A Comparison Of Segmentation Programs For High Resolution Remote .
A COMPARISON OF SEGMENTATION PROGRAMS FOR HIGH RESOLUTIONREMOTE SENSING DATAG. Meinel, M. NeubertLeibniz Institute of Ecological and Regional Development (IOER), Weberplatz 1, D-01217 Dresden, Germany G.Meinel@ioer.de, M.Neubert@ioer.dePS ThS9: Uncertainty, Consistency and Accuracy of Data and ImageryKEY WORDS: Automation, Land cover, Classification, Software, IKONOS, Performance, Comparison, qualitativeABSTRACT:Methods of image segmentation become more and more important in the field of remote sensing image analysis – in particular due tothe increasing spatial resolution of imagery. The most important factor for using segmentation techniques is segmentation quality.Thus, a method for evaluating segmentation quality is presented and used to compare results of presently available segmentationprograms. Firstly, an overview of the software used is given. Moreover the quality of the individual segmentation results is evaluatedbased on pan-sharpened multi-spectral IKONOS data. This is done by visual comparison, which is supplemented by a detailedinvestigation using visual interpreted reference areas. Geometrical segment properties are in the focus of this quantitative evaluation.The results are assessed and discussed. They show the suitability of the tested programs for segmenting very high resolution imagery.1. INTRODUCTION2. EVALUATED SEGMENTATION SOFTWARESegmentation means the grouping of neighbouring pixels intoregions (or segments) based on similarity criteria (digitalnumber, texture). Image objects in remotely sensed imagery areoften homogenous and can be delineated by segmentation.Thus, the number of elements as a basis for a following imageclassification is enormously reduced. The quality ofclassification is directly affected by segmentation quality. Hencequality assessment of segmentation is in the focus of thisevaluation of different presently available segmentationsoftware.Recently there exists a multitude of implemented segmentationalgorithms for remote sensing tasks, partially having verydifferent characteristics. Only some of them are availablecommercially. Often they are developed by research institutionsor universities. For evaluating capabilities of differentalgorithms the following programs were compared: eCognition 2.1 resp. 3.0 (Definiens Imaging GmbH,Munich, Germany); Data Dissection Tools (INCITE, Stirling University,UK); CAESAR 3.1 (N.A.Software Ltd., Liverpool, UK) InfoPACK 1.0 (InfoSAR Ltd., Liverpool, UK); Image segmentation for Erdas Imagine (USDA ForestService, Remote Sensing Applications Center, SaltLake City, USA); Minimum Entropy Approach to Adaptive ImagePolygonization (University of Bonn, Institute ofcomputer science, Bonn, Germany); SPRING 4.0 (National Institute for Space Research,São José dos Campos, Brasilia).Despite some early research activities (e.g. Kettig & Landgrebe,1976), image segmentation was established late in the field ofremote sensing. First beginning with the availability of veryhigh resolution imagery ( 1 m) and their characteristics (highlevel of detail, spectral variance etc.) this method has becomepopular as a common variant of data interpretation.Recent investigations have shown that a pixel-based analysis ofsuch high resolution imagery has explicit limits. Usingsegmentation techniques some problems of pixel-based imageanalysis could be overcome (e.g. Meinel, Neubert & Reder,2001).This paper is not related to more mathematical surveys ofsegmentation, like Haralick & Shapiro (1985) or Pal & Pal(1993). It is rather a more application-oriented comparisonbased on real remote sensing data.Among segmentation software there is a growing number offeature extraction programs. In contrast they do not fractionalisethe whole image but rather select specific objects from imagery.APEX (PCI Geomatics), FeatureXTR (Hitachi Software GlobalTechnology) or Feature Analyst (Visual Learning Systems)belonging to this group of tools, which are not consideredherein.All programs are described in brief in table 1. The choice ofapproaches was based on the software segmentation suitabilityfor remote sensing imagery. On the other hand cooperativenessof the developers was a precondition for this survey.3. METHOD3.1Used imageryPan-sharpened multi-spectral IKONOS data (1 m groundresolution, principle component algorithm) of two test areaswere segmented by the software above. Each test area has a sizeof about 2000 by 2000 Pixel, representing an urban and a rurallandscape. The procedure was aimed at the extraction of
eCognition2.1 resp. e.org.uk/projects/AlgorithmRegiongrowingField ingImageanalysis,statisticalphysicsFerber rDistributionIn- and referenceBaatz &Schäpe 2000State r ofparametersCa. /2002 resp.11/2002WinStand-alone3Win, Linux,Unix, SGIN/a (maybeMatLab)310 min10 minNo resp.Noyes (ver. 3.0)Yes (FuzzyNoLogic, Near.Neighbour)Max. image10000 by4000 bysize [ca. Pixel]1100004000Max. bit depth32 bit16 bitInput formatsRaster, Shape Raster (TIFF)Vector outputShapeNo (externalformatconversion)Use of externalYesNodataAvailabilityCommercial CommercialCa. costs7(commercial/research)14.000 /9.000 (non-profit)/2.900 8N/a resp.free (nonprofit)CAESAR 3.1InfoPACK1.0Image SegMinimumSPRING 4.0mentationEntropy(for ErdasApproachImagine)N.A.Software InfoSAR Ltd. USDA Forest University ofNationalLtd.Service, Re- Bonn, Inst. of Institute formote e/ ensing, esp. sensing, esp.sensingn of noisysensingradar dataradar dataimageryCook et al.1996Cook et al.1996Ruefenacht etal. 200209/199803/2003Linux,SolarisStand-alone6Bins et al.199602/2002Hermes &Buhmann200108/2002Linux, WinWinLinuxWin, 222N/a (long)No10 minNo1,5 hNoSeveral hoursNo30 minNoYesYes(MaximumLikelihood)NoLimitations64 bitNetCDF4No (externalconversion)YesNoNoYes2000 by20008 bitIMGArcCoverageYes2000 by15008 bitRasterNo (externalconversion)5Yes2000 by20008 bitGRIB4No (externalconversion)6JaCommercialFreewareOn requestFreeware8.300 /5.800 free-free2000 by200016 bitIMGNo (externalconversion)YesOut ofdistribution-08/20031Specification heavily depends on system resources, particularly main memory; 2 for the used imagery (2000 by 2000 Pixel); 3 whenimage size is modified; 4 convertible from divers raster formats, e.g. GeoTIFF, IMG; 5 Proprietary vector format; 6 Internal vectorconversion was not possible; 7 State April 2003; 8 A simplified version (eCognition elements) is available since May 2003 (3500 )Table 1. Outline about evaluated segmentation software.relevant land cover/use object boundaries. The segmentationswere produced by the developers – except for eCognition,SPRING and the Erdas Imagine extension ‚Image Segmentation’. Thus, optimal segmentation results (e.g. parametersettings) by experts were assured.Results for the software CAESAR were available only for therural test area. The segmentations of the ‚Minimum EntropyApproach’ cover only one fourth of both areas (each 1000 by1000 Pixel) due to a lack of performance. Except for InfoPACKthe segmentations were done in different levels using altered
parameters affecting segment size. When an object was poorlysegmented, coarser or finer segmentation levels could be used.possible. Large homogenous image objects are divided arbitrarily sometimes.3.2eCognition uses a new segmentation algorithm since release 3.0which enables a result not depending on image size. This is animportant improvement because often parameters are tested onsmall subsets. Nevertheless the old algorithm of version 2.1could still be used alternatively in the current release 4.0.Altogether eCognition has a high potential due to its multi-scalesegmentation and the fuzzy logic based image classificationcapabilities. Because of the various interfaces to other GIS andremote sensing software systems important user requirementsare complied.Pre-processing of the segmentation resultsAll segmentation results were converted into vector format(ArcView shape file) for the subsequent comparison ofgeometry. Only eCognition and the Erdas extension ‚ImageSegmentation’ are able to generate a GIS-readable vectorformat. All other results were generated in raster format (TIFF)with a unique value for each segment. Geocoding was restoredby adding a world file (TFW). Then a raster-to-vectorconversion was carried out using Erdas Imagine. Only in thecase of the ‚Minimum Entropy Approach’ this procedure resultsto some negative effects, because the implemented triangulationalgorithm fractionalises the image without respect to rasterboundaries. The preliminary segments are stored in aproprietary vector format, which cannot be saved. Rather thesegmentation result was converted into a raster output, whichadmittedly leads to more partial segments and faultysegmentations (unclosed polygons etc.). These unavoidableeffects have a negative influence to the quality assessment.3.3Quality assessmentFirstly, all results came under an overall visual survey. Generalcriterions, like the delineation of varying land cover types (e.g.meadow/forest, agriculture/meadow, etc.), the segmentation oflinear objects, the occurrence of faulty segmentations and adescription of the overall segmentation quality were in the focusof this first step.Figure 1. Segmentation result of eCognition 2.1.Furthermore a detailed comparison based on visual delineatedand clearly definable reference areas was carried out. Therefore20 different areas (varying in location, form, area, texture,contrast, land cover type etc.) were selected and each wasvisually and geometrically compared with the segmented pendants. The geometrical comparison is a combination of formalfactors (area, perimeter, and Shape Index (area-perimeter-ratio))and the number of segments resp. partial segments (in the caseof over-segmentation). For all features the variances to thereference values were calculated.As partial segments all polygons with at least 50 % area in thereference object were counted. The Shape Index comes fromlandscape ecology and indicates the polygon form. It iscalculated by the quotient of perimeter and four times thesquare root of area. Additionally the quality of segmentationwas visually rated (0 poor, 1 medium, 2 good).A good segmentation quality is reached, when the overalldifferences of all criteria between the segmentation results andthe associated reference objects are as low as possible.Furthermore the objects of interest should not be oversegmented too much.4. RESULTS4.1Overall visual surveyeCognition: Despite their differences albeit using the sameparameters the segmentation the results of eCognition 2.1(figure 1) and 3.0 (figure 2) are of good quality. Indeed theysometimes contain irregular or ragged delineated segments,especially at seam-forming boundaries and in woody areas. Inareas of low contrast the occurrence of faulty segmentations isFigure 2. Segmentation result of eCognition 3.0.Data Dissection Tools: The segmentations of the ‚DataDissection Tools’ (figure 3) offer only partly satisfying results.The software tends to a strong over-segmentation of brightimage areas, whereby a multitude of small segments occur.Homogeneous areas like fields, meadows or water bodies aresegmented almost correct. Only very large areas are dividedarbitrarily sometimes. Explicit mistakes of delineation appear inimage areas of low contrast (e.g. woody areas). As in brightimage areas sometimes single pixels are segmented (comparableto the salt-and-pepper effect). The near infrared was onlymarginally used in consequence of an unfavourable weighting,which impacts especially the separation of deciduous vs.coniferous forest.
Figure 3. Segmentation result of the ‘Data Dissection Tools’.Figure 5. Segmentation result of InfoPACK 1.0.CEASAR: The program CEASAR 3.1 which was developed forradar data leads to results that cannot be used for further processing (figure 4). The produced segments are compact and of asimilar size. This effect occurs even though using different segmentation parameters which yield only to a varying averagesegment size. Thus, small structures and in particular linearelements are often segmented faulty and an over-segmentationis the consequence. Boundaries of low contrast are representedbadly, sometimes boundaries of sufficient contrast too (e.g.forest vs. meadow).Erdas Imagine extension ‘Image Segmentation’: The ErdasImagine extension ‘Image Segmentation’ (figure 6) leads toover- and under-segmentation within the same segmentationresult. Well-contrasted boundaries between main land coverclasses were correctly represented. Areas of low contrast wereoften not segmented. In particular the delineation of fields vs.meadow was problematic. Forested areas were merged intolarge conglomerates, with small island segments inside onlyslightly greater than the parameter minimal segment sizechosen. Linear elements were segmented inadequate and homogeneous image objects were divided frequently.Furthermore, the result contains faulty segmentations in termsof non-explainable horizontal or vertical boundaries. Thedegree of this effect has been slightly reduced by a new versionfrom September 2002. It was mainly a consequence of the blocksize used by the software, which can now be set freely inaccordance to the available system resources or the image size.Thus, the computing time has been reduced too. But thesegmentation quality remained nearly unimproved.Figure 4. Segmentation result of CAESAR 3.1.InfoPACK: The result of InfoPACK 1.0 (figure 5), the furtherdevelopment of CAESAR, shows a good delineation for most ofthe objects, but tends strongly to over-segmentation. Homogeneous areas are thereof less affected and are adequatelyrepresented. In particular especially forests and built-up areaswere much partitioned. At land cover transitions often interfering seam-forming segments were created. Generally lowcontrasted boundaries were segmented correctly. Compact andnearly similar sized segments as in CAESAR exist no longer.For processing scenes of any size the software uses an implemented tiling algorithm. Indeed this leads to additional segmentboundaries at the tile transitions. Furthermore margin effectscan yield to different results on both sides of the tile boundary.As eCognition the software contains additional classificationtools. Thus, a classification based on merging of similar classified and neighbouring segments is possible and this reduces thenumber of elements to be classified significantly. It must bepointed out, that InfoPACK as well as CEASAR have beendeveloped to analyse very noisy radar data. Hence, the segmentation of optical data could be suboptimal.Figure 6. Segmentation result of the Erdas extension ‘ImageSegmentation’.‘Minimum Entropy Approach’: The ‘Minimum Entropy Approach’ (figure 7) was well reproducing straight boundaries ofman-made features (e.g. field boundaries, roads). More complexnatural boundaries (e.g. forest edges) were often imprecisely delineated by the used triangulation algorithm. Large homogeneous areas were divided frequently. Simultaneously, effects ofthe above-mentioned raster-to-vector conversion could befound. Generally it can be pointed out, that the triangulationalgorithm often leads to straight segment boundaries or sections
Figure 7. Segmentation result of the ‘Minimum EntropyApproach’.resp. typical segment shapes which are closer to a humaninterpretation.SPRING 4.0: The segmentation results of the region growingalgorithm implemented in the image processing softwareSPRING are showing a good overall impression (figure 8).Homogeneous areas are delineated well but often oversegmented. Heavily textured areas as forests are mostly undersegmented. Sporadic segmentation mistakes occur. However,the ease of operation as well as the data handling of thesoftware is insufficient. The implemented edge-based watershed-algorithm was also tested, but was leading to worse results(strong over-segmentation) and was therefore not used for hisevaluation.Segmentation programNumber of referenceareas1eCognition eCognitionData2.1Dissection3.0ToolsFigure 8. Segmentation result of SPRING 4.0.4.2Comparison based on reference areasThe overall results of all 20 reference areas are cumulated intable 2. As shown in this table the results of SPRING,eCogniton 2.1 and 3.0, InfoPACK and the ‚Minimum EntropyApproach’ are reaching the best average area conformity.Except for InfoPACK, the same result is shown in the case ofthe average conformity of perimeter and the Shape Index. Thehigh conformity of the Shape Index in the case of the‚Minimum Entropy Aproach’ is affected by the segment shapesresulting from the triangulation algorithm which is closer tohuman interpretation.Especially within the number of segments both versions ofeCognition revealed their strengths. Both led to the slightestover-segmentation in this evaluation. In this regard the resultsof SPRING could be rated as good. The results of the ErdasImagine extension ‚Image Segmentation’ also reached a slightnumber of segments, but due to strong differences of the othervalues the result is indicating under-segmentation.CAESAR3.1InfoPACK Image Seg- Minimum1.0mentationEntropy(for Erdas ApproachImagine)SPRING4.0202020101202011120Average difference ofarea [%]12,515,92100,375,111,1107,013,68,2Average difference ofperimeter [%]15,917,2475,635,130,9177,310,010,8Average difference ofShape Index [%]16,716,238,925,525,587,110,011,7Average number ofpartial segments1,91,8134,610,417,15,99,06,2Average quality, visualevaluated [0 2]1,00,90,20,00,60,20,80,9differing number due to partial incomplete segmentation resultsTable 2.Cumulated results of all 20 reference areas.
Thus, the results of visual compared qualities results of theindividual segmentation programs are reinforced. Only thesegmentations by SPRING as well as eCognition 2.1 and 3.0have reached good overall results. These programs leading tothe slightest differences to the reference areas at all factorsinvestigated. Likewise InfoPack and the ‚Minimum EntropyApproach’ yielded to an acceptable quality, but InfoPack tendsto over-segmentation and the ‚Minimum Entropy Approach’ hassome processing problems as stated above. The results of thethree remaining programs did not reach this quality. Theyprobably failed due to the high complexity of high resolutionremote sensing imagery. Often a strong faulty or oversegmentation is the consequence. Furthermore, the grade ofconformity with the reference objects is only slight. Indeed ithas to be reemphasised, that some of the approaches have notprimarily been developed for (optical) remote sensing imageanalysis.5. CONCLUSIONSDue to the dissimilitude of the software implementations thesegmentation results are naturally varying. It was shown, thatbest results have been calculated using the commercial softwarepackages – eCognition and InfoPACK. The only exception isthe freeware SPRING, but with the disadvantages of a higheroperating expense and a worse handling. However, the use ofInfoPACK leads to more over-segmented results. Anotheralgorithm with a high potential is the ‘Minimum EntropyApproach to Adaptive Image Polygonization’, but there wasalso an over-segmentation. The results of the other programswere not satisfying user’s demands.Image segmentation has become essential for high resolutionremote sensing imagery. The further development of firstpromising segmentation approaches offers a lot of potentials tomake remote sensing image analysis more accurate as well asmore efficient. The use of texture information for segmentationcould improve the results. Indeed at the moment onlyInfoPACK provides this option, which was not used for thisevaluation. Increasing combinations, for instance withalgorithms of feature extraction, edge-oriented or model-basedsegmentation should be aspired for the improvement ofsegmentation quality.Segmentation algorithms respond often very sensitively in thecase of negligible variations, like slight parameter chances, theorder of segmentation hierarchical approaches or the image dataitself (image size, bit depth, etc.). Thus, the user is confrontedwith a high degree of freedom, which should be minimised. Forinstance, when selecting parameters by the trial-and-errormethod the results are highly influenced by subjectivity. Theintegration of instruments for evaluation of segmentationquality appears desirable.In future additional segmentation programs will be evaluated,for instance the image processing systems HALCON andIMPACT. Moreover, this more qualitative evaluation will beadded by a quantitative comparison using the software SEQTool (Delphi IMM GmbH, 2003). This tool compares theidenticalness of polygon outlines (segmented vs. reference).REFERENCESBaatz, M. & Schäpe, A., 2000: Multiresolution Segmentation –an optimization approach for high quality multi-scale imagesegmentation. In: Strobl, J. et al. (eds.): AngewandteGeographische InformationsverarbeitungHeidelberg, pp. 12-23.XII.Wichmann,Bins, L. S.; Fonseca, L. M. G.; Erthal, G. J. & Ii, F. A. M.(1996): Satellite imagery segmentation: a region growingapproach. Proceedings of VIII Brazilian Remote SensingSymposium, Salvador, Bahia: 4 p.Cook, R.; McConnell; I., Stewart, D. & Oliver, C., 1996:Segmentation and simulated annealing. In: Franceschetti, G. etal. (eds.): Microwave Sensing and Synthetic Aperture Radar.Proc. SPIE 2958, pp. 30-35.Delphi IMM GmbH, 2004: Bestimmung der Segmentierungsqualität bei objektorientierten Bildanalyseverfahren mit SEQ.http://www.delphi-imm.de/neu/ Fernerkundung Software SEQ (in German only, accessed 26 Apr. 2003).Haralick, R. & Shapiro, L., 1985: Image segmentationtechniques. Computer Vision, Graphics, and Image Processing,vol. 29, pp 100-132.Hermes, L. & Buhmann, J. M., 2001: A New AdaptiveAlgorithm for the Polygonization of Noisy Imagery. TechnicalReport IAI-TR-2001-3, Dept. of Computer Science III,University of Bonn.Kettig, R. L. & Landgrebe, D. A., 1976: Classification ofMultispectral Image Data by Extraction and Classification ofHomogeneous Objects. IEEE Transactions on GeoscienceElectronics, Vol. GE-14, No. 1, pp. 19-26.Meinel, G., Neubert, M. & Reder, J., 2001: The potential use ofvery high resolution satellite data for urban areas – Firstexperiences with IKONOS data, their classification andapplication in urban planning and environmental monitoring.In: Jürgens, C. (ed.): Remote sensing of urban areas.Regensburger Geographische Schriften 35, pp. 196-205.Pal, N. R. & Pal, S. K., 1993: A review on image segmentationtechniques. Pattern Recognition, vol. 26, pp. 1277-1294.Ruefenacht, B., Vanderzanden; D., Morrison, M. & Golden, M.,2002: New Technique for Segmenting Images. Technicaldocumentation.von Ferber, C. & Wörgötter, F., 2000: Cluster update algorithmand recognition. Phys. Rev. E 62, Nr. 2, Part A, pp. 1461-1464.ACKNOWLEDGEMENTSThe authors wish to thank the German Research Foundation forclaiming the project “Use and application of high resolutionsatellite imagery for spatial planning” (Me 1592/1-2). Forprocessing the different segmentations we thank Ms. Prietzsch(Infoterra), Mr. Oliver (InfoSAR), Mr. von Ferber (Universityof Freiburg) and Mr. Hermes (University of Bonn).
Methods of image segmentation become more and more important in the field of remote sensing image analysis - in particular due to . The most important factor for using segmentation techniques is segmentation quality. Thus, a method for evaluating segmentation quality is presented and used to compare results of presently available .
Internal Segmentation Firewall Segmentation is not new, but effective segmentation has not been practical. In the past, performance, price, and effort were all gating factors for implementing a good segmentation strategy. But this has not changed the desire for deeper and more prolific segmentation in the enterprise.
Internal Segmentation Firewall Segmentation is not new, but effective segmentation has not been practical. In the past, performance, price, and effort were all gating factors for implementing a good segmentation strategy. But this has not changed the desire for deeper and more prolific segmentation in the enterprise.
segmentation research. 2. Method The method of segmentation refers to when the segments are defined. There are two methods of segmentation. They are a priori and post hoc. Segmentation requires that respondents be grouped based on some set of variables that are identified before data collection. In a priori segmentation, not only are the
Fig. 1.Overview. First stage: Coarse segmentation with multi-organ segmentation withweighted-FCN, where we obtain the segmentation results and probability map for eachorgan. Second stage: Fine-scaled binary segmentation per organ. The input consists of cropped volume and a probability map from coarse segmentation.
A comparison of spectral clustering methods is given in [8]. The authors attempted . As mentioned, we will compare three different segmentation techniques, the mean shift-based segmentation algorithm [1], an efficient graph-based segmentation algo-rithm [4], and a hybrid of the two. We have chosen to look at mean shift-based segmen-
Psychographic Segmentation is also referred to as behavioral segmentation. Psychographic segmentation is analyzed in literature as a useful tool to explore the link between satisfaction and revisit intention (Gountas & Gountas 2001; Cole 1997). This type of segmentation divides the market into groups according to visitors' lifestyles.
[12], [13], [16]. Recently, level set-based segmentation methods are introduced in image segmentation [3], [4], [17], [20]. The idea of the level set methods is as follows: For a given image u0, we denote the desired contours of edges by Γ. When a level set function φ : Ω IR [18] is incorporated with a segmentation method, the contours of .
MySQL Quick Start Guide This guide will help you: Add a MySQL database to your account. Find your database. Add additional users. Use the MySQL command-line tools through ssh. Install phpMyAdmin. Connect to your database using PHP. You’ll also find links to further information that will help you make the most of your database. Customer Support MySQL Quick Start Guide Page 1 Contents .
