Close
Metadata

1. Identity statement
Reference TypeConference Abstract (Conference Proceedings)
Sitemtc-m21b.sid.inpe.br
Holder Codeisadg {BR SPINPE} ibi 8JMKD3MGPCW/3DT298S
Identifier8JMKD3MGP5W34M/3GHDDDB
Repositorysid.inpe.br/mtc-m21b/2014/06.25.22.54
Last Update2014:06.25.22.54.06 (UTC) administrator
Metadata Repositorysid.inpe.br/mtc-m21b/2014/06.25.22.54.06
Metadata Last Update2018:06.04.03.04.20 (UTC) administrator
Labelself-archiving-INPE-MCTI-GOV-BR
Citation KeyMusciFeitCostAlme:2014:ImInMe
TitleAn Image Interpretation Methodology Using Independent Class Specific Segmentations
Year2014
Access Date2024, May 18
Secondary TypePRE CI
Number of Files1
Size3017 KiB
2. Context
Author1 Musci, Marcelo
2 Feitosa, Raul Queiroz
3 Costa, Gilson Alexandre Ostwald Pedro da
4 Almeida, Cláudia Maria de
Resume Identifier1
2
3
4 8JMKD3MGP5W/3C9JGS3
Group1
2
3
4 DSR-OBT-INPE-MCTI-GOV-BR
Affiliation1
2
3
4 Instituto Nacional de Pesquisas Espaciais (INPE)
Author e-Mail Address1
2
3
4 almeida@dsr.inpe.br
e-Mail Addressalmeida@dsr.inpe.br
Conference NameConference on Geographic Object-Based Image Analysis, 5 (GEOBIA 2014).
Conference LocationTessalônica, Grécia
Date21-24 maio, 2014
Pages116
Book TitleAbstracts
History (UTC)2014-06-25 22:54:06 :: almeida@dsr.inpe.br -> administrator ::
2018-06-04 03:04:20 :: administrator -> marcelo.pazos@inpe.br :: 2014
3. Content and structure
Is the master or a copy?is the master
Content Stagecompleted
Transferable1
Content TypeExternal Contribution
Version Typepublisher
Keywordssegmentation
hierarchical segmentation
parameter tuning
AbstractHierarchical Segmentation (HS) has been a dominant methodology within the GEOBIA community in the last years. In essence, it involves the segmentation of the input image in a number of distinct scales, whereby each segment at a finer scale lies entirely within a segment of a coarser scale. Such methodology implies at least two difficulties. First, inaccuracies of the first segmentation step propagate to the scale levels segmented subsequently. Second, and even more important, it is generally difficult to find a single set of segmentation parameter values that optimises spatial accuracy for all object classes present at the same scale. In general, the user chooses for each scale level the parameter values that yield an acceptable trade-off for all object-classes coexisting at that level, which are generally non optimal for all those object classes. This work proposes and evaluates an alternative to HS, so-called class specific segmentation (CSS). Basically CSS consists of performing a separate and independent, optimised segmentation for each object class. This may be achieved by tuning the segmentation parameters individually for each object class, by running different segmentation algorithms for each class, or a mixture of both, so as to obtain (near) optimum segmentation for each class. The benefit of CSS relative to HS is twofold. First, image objects of each class are expected to be more precisely delineated, improving, therefore, the overall spatial accuracy. Second, since the borders of image objects will be more accurate, morphological features will be more discriminative, contributing as well to a better thematic accuracy. However, CSS brings about a problem not present in HS. Multiple potentially contradictory segmentation results will coexist in some phases of the interpretation procedure. This problem is handled by CSS in the classification stage in the following way. For each given object class a specific detector is designed, which searches the corresponding segmentation outcome for instances of its object class. As a result some image regions will possibly be assigned to more than one class, giving rise to what we call spatial conflicts. Besides delivering a class/non class logical label for each segment, the class specific detectors also yield a membership or a probability value that expresses how well a segment fits the detectors class. Spatial conflicts are then resolved by assigning regions of conflict to the class with the highest membership or likelihood. Three segmentation algorithms in combination with two alternative detector designs have been tested upon three very-high resolution images. The results indicated unanimously that CSS performs significantly better than, or sporadically as good as, the HS approach, both in terms of spatial as well as thematic accuracy.
AreaSRE
Arrangementurlib.net > BDMCI > Fonds > Produção anterior à 2021 > DIDSR > An Image Interpretation...
doc Directory Contentaccess
source Directory Contentthere are no files
agreement Directory Content
agreement.html 25/06/2014 19:54 1.0 KiB 
4. Conditions of access and use
data URLhttp://mtc-m21b.sid.inpe.br/ibi/8JMKD3MGP5W34M/3GHDDDB
zipped data URLhttp://mtc-m21b.sid.inpe.br/zip/8JMKD3MGP5W34M/3GHDDDB
Languageen
Target FileGeobia_2014_Abstract_Book_Musci et al.pdf
User Groupalmeida@dsr.inpe.br
marcelo.pazos@inpe.br
Reader Groupadministrator
marcelo.pazos@inpe.br
Visibilityshown
Read Permissionallow from all
Update Permissionnot transferred
5. Allied materials
Mirror Repositoryiconet.com.br/banon/2006/11.26.21.31
Next Higher Units8JMKD3MGPCW/3ER446E
Citing Item Listsid.inpe.br/mtc-m21/2012/07.13.14.43.49 2
Host Collectionsid.inpe.br/mtc-m21b/2013/09.26.14.25.20
6. Notes
Empty Fieldsarchivingpolicy archivist callnumber copyholder copyright creatorhistory descriptionlevel dissemination doi edition editor format isbn issn lineage mark nextedition notes numberofvolumes orcid organization parameterlist parentrepositories previousedition previouslowerunit progress project publisher publisheraddress rightsholder schedulinginformation secondarydate secondarykey secondarymark serieseditor session shorttitle sponsor subject tertiarymark tertiarytype type url volume
7. Description control
e-Mail (login)marcelo.pazos@inpe.br
update 


Close