System for classifying seafloor roughness

Claim 1

A system for classifying seafloor roughness using artificial neural network 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 self-organizing map preprocessor ( 20 ), said SOM preprocessor being attached through one or more Learning Vector Quantization variants ( 21 and 23 ) to a memory/display module ( 22 ).

Claim 2

A system for classifying seafloor roughness using artificial neural network 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 self-organizing map preprocessor ( 20 ), said SOM preprocessor being attached through one or more Learning Vector Quantization variants ( 21 and 23 ) to a memory/display module ( 22 ).

Claim 35

A system for seafloor classification using artificial neural network hybrid layout with the use of unprocessed multi-beam backscatter data, comprising an artificial neural network system that consists of a self-organizing map preprocessor [ 20 ], learning vector quantization variants LVQ2 [ 21 ] and LVQ2 [ 23 ], and a memory/display module [ 22 ], wherein the input to the said self-organizing map preprocessor [ 20 ] is derived from the output of an r.m.s. estimator module [ 19 ], which receives its input from an A/D converter [ 16 ], which in turn acquires its input from a beam-former [ 13 ] attached to two identical perpendicularly oriented arrays [ 7 ] and [ 8 ] of a multi-beam acoustic device that provide a signal that is rendered into analog format and filtered with the use of appropriate electronic hardware circuitry [ 14 ] and [ 15 ] respectively; the said self-organizing map preprocessor [ 20 ] receiving unprocessed multi-beam backscatter r.m.s data derived from different pre-formed beams as seafloor data, and classifying the seafloor data into various roughness types that are distinguished based on a 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 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 ]; 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 a the ground truth data set using an appropriate weight-updating criterion, with a correctly-representing 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 LVQ1 [ 21 ] being subsequently input to the memory and display module [ 22 ], wherein the results thereby displayed can be interpreted by a human to enhance the quality of classification; the system of the present invention being capable of incorporating an alternate hybrid layout wherein the LYQ1 [ 21 ] module can be replaced by another supervisable module LVQ2 [ 23 ] 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 layout having 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 tail portions of the weight distributions in the process of classification.

Claim 36

A system for seafloor classification using artificial neural network hybrid layout with the use of unprocessed multi-beam backscatter data, comprising an artificial neural network system that consists of a self-organizing map preprocessor [ 20 ], learning vector quantization variants LVQ2 [ 21 ] and LVQ2 [ 23 ], and a memory/display module [ 22 ], wherein the input to the said self-organizing map preprocessor [ 20 ] is derived from the output of an r.m.s. estimator module [ 19 ], which receives its input from an A/D converter [ 16 ], which in turn acquires its input from a beam-former [ 13 ] attached to two identical perpendicularly oriented arrays [ 7 ] and [ 8 ] of a multi-beam acoustic device that provide a signal that is rendered into analog format and filtered with the use of appropriate electronic hardware circuitry [ 14 ] and [ 15 ] respectively; the said self-organizing map preprocessor [ 20 ] receiving unprocessed multi-beam backscatter r.m.s data derived from different pre-formed beams as seafloor data, and classifying the seafloor data into various roughness types that are distinguished based on a 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 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 ]; 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 a the ground truth data set using an appropriate weight-updating criterion, with a correctly-representing 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 LVQ1 [ 21 ] being subsequently input to the memory and display module [ 22 ], wherein the results thereby displayed can be interpreted by a human to enhance the quality of classification; the system of the present invention being capable of incorporating an alternate hybrid layout wherein the LYQ1 [ 21 ] module can be replaced by another supervisable module LVQ2 [ 23 ] 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 layout having 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 tail portions of the weight distributions in the process of classification.