Building and Vegetation Heights 2012

Methodology

Procedure

The procedure for determining the building and vegetational height involves a high-effort process which can be described only roughly here. For detailed information, especially with regard to the details of particular procedural steps, which nonetheless have a considerable effect on the quality of the results, please refer to the Project Report, the technical literature, e.g. Bayer et al. (in publication), or the appropriate websites, e.g. that of the DLR.

Figure 4 shows the basic course of the procedure, which is then described in the sections below.

The main work steps are shown in various colours: segmentation steps in orange, classification steps in blue, and the export of the results in grey.

Fig. 4 Work procedure for the ascertainment of buildings and vegetation items in the Berlin urban area

Data processing (DSM/TOM generation)

The Digital Surface Model (DSM) and the TrueOrthoMosaic (TOM), which is generated on the basis of the former, are processed on the basis of stereo photography. If the orientation parameters of the GPS, the Inertial Navigation System (INS) and the Satellite Positioning Service of the German Ordnance Survey System (SAPOS) are available, the single overlapping aerial photographs are linked via so-called tie points, compiled into a composite image, and simultaneously transferred to the required coordinate system (Kraus 2004).

The DSM provides coded elevation points of the earth’s surface, including all items located upon it (buildings, roads, vegetation growth, etc.). This is done by adding the terrain elevation to the height of the item. Since the terrain elevation in Berlin, while not showing particularly great differences of topography – the range is between 35 m and 100 m above sea level – is nonetheless not constant, no conclusions regarding the respective absolute item height can be drawn at this time. For this, a Normalized Digital Surface Model (NDSM), in which the terrain is everywhere set to a standard of zero, would be required. The NDSM is accordingly generated by subtracting the Digital Terrain Model (DTM) from the digital surface model (DSM):

  • NDSM = DSM – DGM

With respect to the ensuing segmentation and classification, this simplifies the distinction between elevated and non-elevated objects, and permits direct measurement of object heights. It permits the distinction between streets, elevated vegetation, and structures, and provides precise height information (cf. Fig. 5).

Fig. 5: The principle for the generation of a Normalized Digital Surface Model (NDSM)

The basic material provided by the aerial photography flight prior to data processing is a composite aerial image with a central perspective depiction. This type of projection results in an image of the object which is geometrically distorted toward the edges. Hence, no uniform scale is present in the image, and no realistic distances can be calculated. In order for the image data to be useful in the geo-information system (GIS), a differential rectification must be carried out with the aid of a DSM. This procedure results in a TrueOrthoMosaic (TOM). With the aid of TrueOrtho image generation, it is then possible to determine the geometrically exact position of pixel precision, since each pixel only has one position within the DSM. The result is that there is no longer any tilting, which, e.g., might permit building façades to be visible; at the same time, it is only this correction of the image data of the measurement that makes real distances possible (Heipke 2003, cf. Figure 7). The existence of a DSM and its TOM is accordingly the basic precondition for precise analysis in densely built-up urban areas, and considerably increases the reliability of statements.

Fig. 6: A comparison of the DOP and the TOM: a) DSP 20RGB 2011, with tilted building images, b) TOM RGB, with ALK building contours and building edges superimposed and corrected true to the position, c) NDSM with ALK building contours superimposed

Fig. 6: A comparison of the DOP and the TOM: a) DSP 20RGB 2011, with tilted building images, b) TOM RGB, with ALK building contours and building edges superimposed and corrected true to the position, c) NDSM with ALK building contours superimposed

The further procedure described below refers, inasmuch as values are stated, to the consistently qualitatively better input material from Phase 1. Hence, not all corrective steps, e.g. for the ascertainment of vegetation in shaded areas, can be carried out in the areas of Phase 2.

The existing TOM database consists of four channels: red, green, blue (true colours RGB) and near-infrared (NIR) in a geometric resolution in x/y of 15 cm. Similar to the NDSM, spectral data are reduced to a resolution of 30 cm. This primarily serves to reduce the data quantity, but also takes into account the spectral geometrical heterogeneity of the item being ascertained (Trosset et al. 2009). Due to the fact that the aerial photography for both phases was conducted relatively late in the year (cf. flight parameters), the photographic products show a large amount of shading. Thanks the high radiometric resolution of more than 12 bits, and the simultaneously generated DSM, many objects in the shaded areas can nonetheless be correctly identified; however, this is true only for Project Phase 1. Particularly by means of the NIR channel, in which the vegetation heights show spectral values, it is possible to reliably ascertain the vegetation share (cf. Figure 7).

Fig. 7: Quality of the TrueOrthoMosaic (TOM) data produced: Use of the near-infrared (NIR) channel (below); the bright areas indicate high spectral values and good ascertainability of the vegetation shares

Fig. 7: Quality of the TrueOrthoMosaic (TOM) data produced: Use of the near-infrared (NIR) channel (below); the bright areas indicate high spectral values and good ascertainability of the vegetation shares

With the increasing availability of Very High Spatial Resolution (VSHR) remote sensing data, the suitable methods for processing them and the yield of information from these data also increases. Particularly in densely built-up urban areas characterized by a high degree of heterogeneity of structures, this task has proven to be very complex. Even though it has been the VHRS data which have made possible the satisfactory analysis of urban spaces, the high degree of resolution and the resulting greater degree of heterogeneity of the data has generated additional problems. Given the large quantity of data, it is therefore necessary to use suitable methods for the large-scale processing of data.

Analysis and segmentation

In remote sensing, two different approaches of image analysis are used: the object based approach and the pixel-based approach. In the object based analysis used here, single items are analysed in their respective contexts. As a result, not only the spectral properties, but also the object forms, textures and especially the neighbourhood relationships – all of that information which cannot be derived from a single pixel – are taken into account. In this way, the object based procedure attempts to model itself upon the human method of perception. Unlike with pixel-based analysis, it is not single image elements, but the more expressive homogenous segments consisting of a number of pixels which form the basis of classification (Blaschke 2000).

The data driven procedure used in this project thus segments the entire image scene with the aid of certain statistical procedures and defined parameters. The resulting segments are pixel clusters which do not yet have any particular semantical significance. Only the ensuing classification assigns these items to those classes with the class description of which they most closely correspond. The quality of segmentation is determinant for the following classification. For this reason, the software used also contains various segmentation algorithms which can be applied.

The segmentation method known as Multi-Resolution Segmentation (MRS) is of exceptional importance, especially for the generation of the vegetation masks. Here, adjacent pixels which fulfil certain homogeneity criteria are compiled into ever larger segments. This is carried out as long as a certain threshold value of homogeneity is not exceeded. The more precisely the desired item or item segment reflects its real shape, the easier it is to carry out the ensuing classification (Baatz & Schäpe 2000).

Figure 8 shows this segmentation procedure via a number of object hierarchy stages, in which smaller items (e.g. parts of buildings) are classed at the lower levels and the larger ones (ALK buildings) at the upper levels. By means of this procedure, the item segments “inherit” the properties (e.g. item keys) of the respective “mother” items, e.g., of the ALK buildings.

Fig. 8: Abstract depiction of the object hierarchy and several scale levels (left), and with image examples (right)

Object-based classification

Based on the existing digital aerial photography images and the NDSM, it was possible to define certain item classes with the aid of ALK building data. Object based classification is a multi-stage procedure. The particular steps are processed hierarchically along the so-called process tree (cf. Figure 9), and then stored; they are thus available for use in later procedures (e.g. in the context of monitoring by means of change analyses).

Fig. 9: Process tree developed for object-based classification (depiction of a segment)

Fig. 9: Process tree developed for object-based classification (depiction of a segment)

The classification is structured for the following areas of evaluation emphasis:

First, two main classes are distinguished (Buildings and Vegetation). In addition, other sub-classes are defined for the vegetation and for the buildings, respectively, in the form of height levels; moreover, the buildings are also broken down into further semantical classes (e.g. garages, sheds, garden houses and “Buildings planned or under construction”).

Ascertainment of the buildings

In the area of buildings, various item classes are distinguished, which are also to be kept separate in the further procedure, and in the data storage in the GIS:

  • ALK-Level Buildings, in the categories:
    • Existing
    • Planned/under construction
    • Special ascertainment of the following use types: garages, sheds, garden and weekend cottages
    • Special category: “Roof area under trees”
  • ALK-Level Topography (bridges, above-ground railway areas)
  • Above-ground structures which are not part of the ALK, in the categories
    • Sheds/garden cottages under the ALK building key system “Allotment gardens or cottage colonies”
    • Buildings
  • Greened roofs, distinguished by location as:
    • ALK buildings
    • Non-ALK buildings.

In a generalized segmentation process, the higher buildings are separated from the lower ones on the basis of certain homogeneity criteria. Next, a more detailed segmentation of the vegetation from the anthropogenic items is carried out. For this purpose, adjacent pixels which fulfil certain homogeneity criteria are compiled into ever greater segments. This is continued until a pre-established homogeneity threshold value is no longer exceeded. Distributed across the entire scene, all segments grow simultaneously, which ensures that they are of equal size, and hence comparable (Baatz & Schäpe 2000).

The classification of the measured buildings is carried out with the aid of the ALK building layer, so that, e.g., even large roof eves ascertained in the aerial photography data can be corrected for. In addition, bridge-type items and above-ground railway areas of various types – especially railway stations and elevated railway tracks – can be ascertained as part of the ALK topography.

Low buildings such as sheds, garages or cottages which can often not be masked out under the interpolation rules of the surface model, can subsequently be ascertained in the ensuing step on the basis of the respective key from the ALK item key catalogue. The later assignment of heights is in this case carried out by reference to type of use (number of stories = 1 × 2.8 m).

The ALK level “Planned buildings” is also a separate item class which includes surface or subsurface buildings planned or under construction. For these items – some 1400 in the city, according to the ALK as of June 2010 – no conclusive heights can be determined. There are likely candidates for later change analyses (cf. Figure 10). However, it should be noted in this respect that structural changes or new structures are far more numerous than those documented in the ALK data.

Fig. 10: Two classification examples for buildings of the item class “Buildings planned or under construction” (as of June 2012); left and middle: certain classified segments, and, right, as compiled items; in the top row is an example of the construction project in progress; below it is a building already completed at the time of the aerial photography flight in September 2010

Fig. 10: Two classification examples for buildings of the item class “Buildings planned or under construction” (as of June 2012); left and middle: certain classified segments, and, right, as compiled items; in the top row is an example of the construction project in progress; below it is a building already completed at the time of the aerial photography flight in September 2010

All building items which are not part of the ALK structures but which are nonetheless structural entities, have been ascertained just as comprehensively. The example of the item class “Sheds/garden cottages” in allotment garden and cottage colonies illustrates the detailed precision of the procedure with which all built-up areas are to be largely classified without error, even prior to the generation of the vegetation mass. In the case of low garden cottages, the shape characteristics have been used to delimit them from road and pathway surfaces (cf. Figure 11).

Fig. 11: Classification steps for the determination of the item classes “Garden cottages” outside the ALK system: a) The classification of the vegetation mask attained prior to this step, with real colour RGB aerial photography data superimposed; b) Situation-precise section of the surface model NDSM – no heights of garden cottages can be directly derived; c) Initial classification attempt yields errors in the area of streets; d) Ensuing classification step, with correction for the street areas

Fig. 11: Classification steps for the determination of the item classes “Garden cottages” outside the ALK system: a) The classification of the vegetation mask attained prior to this step, with real colour RGB aerial photography data superimposed; b) Situation-precise section of the surface model NDSM – no heights of garden cottages can be directly derived; c) Initial classification attempt yields errors in the area of streets; d) Ensuing classification step, with correction for the street areas

For buildings listed in the ALK, building heights are conclusively presented on the basis of the ALK building outlines and their storey lines. For this purpose, the ALK building polygons are first of all separated into ALK building portions by means of the storey lines. The heights of the building segments are then aggregated proportionately to these ALK building portions (Fig. 12).

Thus, the mean heights of ALK buildings are represented on the basis of ALK geometries. Buildings not listed in the ALK are therefore still presented by means of building segments.

Fig. 12: Aggregation of heights of building segments to the ALK building portions; left: building segments; right: ALK building portions

Fig. 12: Aggregation of heights of building segments to the ALK building portions; left: building segments; right: ALK building portions

Ascertainment of vegetation

After the classification of buildings, the classification of the vegetation mask is carried out. By means of the preceding multi-resolution segmentation (MRS), the existing items are very finely structured, so that the differentiated diversity of the existing vegetation can be shown. The usual method of vegetation extraction used in digital remote sensing is based on a vegetation index calculated by means of the channels red (R) and infrared (NIR) ascertained by aerial photography. This Normalized Difference Vegetation Index (NDVI) is derived as follows:

  • NDVI = (nIR – R) / (nIR + R).

The index uses the fact that vital vegetation yields particularly high values in the near-infrared range, and much lower values in the red spectral range, a constellation not found in any other class of items, which makes possible a simple separation between vegetation and other classes (Albertz 2001).

In order to ascertain the existing vegetation structures and heights, a Contrast Split Segmentation (CSS) process is carried out. In this way, the vegetation mask can be subdivided into lower and higher vegetation segments. The resulting detailed results provide a very suitable basis for the ensuing calculations with regard to further differentiation of vegetation height data. For this purpose, the Multi-Threshold Segmentation (MTS) method is used. This procedure uses the pixel values of the NDOM, and subdivides the existing segments according to the stipulated nine height levels, comparable to the drafting of a contour line map (cf. Figure 13).

Fig. 13: Results of Multi-Threshold Segmentation (MTS) – Subdivision of the vegetation into nine height levels

In order to address the differentiated diversity of vegetation on a still more detailed scale, the existing nine vegetation height layers are further subdivided by taking their height structures into account. With the aid of the already described MRS, it is thus possible to ascertain which different vegetation structures exist (cf. Figure 14).

Fig. 14: Results of the subdivision into vegetation structures with the aid of multi-resolution segmentation (b & c) within particular height levels, from the preceding multi-threshold segmentation (a, cf. Figure 13)

Greened roof surfaces

In connection with the measures for the adaptation to climate change or efforts for sustainable concepts for rainwater use [German only], the ecological value of greened roofs has increasingly been discussed. While Berlin has a long tradition of greening roof surfaces, there has to date been no full scale mapping of these types of roofs.

In the initial ascertainment carried out in the context of this project, no differentiation has been made with regard to varying intensities of roof greening, such as intensively used roof gardens or exclusively moss-covered roof surfaces (cf. Figure 15). The correlation of the vegetation mask with the building layer based on the ALK was to serve the purpose of categorization of surfaces as greened roofs, but it provided no conclusive result. Here, no account was taken of adjacent vegetation which might partially cover the roof and hence distort the result somewhat.

Fig. 15: Greened roof surfaces, as a result of the correlation of the vegetation mask with the ALK building layer

Fig. 15: Greened roof surfaces, as a result of the correlation of the vegetation mask with the ALK building layer

In order to classify the vegetation items which cover roof segments, but are not themselves greened roof surfaces, the class of greened roofs was examined in NDOM and in NDVI for its differences from the adjacent items. If the items in the class “Greened roofs” showed only a slight difference in height with respect to adjacent tree items, and a greater difference in height with respect to the buildings, as well as very similar spectral properties to the trees, they were iteratively shifted to the class “Roof segments covered by vegetation”. In Figure 16, the necessity of this formation of classes can be clearly seen with reference to the attached NDOM image. If they were to remain classified as buildings, they would lead to a distortion of the figures for average building heights.

Fig. 16: Classification results of the class “Roof segments covered by vegetation” (conclusive representation below right)

Fig. 16: Classification results of the class “Roof segments covered by vegetation” (conclusive representation below right)

The classification of greened roofs also yielded several further errors which could be largely eliminated via spectral analysis of adjacent segments. The analyses carried out in the context of this project are insufficient to provide a very precise detection of greened roof surfaces with regard to their delimitation, and possibly also there vegetational stock; however, they do provide a largely complete overview of the generally greened roof surfaces. The initial ascertainment in the inner-city area shows that in the 445 sq km area processed, there were approx. 10,000 roof surface areas (a figure which does not equal the number of buildings) which have been greened to various degrees of intensity. Due to the poor resolution of the data obtained from the outlying areas, the information on greened roofs obtained in Project Phase 2 was not of satisfactory precision. For this reason, greened roofs in the outlying areas have not been marked.

Overall assessment of the quality of ascertainment

The assessment of the quality of the results requires differentiation between the two project phases. Overall, the resulting buildings and vegetation items in the inner-city area have been ascertained satisfactorily, and in some cases, particularly for buildings in the ALK inventory, the ascertainment has been very good. However, they have also shown certain weaknesses, which will be examined more closely in the following, and which are considerably more evident in the results from Project Phase 2, due to the less precise input data, which could only be corrected to a certain degree.

Exactness for buildings

h5. Project Phase 1

The methods used involve exclusively computer-based and automated procedures of item extraction, which has a major effect on the quality of the results. The completeness and correlation of ascertained buildings can always be assumed to be very good if it can be supported by the ALK inventory of buildings.

The precision of ascertainment of buildings which are not part of the ALK can also be evaluated as very good, with some exceptions. These involve temporary objects such as circus tents or large construction containers, the ascertainment of which can only be corrected by manual post-processing.

Further subdivision of buildings according to their roof segments, which is based on the respective roof geometry, can often be implemented to a degree which is only satisfactory. The sensitivity of the MRS method achieved to date, using the spectral channels, is insufficient for an exact determination of problematic areas involving shading. The problems of these limits of ascertainment procedures can be clearly seen in the context of the ascertainment of greened roofs in the shadows of overhanging trees and vegetation objects with visible shadows of buildings (cf. precision of vegetation heights).

The greened roofs have also been satisfactorily determined with respect to the targeted initial ascertainment. The completeness of their extraction has however on occasion been distorted by shading. With the aid of an initial relatively precise investigation of 20 selected greened roofs, a high degree of correspondence with regard to correctness and completeness of the results can be demonstrated. Approx. one third of the correctly detected surfaces show a completeness of the ascertained surface of below 50 %, while half show a completeness of more than 75 %. Taking into account the various factors which can interfere with this classification, this result can be considered good.

In addition to the realization of an ascertainment of the surface extent of items which is as close as possible to reality, the quality of determination of item types is of particular importance. With regard to the “average value of ascertained heights” indicated in the factual data representation, it should be noted that such average values can in some cases in which several areas of the building have very different heights and expanses, lead to arithmetic combined values which deviate considerably from the maximum height of one part of the building. In the context of the item determination of buildings, therefore, maximum heights are also ascertained which are, however, of less importance for the major part of the building, e.g., if they involve only small smokestack areas, etc. On the other hand, the Victory Column is given with an average height of only approx. 20 m, although its maximum height is almost 70 m (cf. Figure 17). In other cases, such as, e.g., ordinary residential buildings, this type of calculation is, however, advantageous, since the maximum heights here are much less significant than the calculated average heights.

Fig. 17: Building height ascertainment, using the example of the Victory Column: Here, the calculated average height of approx. 20 m deviates considerably from the maximum height of the item, which is 67 m

h5. Project Phase 2

Although the quality of the ascertained building items has indeed been noticeably improved by means of later correction procedures, it is still conspicuously different from that of Phase 1. As it was not possible to carry out manual corrections, certain buildings, particularly not in the ALK inventory, have inevitably been detected which do not really exist. Moreover, particularly with lower objects, heights have tended to be underestimated. For this reason, the information from the “Maximum Number of Full Storeys as per ALK” have been integrated into the data display in order to provide additional information (in the context of this investigation, heights of 2.8 m per full storey have been assumed).

Moreover, it is recommended, especially in the case of small-scale analyses in the FIS Broker presentation, to use the possibility of overlaying current aerial photography.

Exactness for vegetation items

h5. Project Phase 1

The class Vegetation includes a total of nine height levels. These are also exported together with the height attributes, so that vegetation, too, can be assigned to its own system of height differentiation. This categorization into height levels and structures it is very precise and true to nature, and the phenomenology of the segments is homogeneous. Only in the interpolated areas of the digital terrain model are there cases of misclassification of vegetation, which is, however, due to the inadequate input data in these areas, and not to the determination methodology. In these cases, too, imprecision can only be corrected by manual follow-up correction.

h5. Project Phase 2

The quality restrictions in the interpolation area of the terrain model that also apply to Phase 1 are, depending on the location, even more apparent in the outlying areas, due to the less precise input data of 50 cm ground resolution, as opposed to 10 cm in the inner city area. The greatest distortions of altitude occur in orographically moving terrain; they have been largely corrected by means of manual altitude point setting, so that they ultimately ascertained height values for vegetation largely conform to reality. However, reduced output quality has a negative effect on the segmentation of vegetational objects, which can be seen in the large number of ‘error points’, where no statement on the vegetational stock can be made.

Figure 18 shows this very clearly in the transitional area between Phases 1 and 2, in an area in the western portion of the Grunewald Forest.

Fig. 18: Segmentation gaps in Project Phase 2, due to reduced quality of the image data material (using the example of Grunewald Forest)

Fig. 18: Segmentation gaps in Project Phase 2, due to reduced quality of the image data material (using the example of Grunewald Forest)

Data display

The data display in the FIS Broker includes the following information on the selected buildings and building segments, and the respective vegetational areas:

  • Buildings
    • Building ID
    • Building key as per ALK
    • Classification
    • Berlin item key catalogue (buildings use code)
    • Use as per ALK item key catalogue as of June 30, 2012)
    • Maximum number of full storeys, as per ALK
    • Ascertained average height, in meters
    • Greened Roof [yes/no] (only in the Phase 1 area; see Fig. 2)
    • Building is located in the block/block segment with the key…
    • Borough
    • Area size (sq m)
  • Vegetation items
    • Item ID
    • Item key
    • Area size, in sq m
    • Ascertained average height, in meters
    • Vegetation is located in the block/block segment with the key…

Moreover, it is still useful to call up the current aerial photography images as ortho-photos or colour infrared (CIR) versions, via the corresponding button in the FIS Broker map presentation, as a background map for the selected map segment.