Object-Based Level Set Model for Building Detection

Object-Based Level Set Model for Building Detectionin Urban AreaYun YangCollege of Geology Engineering and GeomaticsChang'an UniversityXi’an, Chinayangyunbox@Ying LinCollege of Computer Science and Technology Harbin Engineering UniversityHarbin, Chinalinyingbox@Abstract—This paper studies an new approach to creating a variational level set model for buildings detection by combining LiDAR point clouds and Aerial image data. The level set model introduces an object-based im age analysis technique. Firstly, a fundam ental object-based level set fram ework is built by neighbor analysis of remote sensing image. Then, several derived products directly or indirectly from LiDAR raw point cloud data, like nDSM and absolute roughness data, are used to construct a novel energy term in relation to height and roughness of non-terrain objects, in order to make up the disadvantages caused by insufficient information only from remote sensing image. Thus, a closely combined model for buildings extraction has formed. The m odel can well fuse spectral feature, height and roughness infor ation of objects fro different sensors. Finally, experiments on pairs of Aerial image and LiDAR 3D point cloud data are carried out, and conclusions can be drawn that our model can effectively separate various small or high building in urban area from other land covers, including trees, grass, ground etc., and alleviate those influence caused by shadow, occlusions or spectral inhomogeneity.I.I NTRODUCTION3D city modeling has recently been a lively research areawithin the remote sensing photogrammetric community. Buildings, as the most prominent features of the urbanlandscape, receive special attention. The new developments in sensor technology allow for increased automation in building extraction and 3D model reconstruction[1], which is very useful information in telecommunication, city planning, flood control, and tourism.Airborne LiDAR data have been taken into considerationsince mid-1990s. Airborne laser scanning represents a new and independent technology for the highly automated generation of digital terrain models (DTM) and surface models[2]. The sensor data have some advantages to detect buildings from high spatial resolution image(e.g.,Quickbird, IKONOS, SPOT5 satellite imagery and Aerial imagery), compared to traditional photogrammetric method by images matching[2,3].To detect building roof candidate regions is the first step in building reconstruction. While LiDAR point cloud data can well describe 3D spatial information, and other information which high resolution imagery can’t reflect. Thus, building detection is basically a filtering problem in the DSM (raw orinterpolated) data. This is essentially a classification task, commonly solved by applying filters for computing a digitalterrain model (DTM) and applying a threshold to the height differences between the DSM and the DTM, referred to [4]. Therefore, there are studies, e.g., [4~9] exploring to only use 3D LiDAR scattered data or its various derived products for building extraction and reconstruction, and have validated the powerful advantages of LiDAR technology. However, studies also have found that the way always can not accurately and conveniently to obtain a good segmentation result of building roof because of irregularity of point cloud data. Therefore, multi-source fusion is a way to solve this problem.Although major difficulty using multispectral imagery with high spatial resolution(e.g., Quickbird, IKONOS, SPOT5 satellite imagery, and Aerial imagery) is the complexity and variability of objects and their form, especially in suburban and densely populated urban regions, imagery allow a more detailed spectral and geometrical analysis, which only LiDAR point cloud data can’t do. Regarding image-based analysis, multispectral classification uses spectral properties and spatial analysis of objects and tries to extract building. The mostly used methods are based on region-based methods, like mean shift clustering proposed by [10], object-based classification method like in [11,12], and level set methods (e.g., [13-18]) as well.In recent years, the availability of simultaneously recordheight data and color Aerial imagery have directed the attentionof many researchers towards the extensive application of LiDAR technology and the integration of Aerial images[1]. The trend has been towards the incorporation of the combined analysis of remotely sensed data and DTM into mixed environmental models[19]. The joint use of imagery and LiDAR products such as DSM, DTM, nDSM etc., can increase the total amount of information extracted from both types of data.The increasing amount studies are concerning the topics for target detection. For example, Rottensteiner et al. in [20], firstly detected building regions in raw data, then, roofs were detected using a curvature-based segmentation technique. Additional planar faces were estimated with Aerial images. Also, in [21], Dempster–Shafer method was proposed for the fusion of LiDAR data and multi-spectral images for buildingdetection. Sohn and Dowman in [22] used IKONOS images to find building regions before extracting them from LiDAR data. Haala et al. in [23] focused on building reconstruction through the combination of image and height data, uses a raster DSM, obtained through image matching. Then 3D intensity or DSM edges are compared to a building model. The data integration is mainly limited to the detection of regions, corresponding to buildings in the height data. The paper in [24] also combined RGB and color infrared Aerial image and L iDAR data for building and trees detection using four methods, i.e., “nDSM + NDVI” “Classification + nDSM” “Voids + NDVI” “L iDAR”. Currently, commercial tools (e.g. ENVI, ERDAS, TerraSolid, etc.) for building extraction and reconstruction have been developed for the task.In summary, the previous studies have made much or less contributions on L iDAR technology. Most approaches try to find the interested object using different methods from different data. However, some proposed schemes contain overfull steps to operate easily for users. In addition, high spatial resolution imagery, with rich spectral information (e.g., Quickbird, IKONOS, SPOT5 satellite imagery) with the same land cover, with similar acquisition time as LiDAR height data, always can’t be easily obtained for the scholars especially from the developing countries. Most importantly, the post-processing method L iDAR data is still not mature, and it is very necessary to explore further how to make best use of LiDAR for information extraction. In addition, Aerial imagery simultaneously acquired with LiDAR 3D point cloud data can be used for building extraction and reconstruction.In the context, the paper also concerns on fusion technique of multi-source data. Considering that the result of building detection will directly bring great influence on post-processing reconstruction process of 3D buildings, the paper mainly focus on the process of building segmentation and detection. The paper presents a variational level set method using the derived products from LiDAR raw point cloud data and Aerial images and simultaneously acquired, with the main aim to extract a variety of buildings and to improve the results.II.M ETHODOLOGYA.Basic Level Set Segmentation ModelOver the years, variational level set method based on curve evolution theory has been widely applied to image segmentation fields, owing to the advantage of free topology change and the capability of multi-source information fusion(i.e., Total energy are updated when information source are added). In remote sensing field, among various methods, e.g., [13-18], processing schemes and systems, which have been reported, level set methods also have revealed promising results. Many model-free level sets have been employed to account for the general task of satellite images segmentation, for the detection of roads, buildings and other manmade objects, in a semi-automatic or automatic level set framework.Considering its obvious advantage in image segmentation, level set method based statistical analysis in [25] is adopted as our foundation in the paper. Firstly, we define region of interest as an open subset 1 of (i.e. 1 ) where is the entire image domain. The evolving curve C is defined as theboundary of the open subset 1 (i.e. C = 1). Thus, the area inside C denotes the region 1 which is the interested object,and the area outside C denotes the region 2 which is the background. Then, our basic level set model(i.e., energy function solved by level set theory) based on region competition can be formulated as:++=++=coutsideinsidedsdIpdIpCECECECE21x))),x((log(x))),x((log()()()()(22221111(1)Where )(CEinsideand )(CEoutsiderepresent the statistical energy of inside or outside C.)(CEcis curve length constraint energy of curve C. I i(x),i=1,2 denotes the object( if i=1) or background region( if i=2) in a given image I(x). , 1 or 2are adjustable parameters to balance the power of each energyterm. )),x((iiiIp represents the probability density of I i(x). In many relevant works, e.g., in [13,15,25], the probability densityare often defined under the assumption of Gaussian distributionin image I(x) as follows:()2,1),)x()x()x(exp()),x((22i=iIIpiiiµ (2)Whereiµ,idenote means and variance of region i.However, there are complex spectral variation and rich details inside a category of land cover for remote sensing imagery. It causes the above level set segmentation model(i.e.,(1)), pure image-based method, were vulnerable tobe misled by low-level information, such as spectral or gray-level [26]. So how to overcome these drawbacks becomes themain work of this paper.B.Object-based Level Set Segmentation FramworkIn recent years, the limitations of pixel-based image analysis (i.e., single pixel is not geographically object, resultingin unavailable geometrical feature of the interested object) is more widely known in remote sensing field. So ‘beyond pixels’ image analysis strategies have been promoted to overcome those drawbacks[12,17]. Relevant image analysis methods suchas block-based methods, patch-based methods and object-based methods have been proposed. OBIA (Object-based Image Analysis) is the latest development in this field, which can divide a given image into meaningful objects through spatial, spectral or geometrical analysis. In the context, this paper introduces the idea of object-based image analysis under levelset framework in order to improve the performance of target detection. Although there have been many complex schemes about how to construct ‘object’ in previous literatures, but this paper proposes a novel construction method of ‘object’ by spatial neighborhood information analysis of image. What thispaper concerns is to suggest an effective object-based image analysis technique in the level set framework.The object-based analysis idea is inspired by the AN(adaptive neighborhood ) methods in [27], which adaptivelycaptures local properties of image to improve accuracy of image segmentation. In the paper, we adopt a scheme ofneighborhood image analysis in level set framework to build an object-based segmentation model, in order to make our modelmore robust to local variations of region-of-interest. With the idea, we can re-calculate i µ, i in (2) as follows:[]})',('|'{)()()(1)()(1)()(2)( µ µ = ==X X d andX X X X X I N X Y I N X iX Y i i i X Y ii(3)Where )(Xrepresents a set of pixels which is a neighborhood of pixel X . represents the size of neighborhood, and is set as 5 in this paper. i µ, i denotes means and variance of neighborhood )(X in the region i . Refine (2) using i µ, i calculated by (3), we can construct our basic object-based segmentation model.However, the basic model can’t perform very well, always resulting in sub-object segmentation(i.e., splitting the target region into several small segments labeled by different colors, as in Fig.1 (c)). This is caused by a limitation of neighborhood analysis and regardless of global information. To solve the problem, we should re-cluster to the each sub-region partitioned using the above basic model. Minimum distance clustering approach is adopted to again cluster small segments ignorance of initial class label of each sub-region. With this amendment, the basic object-based model(called as OLS-I) can well separate man-made objects and natural objects from Fig. 1(a) (as shown in Fig. 1(e)).C. Combining LiDAR Data for Building ExtractionThis section is to further seek the solution of image segmentation in complex urban scene. For building detection, , we often find segmentation errors caused by shadow, occlusions or spectral inhomogeneity, etc ., referred to [28], . To reduce the errors, 3D point cloud data from airborne LiDAR device are introduced into our object-based model in Section B.To overcome the disadvantages caused by shadow, occlusions or spectral inhomogeneity, which often exist for those methods purely depending on spectral information of remote sensing image, elevation and absolute roughness feature derived from LiDAR raw point cloud data are introduced into our segmentation model. Elevation information of object and background can be indirectly reflected by DSM from LiDAR 3D point cloud. While normalized digital surface model derived from the result of DSM subtracting DTM are oftenused to acquire height of objects ignorance of terrain fluctuation. On the other, absolute roughness derived from raw point cloud can well show the roughness (standard deviation)of all elevation points within each pixel[29]. Generally, trees have more strong roughness then building roof. In rasterizedabsolute roughness from ENVI LiDAR Toolkits in [30], the region of trees always is very light, while building roof regioninstead of edge always is very dark. Therefore, it is effective feature to separate tree and building, both of which are hard tobe discriminate only using height information. Therefore, we combine elevation and absolute roughness data to build a newenergy term in order to well extract the man-made targets(e.g., buildings) and also to modify segmentation error resulting fromabove basic model.(a) Arial image with automatic initialcontour (b) Segmentation result of image(a) without amendment(c) Color marking for differentregions of image (b)(d) Color marking for differentregions of image (e)(e) Segmentation result of image (a) using OLS-IFigure 1. Analysis about OLS-I modelTo acquire the derived products from LiDAR raw point cloud data, noises should firstly be removed from raw LiDAR 3D point cloud. Noises include dusts in air, clouds, birds etc. Then we can create digital surface model (DSM) by interpolated over a grid. To conveniently use elevation information of categories, especially building, tree and ground, a height filtering process of raw point cloud is carried out in order to yield DTM, so as to divide points into terrain points and non-terrain points(e.g., building, tree, grass etc .) usingiterative polynomial fitting filter, referred to [31]. Then the normalized Digital Surface Model (nDSM) is created by DSM subtracting DTM (i.e., DSM – DTM). The purpose is to exclude the influence of topography. Hence all objects in the nDSM stand on elevation height zero, that is, height of pure non-terrain objects can be obtained. A detailed description can be found in [32].With the derived nDSM and absolute roughness products, we can add a constraint energy term relative to height and roughness information of land covers.Here, Mahalanobis distance is used to build the constraint energy owing to its invariance feature to any non-singular linear transformations. Consistent with the idea of region competition, the energy term can be formulated as:dXX I M X I M dXX I M X I M C E C E C E l l Tl l ll Tl lroutside r inside r + =+=211121222111))(())(())(())(()()()(µµ µµ (4)Where )(X I lrepresents a two-band image, one is the raster nDSM, another is absolute roughness. li µ, i denotes global mean and covariance matrix of region i on )(X I l.))((X I M lrepresents a local statistic of each pixel in )(X I l , which is calculated by:=dY X Y G dY X I X Y G X I M ll )()()())(((5) Where, )( Gdenotes Gaussian window function with variance and the size of 3 3.For variational level set methods, curve C is implicitly represented by zero level set of level set function . Adding energy term (4) to (1), the level set formulation of a finalmodel(called as O LS-II) we proposed in the paper can beexpressed as follows:x|)(|x))(1())(())((x )())(())((x ))(1)(2))x (()(log(x )()2))x (()(log()(2122211122222221211111d ÖH v d ÖH X I M X I M d ÖH X I M X I M d ÖH I d ÖH I E l l T l l l l T l l+ + + ++ += µµ µµ µ µ (6) D. Numeric ImplementationThe energy function )( E can be minimized using the first variation and gradient descent method under the level set framework[26]. And a partial differential equation (i.e.,Euler-Lagrange equation) can be deduced where the update of moving curve C be conducted on the level set function in order to approximate to boundary of the interested object. Parameterize the descent direction by an artificial time t 0, the gradient flow for ),(x t can be formulated as:}))(())(())(())(()2)x ()(log(1)2)x ()(log(1)(){(2122211122222221111111l l Tl l l l T l l X I M X I M X I M X I M I p I p div v tµµ µµ µ µ + ++ + = (7)III. E XPERIMENTSA. Experimental Conditions• Parameter of models: For our model, v = 0.01 2552, t =1 i = 1(i =1,2), =0.8, 1 =2 = e (e denotes theparameter to keep energy balance between the energy of (2) and (4), and is set as 103 in this paper), =5 forall images. And comparative experiments are carried out among relevant work: TRS model(two-region segmentation model in [25]), OLS-I model and OLS-IImodel, the parameters of TRS model are set correspondent with [25] correspondently. • Experimental data: a pair of simultaneously acquired LiDAR point cloud data and Aerial imagery is tested here. The data is acquired in 2007, covering a sub-scene in Hubei province, China, using Leica ALS40. The study area has an uneven terrain. The size of Aerial image is 874 900, with 6.8 m per pixel. While LiDAR raw point cloud data have 101,933points. The averaged density of 3D point cloud isaveragely 1 point per square meter. In Fig.2(a), we display original Aerial image with automaticinitialization contour (i.e., circles with red color) to extract buildings using our proposed model). A 3-Dimension display of LiDAR raw point cloud is given in Fig. 2(b)).Moreover, two derived products ofLiDAR are listed, i.e., raster nDSM (as in Fig. 2(c)) and raster absolute roughness (as in Fig. 2(d)),.B. Result AnalysisTo verify the effectiveness of buildings extraction, comparative experiments are carried out without and with LiDAR respectively under the same initial condition (shown in Fig.2(a)). For the disturbance from shadow, occlusions or spectral inhomogeneity, TRS model can’t work well (shown in Fig. 3(a)). By a use of spatially neighborhood information, improvements can be obtained, as shown in Fig. 3(b). However, the improved scheme still can’t extract buildings satisfactorily. When nDSM and absolute roughness information are combined, the result of segmentation becomes better, as shown in Fig.3(d). The effectiveness owes to height information used to search building and tree, and absolute roughness used to differentiate trees from buildings. Note that we have extracted many low buildings which haven’t been extracted if onlynDSM data is fused, as shown in Fig. 3(b).(a) Aerial image with automaticinitial contour(b) LiDAR raw point cloud data(c) nDSM of LiDAR derived from ofLiDAR raw point cloud(d) absolute rough derived from ofLiDAR raw point cloudFigure 2. Aerial image and raw or derived LiDAR dataIntegrating spectral feature and the two derived products of LiDAR into one model, OLS-II can obtain better result(shown in Fig. 3(d)). A 3D surface of the final extracted building (i.e., Fig. 3(d)) have been demonstrated in Fig. 3(e).In the experiment, DTM greatly have influence on the accuracy of nDSM product because there are still no ideal filtering algorithm of LiDAR 3D point cloud in uneven terrain. Too often, small objects in the ground like cars etc. can not be removed completely. Thus, nDSM still can not be eliminated the influence from terrain. In addition, the geometrical correction of Aerial image is a key to improve accuracy ofbuildings segmentation.(a) Segmentation result of TRS modeusing Aerial image l(b) Segmentation result of TRS model using raster nDSM(c) Segmentation result of OLS-I model without nDSM and roughness(d) Segmentation result of OLS-II model by a joint use of Aerial image, nDSM and roughness(e) the 3D surface reconstruction from result (d)Figure 3. Comparative analysis about TRS model, OLS-I model and OLS-IImodelIV. C ONCLUSIONIn this paper, a new variational level set model is proposed to implement man-made target (buildings) extraction. Firstly, a fundamental object-based level set framework is built by using the neighbor analysis technology. Then, several derived products of LiDAR raw point cloud data like nDSM, absolute roughness are used to create a novel energy constraint term in order to make up the disadvantages of insufficient information only image-based analysis. Thus, a closely combined model to extract building roof can be obtained by a joint use of multispectral image and LiDAR point cloud data. Experiments have demonstrated that our model can separate low or high building in urban area from other land covers including trees, grass, ground etc ..Considering the fact that point density of raw LiDAR data is generally much higher at buildings or trees than at open terrain or road. Meanwhile, tree areas have low density in DTM data. So, Future work will use these features of LiDAR data to implement special man-made target extraction task by LiDAR point density analysis. Also we will perform 3D building reconstruction in urban environment.R EFERENCES[1] A. Novacheva, Building Foor R econstruction from LiDAR and AerialImages Through Plane Extraction and Colour Edge Detection, The International Archives of the Photogrammetry, R emote Sensing andSpatial Information Sciences. Vol. XXXVII. Part B6b. Beijing2008,pp.53-57[2] F.Ackermann, Airborne L aser Scanning - Present Status and FutureExpectations. ISPRS Journal of Photogrammetry and Remote Sensing,54 (2-3), 1999, pp. 64-67[3] M. Ruijin, DEM Generation and Building Detection from L idar Data,Photogrammetric Engineering & Remote sensing, July 2005, pp.847-854 [4] U. Weidner, Digital Surface Models for Building Extraction. In: A.Grünet al. (Ed.): Automatic Extraction of Man-Made Objects from Aerial andSpace Images (II), Birkhäuser, Basel, S. ,1997,193–202, [5] A. Brunn, U. Weidner, Extracting buildings from digital surface models,http://www.ipb.uni-bonn.de/publications/papers97/, 1997[6] H.G.Maas,G. Vosselman,Two algorithms for extracting building modelsfrom raw laser altimetry data ,ISPRS Journal of Photogrammetry andRemote Sensing,Vol. 54, Issues 2-3, July 1999, pp. 153-163[7] G. Priestnall, J. Jaafar, A. Duncan, Extracting urban features fromL iDAR digital surface models,Computers, Environment and UrbanSystems,Vol.24, Issue 2, pp. 65-78, 2000[8] G. Forlani, C. Nardinocchi, M. Scaioni and P. Zingaretti: Completeclassification of raw L iDAR data and 3D reconstruction of buildings.Pattern Analysis Application, 2006, 8(4): 357-374,[9] F. Tarsha-Kurdi, T. L andes, P. Grussenmeyer, Joint combination ofpoint cloud and DSM for 3D building reconstruction using airborne laserscanner data, 4th IEEE GRSS, Paris : France,2007[10] aniciu, P. Meer, Mean shift:a robust application toward featurespace analysis[J].IEEE Transactions on Pattern Analysis and MachineIntelligence,2002,24(5):603-619.[11] G..J.Hay, G. Castilla, “Object-based image analysis: strengths,Weakness, Opportunities and Threats(SWOT),” OBIA2006:International Archive of Photogrammetry, Remote Sensing and SpatialInformation Sciences, July, 2006./obia06/papers.htm.[12] T. Blaschke, S. L ang, “Object Based Image Analysis for AutomatedInformation Extraction – A Synthesis,” MAPPS/ASPRS, November,2006. / fall2006/[13] A.R. Mansouri, A. Mitiche, C. Va zquez, Multiregion competition: Alevel set extension of region competition to multiple region imagepartitioning, Computer Vision and Image Understanding, no.101,October, 2006, pp.137-150[14] O. Besbesy, Z. Belhadj, N. Boujema, “A Variational Framework forAdaptive Satellite Images Segmentation,” Springer –Verlag Berlin Heidelberg, LNCS, vol.4485, 2007, pp.675-686 [15] C. Samson, L. Blanc-Féraud, G. Aubert, J.Zerubia, “A Level Set Model for Image Classification,” International Journal of Computer Vision, vol.40, no.3, Decemeber,2000, pp.187-197,. [16] Qiu Zhen Ge, Li. Qiong, Zhang Chun Ling, et.al, Stochastic Motion and The L evel Set Method in Semi-automatic Building Detection, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008,pp.431-434 [17] G.Z. Cao, P. Hou, Y.Q. Jin, Data Fusion of Multi-source Remote sensing Based on L evel Set Method and Application to Urban RoadExtraction, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 [18] T. Keaton, J. Brokish, “A Level Set Method for The Extraction of Roads from Multispectral Imagery”, Proceedings of the Applied Imagery Pattern Recognition Workshop, 2002, pp.141–147 [19] I. V. Florinsky, Combined analysis of digital terrain models andremotely sensed data in landscape investigations, Progress in Physical Geography 22,1 ,1998, pp. 33-60 [20] F.Rottensteiner , C.Briese,. Automatic generation of building modelsfrom L idar data and the integration of aerial images. IAPRS, Vol. 34, Part 3-W13, 2003, pp. 174-180 [21] F. Rottensteiner, J. Trinder, S. Clode and K. Kubik. Using the Dempster–Shafer method for the fusion of L iDAR data and multi-spectral images for building detection. Information Fusion, 6(4), 2005,pp. 283-300[22] G.Sohn, I.Dowman, Data fusion of high-resolution satellite imagery and L iDAR data for automatic building extraction. ISPRS Journal of Photogrammetry and Remote Sensing, 62(1), 2007,pp. 43-6..[23] N.Haala, C.Brenner, Generation of 3D city models from airborne laser scanning data. Proc. EARSEL workshop on L idar remote sensing on land and sea, Tallinn, Estonia, 8 p, 1997.[24] N. Demir, D. Poli, E. Baltsavias, Extraction of Buildings and Trees Using Images and L iDAR Data,The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences.Vol. XXXVII. Part B4. Beijing 2008,pp.313-318[25] T. Brox and J. Weickert, L evel Set Based Image Segmentation with Multiple Regions, In Pattern Recognition, Springer L NCS 3175, pp.415-423, Tübingen, Germany, Aug. 2004.[26] L.A. Vese and T.F. Chan, A multiphase level set framework for image segmentation using the Mumford-Shah model, International Journal of Computer Vision 50 ,2002, 271-293. [27] S. Rital, H. Cherifi, S. Miguet, Weighted adaptive neighborhood hypergraph partitioning for Image Segmentation, in: ICAPR 2005, Lecture notes in Computer Science, vol. 3687, 2005, pp. 522-531 [28] K. Karantzalos, N. Paragios. Automatic Model-Based Building Detection From Single Panchromatic High Resolution Images, International Archives of the Photogrammetry, Remote Sensing & Spatial Information Sciences, Com. VII/3, pp.225-230, 2008. [29] LiDAR Tools for ENVI, /BCAL/tools/ EnviTools/ [30] K. M. L austen, ENVI L iDAR Toolkit, / portals/0/whitepapers/ ENVI_LiDAR_WhitePaper.pdf [31] K. Zhang, Z. Cui, Airborne LiDAR data processing and analysis tools.Miami, FL : International Hurricane Research Centre, Department of Environmental Studies, Florida International University,2007 [32] G. Vozikis, Automated Generation and Updating of Digital City Modelsusing High-Resolution L ine Scanning Systems. 20th ISPRS Congress, Istanbul. In: International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. 35,2004 [33] M. Young, The Technical Writer's Handbook . Mill Valley, CA: University Science, 1989.。