct-2008
Grant Date 17-Sep-2008
Date of Filing 22-Jul-2004
Name of Patentee COUNCIL OF SCIENTIFIC AND INDUSTRIAL RESEARCH
Applicant Address RAFI MARG, NEW DELHI-110 001, INDIA.
Inventors:
# Inventor's Name Inventor's Address
1 CHAKRABORTY BISHWAJIT NIO, INDIA.
2 KODAGALI, VIJAY NIO, INDIA.
3 BARACHO,JENNIFER NIO, INDIA.
4 JOSEPH, ANTHONY NIO, INDIA.
PCT International Classification Number G01V 1/38
PCT International Application Number PCT/IB02/01073
PCT International Filing date 2002-03-25
PCT Conventions:
# PCT Application Number Date of Convention Priority Country
1 10/106,706 2002-03-25 U.S.A.