A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice
Identifieur interne : 001643 ( Istex/Corpus ); précédent : 001642; suivant : 001644A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice
Auteurs : Ute Christina Herzfeld ; Oliver ZahnerSource :
- Computers and Geosciences [ 0098-3004 ] ; 2001.
Abstract
A combination of geostatistical methods for data reduction and the neural-network approach for association of input information to object classes forms the basic idea for a connectionist-geostatistical approach to automated image classification. The investigation of the posed glaciological question — understanding a surging glacier —, which presented itself as a natural catastrophe, leaving the onlooking geophysicist without ‘real’ data and only video coverage for analysis, calls for new analysis methods. Deformation states of the rapidly changing ice surface of a surging glacier manifest themselves in crevasse patterns, which were observed in GPS-referenced video images for Bering Glacier, Alaska. It can be shown that directional variograms calculated from video data at appropriate scaling can be used to characterize crevasse patterns. Directional variograms also provide a suitable filter to reduce the data volume necessary for a successful training of a neural network. The final network, a multilayer feedforward perceptron with backpropagation of errors, has a very high success rate for classification of (unknown) crevasse patterns of surging ice from video images.
Url:
DOI: 10.1016/S0098-3004(00)00089-3
Links to Exploration step
ISTEX:F58F80E51A6A3AE890E6AF47A46FD9E51A790B00Le document en format XML
<record><TEI wicri:istexFullTextTei="biblStruct"><teiHeader><fileDesc><titleStmt><title>A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice</title><author><name sortKey="Christina Herzfeld, Ute" sort="Christina Herzfeld, Ute" uniqKey="Christina Herzfeld U" first="Ute" last="Christina Herzfeld">Ute Christina Herzfeld</name><affiliation><mods:affiliation>E-mail: ute@mpl.ucsd.edu</mods:affiliation></affiliation><affiliation><mods:affiliation>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</mods:affiliation></affiliation></author><author><name sortKey="Zahner, Oliver" sort="Zahner, Oliver" uniqKey="Zahner O" first="Oliver" last="Zahner">Oliver Zahner</name><affiliation><mods:affiliation>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</mods:affiliation></affiliation></author></titleStmt><publicationStmt><idno type="wicri:source">ISTEX</idno><idno type="RBID">ISTEX:F58F80E51A6A3AE890E6AF47A46FD9E51A790B00</idno><date when="2001" year="2001">2001</date><idno type="doi">10.1016/S0098-3004(00)00089-3</idno><idno type="url">https://api.istex.fr/document/F58F80E51A6A3AE890E6AF47A46FD9E51A790B00/fulltext/pdf</idno><idno type="wicri:Area/Istex/Corpus">001643</idno><idno type="wicri:explorRef" wicri:stream="Istex" wicri:step="Corpus" wicri:corpus="ISTEX">001643</idno></publicationStmt><sourceDesc><biblStruct><analytic><title level="a">A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice</title><author><name sortKey="Christina Herzfeld, Ute" sort="Christina Herzfeld, Ute" uniqKey="Christina Herzfeld U" first="Ute" last="Christina Herzfeld">Ute Christina Herzfeld</name><affiliation><mods:affiliation>E-mail: ute@mpl.ucsd.edu</mods:affiliation></affiliation><affiliation><mods:affiliation>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</mods:affiliation></affiliation></author><author><name sortKey="Zahner, Oliver" sort="Zahner, Oliver" uniqKey="Zahner O" first="Oliver" last="Zahner">Oliver Zahner</name><affiliation><mods:affiliation>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</mods:affiliation></affiliation></author></analytic><monogr></monogr><series><title level="j">Computers and Geosciences</title><title level="j" type="abbrev">CAGEO</title><idno type="ISSN">0098-3004</idno><imprint><publisher>ELSEVIER</publisher><date type="published" when="2001">2001</date><biblScope unit="volume">27</biblScope><biblScope unit="issue">5</biblScope><biblScope unit="page" from="499">499</biblScope><biblScope unit="page" to="512">512</biblScope></imprint><idno type="ISSN">0098-3004</idno></series><idno type="istex">F58F80E51A6A3AE890E6AF47A46FD9E51A790B00</idno><idno type="DOI">10.1016/S0098-3004(00)00089-3</idno><idno type="PII">S0098-3004(00)00089-3</idno></biblStruct></sourceDesc><seriesStmt><idno type="ISSN">0098-3004</idno></seriesStmt></fileDesc><profileDesc><textClass></textClass><langUsage><language ident="en">en</language></langUsage></profileDesc></teiHeader><front><div type="abstract" xml:lang="en">A combination of geostatistical methods for data reduction and the neural-network approach for association of input information to object classes forms the basic idea for a connectionist-geostatistical approach to automated image classification. The investigation of the posed glaciological question — understanding a surging glacier —, which presented itself as a natural catastrophe, leaving the onlooking geophysicist without ‘real’ data and only video coverage for analysis, calls for new analysis methods. Deformation states of the rapidly changing ice surface of a surging glacier manifest themselves in crevasse patterns, which were observed in GPS-referenced video images for Bering Glacier, Alaska. It can be shown that directional variograms calculated from video data at appropriate scaling can be used to characterize crevasse patterns. Directional variograms also provide a suitable filter to reduce the data volume necessary for a successful training of a neural network. The final network, a multilayer feedforward perceptron with backpropagation of errors, has a very high success rate for classification of (unknown) crevasse patterns of surging ice from video images.</div></front></TEI><istex><corpusName>elsevier</corpusName><author><json:item><name>Ute Christina Herzfeld</name><affiliations><json:string>E-mail: ute@mpl.ucsd.edu</json:string><json:string>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</json:string></affiliations></json:item><json:item><name>Oliver Zahner</name><affiliations><json:string>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</json:string></affiliations></json:item></author><subject><json:item><lang><json:string>eng</json:string></lang><value>Neural network</value></json:item><json:item><lang><json:string>eng</json:string></lang><value>Variogram</value></json:item><json:item><lang><json:string>eng</json:string></lang><value>Videography</value></json:item><json:item><lang><json:string>eng</json:string></lang><value>Glaciology</value></json:item><json:item><lang><json:string>eng</json:string></lang><value>Surge</value></json:item><json:item><lang><json:string>eng</json:string></lang><value>Glaciology</value></json:item></subject><language><json:string>eng</json:string></language><originalGenre><json:string>Full-length article</json:string></originalGenre><abstract>A combination of geostatistical methods for data reduction and the neural-network approach for association of input information to object classes forms the basic idea for a connectionist-geostatistical approach to automated image classification. The investigation of the posed glaciological question — understanding a surging glacier —, which presented itself as a natural catastrophe, leaving the onlooking geophysicist without ‘real’ data and only video coverage for analysis, calls for new analysis methods. Deformation states of the rapidly changing ice surface of a surging glacier manifest themselves in crevasse patterns, which were observed in GPS-referenced video images for Bering Glacier, Alaska. It can be shown that directional variograms calculated from video data at appropriate scaling can be used to characterize crevasse patterns. Directional variograms also provide a suitable filter to reduce the data volume necessary for a successful training of a neural network. The final network, a multilayer feedforward perceptron with backpropagation of errors, has a very high success rate for classification of (unknown) crevasse patterns of surging ice from video images.</abstract><qualityIndicators><score>7.028</score><pdfVersion>1.2</pdfVersion><pdfPageSize>544 x 743 pts</pdfPageSize><refBibsNative>true</refBibsNative><keywordCount>6</keywordCount><abstractCharCount>1185</abstractCharCount><pdfWordCount>6426</pdfWordCount><pdfCharCount>39680</pdfCharCount><pdfPageCount>14</pdfPageCount><abstractWordCount>169</abstractWordCount></qualityIndicators><title>A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice</title><pii><json:string>S0098-3004(00)00089-3</json:string></pii><refBibs><json:item><host><author></author></host></json:item><json:item><host><author></author></host></json:item><json:item><host><author></author></host></json:item><json:item><host><author></author></host></json:item><json:item><author><json:item><name>U.C. Herzfeld</name></json:item><json:item><name>C.A. Higginson</name></json:item></author><host><volume>22</volume><pages><last>52</last><first>35</first></pages><issue>1</issue><author></author><title>Computers & Geosciences</title></host><title>Automated geostatistical seafloor classification — principles, parameters, feature vectors, and discrimination criteria</title></json:item><json:item><author><json:item><name>U.C. Herzfeld</name></json:item><json:item><name>H. Mayer</name></json:item></author><host><volume>43</volume><pages><last>434</last><first>427</first></pages><issue>145</issue><author></author><title>Journal of Glaciology</title></host><title>Surge of Bering Glacier and Bagley Ice Field, Alaska: an update to August 1995 and an interpretation of brittle-deformation patterns</title></json:item><json:item><author><json:item><name>U.C. Herzfeld</name></json:item><json:item><name>M. Stauber</name></json:item><json:item><name>N. Stahl</name></json:item></author><host><volume>30</volume><pages><last>234</last><first>224</first></pages><author></author><title>Annals of Glaciology</title></host><title>Geostatistical characterization of ice surfaces from ERS-2 SAR data of Jakobshavns Isbræ, Greenland</title></json:item><json:item><author><json:item><name>K. Hornik</name></json:item><json:item><name>M. Stinchcombe</name></json:item><json:item><name>H. White</name></json:item></author><host><volume>2</volume><pages><last>366</last><first>359</first></pages><issue>5</issue><author></author><title>Neural Networks</title></host><title>Multilayer feedforward networks are universal approximators</title></json:item><json:item><host><author></author></host></json:item><json:item><author><json:item><name>W.B. Kamb</name></json:item><json:item><name>C.F. Raymond</name></json:item><json:item><name>W.D. Harrison</name></json:item><json:item><name>H. Engelhardt</name></json:item><json:item><name>K.A. Echelmeyer</name></json:item><json:item><name>N. Humphrey</name></json:item><json:item><name>M.M. Brugman</name></json:item><json:item><name>T. Pfeffer</name></json:item></author><host><volume>227</volume><pages><last>479</last><first>469</first></pages><issue>4686</issue><author></author><title>Science</title></host><title>Glacier surge mechanism: 1982–1983 surge of Variegated Glacier, Alaska</title></json:item><json:item><host><author></author></host></json:item><json:item><host><author></author></host></json:item><json:item><host><author></author><title>Learning vector quantization for pattern recognition. Technical Report TKK-F-A601</title></host></json:item><json:item><host><author></author></host></json:item><json:item><host><author></author></host></json:item><json:item><author><json:item><name>G. Matheron</name></json:item></author><host><volume>58</volume><pages><last>1266</last><first>1246</first></pages><issue>8</issue><author></author><title>Economic Geology</title></host><title>Principles of geostatistics</title></json:item><json:item><author><json:item><name>H. Mayer</name></json:item><json:item><name>U.C. Herzfeld</name></json:item></author><host><volume>30</volume><pages><last>249</last><first>243</first></pages><author></author><title>Annals of Glaciology</title></host><title>Structural glaciology of the fast-moving Jakobshavns Isbræ, Greenland, compared to the surging Bering Glacier, Alaska</title></json:item><json:item><host><author></author></host></json:item><json:item><host><author></author></host></json:item><json:item><host><author></author></host></json:item><json:item><host><author></author></host></json:item></refBibs><genre><json:string>research-article</json:string></genre><host><volume>27</volume><pii><json:string>S0098-3004(00)X0064-7</json:string></pii><pages><last>512</last><first>499</first></pages><issn><json:string>0098-3004</json:string></issn><issue>5</issue><genre><json:string>journal</json:string></genre><language><json:string>unknown</json:string></language><title>Computers and Geosciences</title><publicationDate>2001</publicationDate></host><categories><wos><json:string>science</json:string><json:string>geosciences, multidisciplinary</json:string><json:string>computer science, interdisciplinary applications</json:string></wos><scienceMetrix><json:string>natural sciences</json:string><json:string>earth & environmental sciences</json:string><json:string>geochemistry & geophysics</json:string></scienceMetrix></categories><publicationDate>2001</publicationDate><copyrightDate>2001</copyrightDate><doi><json:string>10.1016/S0098-3004(00)00089-3</json:string></doi><id>F58F80E51A6A3AE890E6AF47A46FD9E51A790B00</id><score>0.8191762</score><fulltext><json:item><extension>pdf</extension><original>true</original><mimetype>application/pdf</mimetype><uri>https://api.istex.fr/document/F58F80E51A6A3AE890E6AF47A46FD9E51A790B00/fulltext/pdf</uri></json:item><json:item><extension>zip</extension><original>false</original><mimetype>application/zip</mimetype><uri>https://api.istex.fr/document/F58F80E51A6A3AE890E6AF47A46FD9E51A790B00/fulltext/zip</uri></json:item><istex:fulltextTEI uri="https://api.istex.fr/document/F58F80E51A6A3AE890E6AF47A46FD9E51A790B00/fulltext/tei"><teiHeader><fileDesc><titleStmt><title level="a">A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice</title></titleStmt><publicationStmt><authority>ISTEX</authority><publisher>ELSEVIER</publisher><availability><p>©2001 Elsevier Science Ltd</p></availability><date>2001</date></publicationStmt><notesStmt><note type="content">Fig. 1: Map of Bering Glacier, Alaska, with tracks of video flights.</note><note type="content">Fig. 2: Example of (A) original (nontransformed) video image and (B) vertically mirrored image of (A) with three-directional variograms.</note><note type="content">Fig. 3: Classes of crevasse types in a surging glacier. Examples from video images of Bering Glacier, Alaska.</note><note type="content">Fig. 4: Individual video frame with the two 150×150 pixel subsets that were used in classification.</note><note type="content">Fig. 5: Idealized learning curve and validation curve for neural network. Lower curve — learning curve. Upper curve — validation curve.</note><note type="content">Fig. 6: Typical learning and validation curves encountered in modeling and training a neural network. Lower curves — learning curves. Upper curves — validation curves. (A) Nonconvergent learning process, resultant from training net with two hidden layers. (B) Training a net with too few (10) hidden layer neurons only leaves overall too high error in both learning and validation curves (see Fig. 7). (C) Poor generalization after overfitting with too many (42) intermediate layer neurons (see Fig. 7). (D) Unstable fit caused by too high learning rate of 0.4. (E) Good learning behavior (cf. idealized curve, Fig. 5). Final net topology. (F) Nonconvergent, heavily oscillating learning with high errors, resultant from application of quickpropagation algorithm.</note><note type="content">Fig. 7: Determination of optimal number of intermediate layer neurons. Training versus validation error. 25 to 40 neurons are best.</note><note type="content">Fig. 8: Topology of MLP-BP with 42 neuron input layer, one intermediate layer with 32 neurons, and 6 neuron output layer, layerwise fully connected (1536 connections, no shortcut connections). NET 1.</note><note type="content">Fig. 9: Topology of MLP-BP with 42 neuron input layer, one intermediate layer with 28 neurons, and 9 neuron output layer, fully connected with shortcut connections (1606 connections). NET 2.</note><note type="content">Fig. 10: Classification results and generalization capability of MLP-BP (42-28-9) (NET 2). Examples from test set 2. (A) Test image 14. Bidirectional pattern (class 4) is clearly recognized. (B) Test image 2. En-échelon type (class 8) classified correctly. (C) Test image 10. Chaos (class 1) classified correctly. (D) Test image 11. Class 0 (undisturbed ice surface, moraines, rock) recognized with weight 1 despite small disturbance in image. (E) Test image 19. Class 0 (undisturbed ice surface, moraines, rock) recognized correctly despite large disturbance. (F) Test image 4, showing new pattern, is associated to class 2 (parallel crevasses), which is most closely related. (G) Test image 16, not very distinct image of square-top blocky crevasses, is associated incorrectly to class 2 ‘parallel crevasses’. Parallel structures dominate, however. (H) Test image 8. Correct class 5 has increased weight, but class 3 (parallel crevasses filled with snow) wins. Parallel structures dominate also visually.</note><note type="content">Table 1: Correctly classified crevasse patterns for NET 1 and WTA rule</note><note type="content">Table 2: Correctly classified crevasse patterns for NET 2 and WTA rule</note><note type="content">Table 3: Error matrix NET 2, test with nontransformed original scenes of validation set 2a</note><note type="content">Table 4: Error matrix NET 2, test with scenes of validation set 2a</note></notesStmt><sourceDesc><biblStruct type="inbook"><analytic><title level="a">A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice</title><author xml:id="author-1"><persName><forename type="first">Ute</forename><surname>Christina Herzfeld</surname></persName><email>ute@mpl.ucsd.edu</email><note type="correspondence"><p>Correspondence address. Institute of Arctic and Alpine Research, University of Colorado Boulder, Campus Box 450, Boulder, CO 80309-0450, USA. Tel.: 1-303-492-6198; fax: 1-303-492-6388</p></note><affiliation>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</affiliation></author><author xml:id="author-2"><persName><forename type="first">Oliver</forename><surname>Zahner</surname></persName><affiliation>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</affiliation></author></analytic><monogr><title level="j">Computers and Geosciences</title><title level="j" type="abbrev">CAGEO</title><idno type="pISSN">0098-3004</idno><idno type="PII">S0098-3004(00)X0064-7</idno><imprint><publisher>ELSEVIER</publisher><date type="published" when="2001"></date><biblScope unit="volume">27</biblScope><biblScope unit="issue">5</biblScope><biblScope unit="page" from="499">499</biblScope><biblScope unit="page" to="512">512</biblScope></imprint></monogr><idno type="istex">F58F80E51A6A3AE890E6AF47A46FD9E51A790B00</idno><idno type="DOI">10.1016/S0098-3004(00)00089-3</idno><idno type="PII">S0098-3004(00)00089-3</idno></biblStruct></sourceDesc></fileDesc><profileDesc><creation><date>2001</date></creation><langUsage><language ident="en">en</language></langUsage><abstract xml:lang="en"><p>A combination of geostatistical methods for data reduction and the neural-network approach for association of input information to object classes forms the basic idea for a connectionist-geostatistical approach to automated image classification. The investigation of the posed glaciological question — understanding a surging glacier —, which presented itself as a natural catastrophe, leaving the onlooking geophysicist without ‘real’ data and only video coverage for analysis, calls for new analysis methods. Deformation states of the rapidly changing ice surface of a surging glacier manifest themselves in crevasse patterns, which were observed in GPS-referenced video images for Bering Glacier, Alaska. It can be shown that directional variograms calculated from video data at appropriate scaling can be used to characterize crevasse patterns. Directional variograms also provide a suitable filter to reduce the data volume necessary for a successful training of a neural network. The final network, a multilayer feedforward perceptron with backpropagation of errors, has a very high success rate for classification of (unknown) crevasse patterns of surging ice from video images.</p></abstract><textClass><keywords scheme="keyword"><list><head>Keywords</head><item><term>Neural network</term></item><item><term>Variogram</term></item><item><term>Videography</term></item><item><term>Glaciology</term></item><item><term>Surge</term></item><item><term>Glaciology</term></item></list></keywords></textClass></profileDesc><revisionDesc><change when="2000-03-28">Modified</change><change when="2001">Published</change></revisionDesc></teiHeader></istex:fulltextTEI><json:item><extension>txt</extension><original>false</original><mimetype>text/plain</mimetype><uri>https://api.istex.fr/document/F58F80E51A6A3AE890E6AF47A46FD9E51A790B00/fulltext/txt</uri></json:item></fulltext><metadata><istex:metadataXml wicri:clean="Elsevier, elements deleted: ce:floats; body; tail"><istex:xmlDeclaration>version="1.0" encoding="utf-8"</istex:xmlDeclaration><istex:docType PUBLIC="-//ES//DTD journal article DTD version 4.5.2//EN//XML" URI="art452.dtd" name="istex:docType"><istex:entity SYSTEM="gr1" NDATA="IMAGE" name="gr1"></istex:entity><istex:entity SYSTEM="gr2" NDATA="IMAGE" name="gr2"></istex:entity><istex:entity SYSTEM="gr3" NDATA="IMAGE" name="gr3"></istex:entity><istex:entity SYSTEM="gr4" NDATA="IMAGE" name="gr4"></istex:entity><istex:entity SYSTEM="gr5" NDATA="IMAGE" name="gr5"></istex:entity><istex:entity SYSTEM="gr6" NDATA="IMAGE" name="gr6"></istex:entity><istex:entity SYSTEM="gr7" NDATA="IMAGE" name="gr7"></istex:entity><istex:entity SYSTEM="gr8" NDATA="IMAGE" name="gr8"></istex:entity><istex:entity SYSTEM="gr9" NDATA="IMAGE" name="gr9"></istex:entity><istex:entity SYSTEM="gr10" NDATA="IMAGE" name="gr10"></istex:entity></istex:docType><istex:document><converted-article version="4.5.2" docsubtype="fla"><item-info><jid>CAGEO</jid><aid>929</aid><ce:pii>S0098-3004(00)00089-3</ce:pii><ce:doi>10.1016/S0098-3004(00)00089-3</ce:doi><ce:copyright type="full-transfer" year="2001">Elsevier Science Ltd</ce:copyright></item-info><head><ce:title>A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice</ce:title><ce:author-group><ce:author><ce:given-name>Ute</ce:given-name><ce:surname>Christina Herzfeld</ce:surname><ce:cross-ref refid="AFF1"><ce:sup>a</ce:sup></ce:cross-ref><ce:cross-ref refid="AFF2"><ce:sup>b</ce:sup></ce:cross-ref><ce:cross-ref refid="CORR1">*</ce:cross-ref><ce:e-address>ute@mpl.ucsd.edu</ce:e-address></ce:author><ce:author><ce:given-name>Oliver</ce:given-name><ce:surname>Zahner</ce:surname><ce:cross-ref refid="AFF1"><ce:sup>a</ce:sup></ce:cross-ref></ce:author><ce:affiliation id="AFF1"><ce:label>a</ce:label><ce:textfn>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</ce:textfn></ce:affiliation><ce:affiliation id="AFF2"><ce:label>b</ce:label><ce:textfn>Institute of Arctic and Alpine Research, University of Colorado Boulder, Campus Box 450, Boulder, CO 80309-0450, USA</ce:textfn></ce:affiliation><ce:correspondence id="CORR1"><ce:label>*</ce:label><ce:text>Correspondence address. Institute of Arctic and Alpine Research, University of Colorado Boulder, Campus Box 450, Boulder, CO 80309-0450, USA. Tel.: 1-303-492-6198; fax: 1-303-492-6388</ce:text></ce:correspondence></ce:author-group><ce:date-received day="19" month="10" year="1999"></ce:date-received><ce:date-revised day="28" month="3" year="2000"></ce:date-revised><ce:date-accepted day="28" month="3" year="2000"></ce:date-accepted><ce:abstract><ce:section-title>Abstract</ce:section-title><ce:abstract-sec><ce:simple-para>A combination of geostatistical methods for data reduction and the neural-network approach for association of input information to object classes forms the basic idea for a connectionist-geostatistical approach to automated image classification. The investigation of the posed glaciological question — understanding a surging glacier —, which presented itself as a natural catastrophe, leaving the onlooking geophysicist without ‘real’ data and only video coverage for analysis, calls for new analysis methods. Deformation states of the rapidly changing ice surface of a surging glacier manifest themselves in crevasse patterns, which were observed in GPS-referenced video images for Bering Glacier, Alaska. It can be shown that directional variograms calculated from video data at appropriate scaling can be used to characterize crevasse patterns. Directional variograms also provide a suitable filter to reduce the data volume necessary for a successful training of a neural network. The final network, a multilayer feedforward perceptron with backpropagation of errors, has a very high success rate for classification of (unknown) crevasse patterns of surging ice from video images.</ce:simple-para></ce:abstract-sec></ce:abstract><ce:keywords class="keyword"><ce:section-title>Keywords</ce:section-title><ce:keyword><ce:text>Neural network</ce:text></ce:keyword><ce:keyword><ce:text>Variogram</ce:text></ce:keyword><ce:keyword><ce:text>Videography</ce:text></ce:keyword><ce:keyword><ce:text>Glaciology</ce:text></ce:keyword><ce:keyword><ce:text>Surge</ce:text></ce:keyword><ce:keyword><ce:text>Glaciology</ce:text></ce:keyword></ce:keywords></head></converted-article></istex:document></istex:metadataXml><mods version="3.6"><titleInfo><title>A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice</title></titleInfo><titleInfo type="alternative" contentType="CDATA"><title>A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice</title></titleInfo><name type="personal"><namePart type="given">Ute</namePart><namePart type="family">Christina Herzfeld</namePart><affiliation>E-mail: ute@mpl.ucsd.edu</affiliation><affiliation>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</affiliation><description>Correspondence address. Institute of Arctic and Alpine Research, University of Colorado Boulder, Campus Box 450, Boulder, CO 80309-0450, USA. Tel.: 1-303-492-6198; fax: 1-303-492-6388</description><role><roleTerm type="text">author</roleTerm></role></name><name type="personal"><namePart type="given">Oliver</namePart><namePart type="family">Zahner</namePart><affiliation>Geomathematik, Fachbereich Geographie/Geowissenschaften, Universität Trier, D-54286 Trier, Germany</affiliation><role><roleTerm type="text">author</roleTerm></role></name><typeOfResource>text</typeOfResource><genre type="research-article" displayLabel="Full-length article"></genre><originInfo><publisher>ELSEVIER</publisher><dateIssued encoding="w3cdtf">2001</dateIssued><dateModified encoding="w3cdtf">2000-03-28</dateModified><copyrightDate encoding="w3cdtf">2001</copyrightDate></originInfo><language><languageTerm type="code" authority="iso639-2b">eng</languageTerm><languageTerm type="code" authority="rfc3066">en</languageTerm></language><physicalDescription><internetMediaType>text/html</internetMediaType></physicalDescription><abstract lang="en">A combination of geostatistical methods for data reduction and the neural-network approach for association of input information to object classes forms the basic idea for a connectionist-geostatistical approach to automated image classification. The investigation of the posed glaciological question — understanding a surging glacier —, which presented itself as a natural catastrophe, leaving the onlooking geophysicist without ‘real’ data and only video coverage for analysis, calls for new analysis methods. Deformation states of the rapidly changing ice surface of a surging glacier manifest themselves in crevasse patterns, which were observed in GPS-referenced video images for Bering Glacier, Alaska. It can be shown that directional variograms calculated from video data at appropriate scaling can be used to characterize crevasse patterns. Directional variograms also provide a suitable filter to reduce the data volume necessary for a successful training of a neural network. The final network, a multilayer feedforward perceptron with backpropagation of errors, has a very high success rate for classification of (unknown) crevasse patterns of surging ice from video images.</abstract><note type="content">Fig. 1: Map of Bering Glacier, Alaska, with tracks of video flights.</note><note type="content">Fig. 2: Example of (A) original (nontransformed) video image and (B) vertically mirrored image of (A) with three-directional variograms.</note><note type="content">Fig. 3: Classes of crevasse types in a surging glacier. Examples from video images of Bering Glacier, Alaska.</note><note type="content">Fig. 4: Individual video frame with the two 150×150 pixel subsets that were used in classification.</note><note type="content">Fig. 5: Idealized learning curve and validation curve for neural network. Lower curve — learning curve. Upper curve — validation curve.</note><note type="content">Fig. 6: Typical learning and validation curves encountered in modeling and training a neural network. Lower curves — learning curves. Upper curves — validation curves. (A) Nonconvergent learning process, resultant from training net with two hidden layers. (B) Training a net with too few (10) hidden layer neurons only leaves overall too high error in both learning and validation curves (see Fig. 7). (C) Poor generalization after overfitting with too many (42) intermediate layer neurons (see Fig. 7). (D) Unstable fit caused by too high learning rate of 0.4. (E) Good learning behavior (cf. idealized curve, Fig. 5). Final net topology. (F) Nonconvergent, heavily oscillating learning with high errors, resultant from application of quickpropagation algorithm.</note><note type="content">Fig. 7: Determination of optimal number of intermediate layer neurons. Training versus validation error. 25 to 40 neurons are best.</note><note type="content">Fig. 8: Topology of MLP-BP with 42 neuron input layer, one intermediate layer with 32 neurons, and 6 neuron output layer, layerwise fully connected (1536 connections, no shortcut connections). NET 1.</note><note type="content">Fig. 9: Topology of MLP-BP with 42 neuron input layer, one intermediate layer with 28 neurons, and 9 neuron output layer, fully connected with shortcut connections (1606 connections). NET 2.</note><note type="content">Fig. 10: Classification results and generalization capability of MLP-BP (42-28-9) (NET 2). Examples from test set 2. (A) Test image 14. Bidirectional pattern (class 4) is clearly recognized. (B) Test image 2. En-échelon type (class 8) classified correctly. (C) Test image 10. Chaos (class 1) classified correctly. (D) Test image 11. Class 0 (undisturbed ice surface, moraines, rock) recognized with weight 1 despite small disturbance in image. (E) Test image 19. Class 0 (undisturbed ice surface, moraines, rock) recognized correctly despite large disturbance. (F) Test image 4, showing new pattern, is associated to class 2 (parallel crevasses), which is most closely related. (G) Test image 16, not very distinct image of square-top blocky crevasses, is associated incorrectly to class 2 ‘parallel crevasses’. Parallel structures dominate, however. (H) Test image 8. Correct class 5 has increased weight, but class 3 (parallel crevasses filled with snow) wins. Parallel structures dominate also visually.</note><note type="content">Table 1: Correctly classified crevasse patterns for NET 1 and WTA rule</note><note type="content">Table 2: Correctly classified crevasse patterns for NET 2 and WTA rule</note><note type="content">Table 3: Error matrix NET 2, test with nontransformed original scenes of validation set 2a</note><note type="content">Table 4: Error matrix NET 2, test with scenes of validation set 2a</note><subject><genre>Keywords</genre><topic>Neural network</topic><topic>Variogram</topic><topic>Videography</topic><topic>Glaciology</topic><topic>Surge</topic><topic>Glaciology</topic></subject><relatedItem type="host"><titleInfo><title>Computers and Geosciences</title></titleInfo><titleInfo type="abbreviated"><title>CAGEO</title></titleInfo><genre type="journal">journal</genre><originInfo><dateIssued encoding="w3cdtf">200106</dateIssued></originInfo><identifier type="ISSN">0098-3004</identifier><identifier type="PII">S0098-3004(00)X0064-7</identifier><part><date>200106</date><detail type="volume"><number>27</number><caption>vol.</caption></detail><detail type="issue"><number>5</number><caption>no.</caption></detail><extent unit="issue pages"><start>499</start><end>616</end></extent><extent unit="pages"><start>499</start><end>512</end></extent></part></relatedItem><identifier type="istex">F58F80E51A6A3AE890E6AF47A46FD9E51A790B00</identifier><identifier type="DOI">10.1016/S0098-3004(00)00089-3</identifier><identifier type="PII">S0098-3004(00)00089-3</identifier><accessCondition type="use and reproduction" contentType="copyright">©2001 Elsevier Science Ltd</accessCondition><recordInfo><recordContentSource>ELSEVIER</recordContentSource><recordOrigin>Elsevier Science Ltd, ©2001</recordOrigin></recordInfo></mods></metadata><serie></serie></istex></record>Pour manipuler ce document sous Unix (Dilib)
EXPLOR_STEP=$WICRI_ROOT/Wicri/Rhénanie/explor/UnivTrevesV1/Data/Istex/Corpus
HfdSelect -h $EXPLOR_STEP/biblio.hfd -nk 001643 | SxmlIndent | more
Ou
HfdSelect -h $EXPLOR_AREA/Data/Istex/Corpus/biblio.hfd -nk 001643 | SxmlIndent | more
Pour mettre un lien sur cette page dans le réseau Wicri
{{Explor lien
|wiki= Wicri/Rhénanie
|area= UnivTrevesV1
|flux= Istex
|étape= Corpus
|type= RBID
|clé= ISTEX:F58F80E51A6A3AE890E6AF47A46FD9E51A790B00
|texte= A connectionist-geostatistical approach to automated image classification, applied to the analysis of crevasse patterns in surging ice
}}
| This area was generated with Dilib version V0.6.31. |  |