Classification of basic roof types based on VHR optical data and digital elevation model Silvia Valero1 , Jocelyn Chanussot1 & Philippe Gueguen2 1 GIPSA-Lab, Grenoble Institute of Technology, France 2 LGIT, Grenoble - LCPC, Paris - France [2] [3]. Some operators can also be derived using morphological operators [4]. Other strategies are mostly based on DEM images [5]. Finally, some methods combine optical image information and altitude data [?]. Further detailing this initial (far from trivial) step lies beyond the scope of this paper. Let us assume that the output of this step (i.e. the input of the proposed study) is the set of all the individual segmented buildings. 2) Feature extraction. The proposed technique analyzes different skeleton features using as input data the roofs extracted in the first step. To obtain the final roof-type, two different parameters are calculated. This algorithm is described in the next Section II. 3) Classification. The proposed pattern recognition is analysed in Section III. Abstract—In the frame of seismic vulnerability assessment in urban areas, it is very important to estimate the nature of the roof of every building and, in particular, to make the difference between flat roofs and gable ones. In order to perform this tedious task automatically on a large scale, remote sensing data provide a useful solution. In this study, we use simultaneously very high resolution panchromatic data, and an accurate digital elevation model. The fusion of these two modalities enables the extraction of two mixed features. Based on these features the classification between the two considered classes becomes a simple linearly separable problem. Index Terms—Data Fusion, Seismic Risk Assessment, Mathematical Morphology, Skeleton , Classification, Urban Areas I. I NTRODUCTION In the frame of seismic vulnerability assessment in urban areas, estimating the nature of every building’s roof is a key issue. In this paper, we present a feature extraction process aiming at discriminating between flat roofs and gable ones. We consider incorporating information in two ways: information from digital elevation model (DEM) and from VHR orthoimage. The two pictures are spatially registered. Using this information, we obtain a novel approach to feature extraction and classification that is described in the following. The proposed method is composed of three main steps, as described on Fig. 1: Fig. 1. General flowchart: we address the two last steps 1) Building footprint extraction (or roof segmentation). We do not develop this step in this paper. It can be performed using published methods using both the DEM and the VHR data. Some techniques are mainly based on edge and line primitives detection from an optical image [1] 978-1-4244-2808-3/08/$25.00 ©2008 IEEE II. FEATURE EXTRACTION The available data sets include a very high resolution panchromatic image (airborne data, 25cm resolution) and a digital elevation model (airborne acquisition, 1m resolution in the three directions). These two modalities provide complementary information. We need to define features that allow to separate flat roofs and gable ones. A. Feature extracted from the DEM Digital elevation model provides a description of the height of the different pixels. In our study, the height difference between the roof outline height and the ridge height is a meaningful feature. This feature is a key in our pattern recognition since higher values are expected for gable roofs: gable roofs have a ridgeline that both roof surfaces meet and corresponding pixels have a considerable height difference when compared to the border pixels. In order to calculate this height difference, we first need to extract an accurate roof outline and perform a ridge line detection. It is possible thanks to creation of one mask called ψ(x, y) presented in Fig. 2 (a). This ideal mask is the binary building footprint. It is defined as follows: 1 if x,y ∈ Roof ψ(x, y) = 0 otherwise The roof outline ψc (x, y) is trivially extracted from this mask using a standard edge detector. It is presented in Fig. 2 (b). It is applied to the DEM leading to Hc (x,y) containing κ1 pixels: IV - 149 IGARSS 2008 Hc (x, y) = DEM (x, y) × ψc (x, y) (1) Once Hc (x,y) has been determined, the estimation of average roof outline height, denoted by Γ, can be derived as follows: Γ= 1 κ1 κ1 Hc (x, y) (2) n=1 κ2 1 Hs (x, y) κ2 n=1 Υ= (4) With Υ and Γ, the difference Δ is computed: Δ=Υ−Γ (5) The Δ computed for different roofs provided the expected results: the magnitude of Δ as a matter of fact depends on the roof-type. Regarding Section III, we notice that flat roofs have lower Δ values than galbe ones. Hence, we propose Δ as one roof feature capable of classifying different roof-types using a simple threshold. B. Feature extracted from the orthoimage (a) (b) (c) (d) VHR orthoimage is a photographic map Fig. 3. In our study, this image provides a representation of a two-dimensional urban area. This kind of information is functional in order to be able to evaluate or analyze building roofs. Searching roof features, we take advantage again of the aspect of the ridgeline: it usually appears with a strong grey level contrast on the picture. This is due either to a change of illumination on the different sides of the roof (hence inducing a shadow), or to the use a metallic corner on top of the tiles. Fig. 2. (a) Binary mask ψ(x, y) (b) Binary mask ψc (x, y) (c) DEM(x,y) image (d) Binary mask ψs (x, y) In a similar way, the extraction of the average ridgeline height requires a mask to achieve a good ridgeline detection in the DEM. To perform this detection, we consider roofs as convex polygons. Consequently, the standard skeleton and the medial axis are the same in the interior of the roof. Therefore, the ridgelines of the roof can be extracted using one mask denoted ψs (x, y) which is the roof skeleton. Note that this feature does not depend on the actual shaoe of the roof. We simply assume that for gable roofs, th ecentral part will be significantly higher that the borders of the roof. This holds even in the case of non perfectly symmetrical roofs. The corresponding mask ψs (x, y) is shown in Fig. 2. The application of ψs (x, y) to the DEM image leads to Hs (x,y), with κ2 pixels: Hs (x, y) = DEM (x, y)ψs (x, y) (3) Using these data points of Hs (x,y), the average ridge height Υ can be estimated: Fig. 3. VHR orthoimage One example of the contrast existing between different roof sides is presented in Fig 4(a). This relation between the roof gradient and the perception of contrast is explained by the grey level gradient: ∂orto(x, y) ∂orto(x, y) i+ j ∂x ∂y orto(x, y) = (6) This is a vector with two components. The corresponding norm is computed as: orto(x, y) = ( ∂orto(x, y) 2 ∂orto(x, y) 2 ) +( ) (7) ∂x ∂y The result of orto(x, y) is shown in Fig 4(c). We can observe that different elements are detected besides the ridgeline. Some details are visible on the top of the roof, such as chimneys, dormers or structures to assist in the maintenance IV - 150 of the building . The presence of these elements is undesirable in our feature extraction since these elements disrupt the gradient image. Unfortunately, they are usually close to medial axis in the case of flat roofs. This problem gives way to one possible error in the pattern classification if this contrast feature is used alone. Therefore, we recommend filtering the roof image before proceeding to the contrast feature extraction. This filter is made by one standard morphological opening, whoch is defined as: ortof (x, y) = orto(x, y) ◦ se = (orto(x, y) se) ⊕ se (8) In the context of morphological transformations, the opening operation on a data set is classically defined as the succession of an erosion and a dilation. (a) Θ= κ3 1 Gs (x, y) κ3 n=1 (10) Theoretically, in the case of a gable roof, Θ should ahave a larger magnitude than for a flat roof. However, results presented in Section III do not perfectly corroborate this expected discrimination. III. CLASSIFICATION The classification step aims at separating the two classes of interest (gable roofs and flat roofs, respectively) using the two extracted features (Δ and Θ). In order to ascertain if these two features allow distinguishing roof-types, it is necessary to make several observations. In our case, we have made sixty different observations that have given rise to a couple of features-specific vectors. To interpret the vectors information, the gathered data are represented using a scatter-plot of the feature values. Thanks to this plot, we can determine graphically how to divide the plane into homogeneous regions where all objects belong to the same class. The result of our scatter-plot is shown in Fig. 5. (b) (a) (c) Fig. 5. (d) Scatter-plot Fig. 4. (a) VHR orthoimage (b) Binary mask ψc (x, y) (c) VHR orthoimage filtre (d) orto(x, y) Applying this morphological transformation, the ridgeline contrast in the image is enhanced satisfactorily as we show in Fig 4(b), which is the result of the Fig 4(a) filtered. If we make a comparison between both images, it is easy to perceive that all the small elements have been removed. In order to select ridge pixels, we use the ridge mask ψs (x, y). However, it is slightly modified using a morphological dilation (leading to the mask ψsd (x, y)) as the contrast of interest is not always very accurately located in the center of the building. Using ψsd (x, y), Gs (x, y) is created as follows: Gs (x, y) = ortof (x, y)ψsd (x, y) (9) Where Gs (x, y) is a set of κ3 pixels which provide information about ortof (x, y) on the roof skeleton. Finally, the average of Gs (x, y) is computed: (b) Fig. 6. (c) (a) Separation according to Θ (b) Separation according to Δ Two different colour markers are used in order to distinguish each roof-type. Blue crosses are used for gable roofs whereas flat roofs are represented as red crosses. Δ values are plot along the horizontal axis while the vertical axis corresponds to Θ values. Looking at fig 5 we can remark that the separation between classes is evident. However, the results obtained from the orthoimage called Θ are not as we expected. Θ values from flat roofs are lower in average than values measured on gable IV - 151 roofs. However, the standard deviation is too high compared to this difference to obtain a simple classifier based on this feature only (see Fig. 6(a)). (a) (b) (c) (d) misunderstanding in the evaluation of the real values of the roof outline edge. Despite of these limitations, the experimental results show the robustness of our classification method. One example of this achievement is the application of this method to Fig. 3. The obtained solution is shown in Fig. 8 where flat roofs are green whereas triangular roofs are represented in red. Fig. 8. Fig. 7. (a) VHR orthoimage (b) orto(x, y) (c) VHR orthoimage filtre (d) ortof (x, y) On the contrary, Δ values enable a very good separation between the two patterns. Therefore, we can establish a good differentiation. This separation is illustrated in Fig. 6(b). Thanks to the effectiveness of Δ, we decided to extrapolate the concept of this feature to make our classification more general. The main reason is the existence of several triangular roof types. Thus, our study does not have to be only focused on gable roofs but used to separate all triangular roofs from flat ones. Therefore, a parallel study of the basic DEM features provided satisfactory results on different triangular roofs such as hipped, saltbox, gambrel and pyramidal roofs. However, we found several Δ limitations. For example, one important limitation was found in hipped roofs because of the irregular geometry of their surfaces. The problem is that ridgelines are slightly out of the roof medial axis and that causes that the highest part of the roof is not situated on the skeleton. Thus, the height difference obtained from Δ is not exactly between roof outline height and ridge height. Being not exactly could not important in roofs with a hard slope on the roof surfaces. In that case, although the height difference is not the real one, it is enough performing to make a good classification. However, the slight movement in ridgelines is a real problem for hipped roofs with a low pitch. Being the height difference so small, the triangular roof could be discriminated as a flat roof. The second limitation comes from the existence of dormer windows that protrude in many triangular roofs. These common features are aligned along the roof surface, normally following the roof outline edge. They can be a problem since dormers are represented in DEM images as pixels with a high height value. In fact, those values could give rise to a Results of Classification IV. C ONCLUSION Starting from two different data such as DEM images and orthoimages, this paper has presented two skeleton features for a satisfactory recognition and classification of roof-types. The idea of two input data led us to develop a distinction between roof-types that takes into account two roof features. However, experimental results have shown that our classification can be approached by using just one feature, Δ from DEM data. This feature effectiveness achieves the purpose of this study and moreover, it makes possible a new classification replacing the pattern gable roof for all triangular roofs. Therefore, our classification can be more general and then more effective to perform a flat roofs isolation. It must also be noticed that this generalisation for triangular roofs is also limited as we mentioned in Section III. R EFERENCES [1] C. Jung and R. Schramm, “Rectangle detection based on a windowed hough transform,” XVII Brazilian Symposium on Computer Graphics and Image Processing (SIBGRAPI04), 2004. [2] W. L. Z. LIU and J. WANG, “Building extraction from high resolution imagery based on multi-scale object oriented classification and probabilistic hough transform,” Geoscience and Remote Sensing Symposium, 2005. IGARSS ’05. Proceedings. 2005 IEEE International, vol. 4, pp. 2250–2253, 2005. [3] Z. Z. Yanfeng Wei and J. Song, “Urban building extraction from highresolution satellite panchromatic image using clustering and edge detection,” Geoscience and Remote Sensing Symposium, 2004. IGARSS ’04. Proceedings. 2004 IEEE International, vol. 3, pp. 2008–2010, 2004. [4] J. Chanussot, J. A. Benediktsson, and M. Fauvel, “Classification of remote sensing images from urban areas using a fuzzy possibilistic model,” IEEE Geoscience and Remote Sensing Letters, vol. 3, no. 1, pp. 40–44, 2006. [5] M. Ortner, X. Descombes, and J. Zerubia, “Building outline extraction from digital elevation models using marked point processes,” International Journal of Computer Vision, vol. 72, no. 2, pp. 107–132, 2007. IV - 152
© Copyright 2026 Paperzz