Remote sensing imagery for damage assessment of buildings after destructive seismic events

Size: px

Start display at page:

Download "Remote sensing imagery for damage assessment of buildings after destructive seismic events"

Julie Williamson
5 years ago
Views:

1 Remote sensing imagery for damage assessment of buildings after destructive seismic events G. Bitelli 1, R. Camassi 2, L. Gusella 1 & A. Mognol 1 1 DISTART Dept., Bologna University, Italy 2 INGV, Bologna, Italy Abstract The adoption of a macroseismic scale to describe the effect of an earthquake requires the collection in a short time of a comprehensive documentation about damage on buildings and structures. In the case of large destructive earthquakes, however, the traditional field surveying methods are time consuming and costly. This is true though new methods and technologies can offer today more effective and rigorous solutions. Satellite imagery can play an important role for monitoring of earthquake effects on the builing estate, representing an extensive and quickly available information on the pre- and post-event situations. Where a cartographic base is not available, this could be the only reachable solution. Furthermore, the new generation of very high resolution imagery, with a ground pixel size on the order of one meter, may permit one to obtain a detailed description of the damage referred to each building, with very significant savings in time and costs in respect to other techniques. The paper presents the results of a research carried out on the Marmara (1999) and Boumerdes (2003) highly destructive earthquakes. Pre- and post-event images acquired by different platforms, from medium to very high resolution, panchromatic and multispectral, were evaluated and compared, and specific considerations made about their characteristics; some damage classifications are presented, where different approaches were applied. Keywords: remote sensing, VHR imagery, earthquake, macroseismic, disaster, multitemporal, change detection, classification.

2 726 Risk Analysis IV 1 Introduction Macroseismic data, describing the effect of a seismic event, play an important role both for social and scientific purposes. These data permit to examine the interaction between shaking and buildings, supplying information on engineering aspects, allowing to estimate the risk for future earthquakes, to operate for an effective damage reduction and to provide adequate resource allocation. Moreover, these data contribute to improve seismological knowledge, supplying information about propagation, dimension and orientation of seismogenetic structures, and allow calibrating historical seismological data, which constitute the main component on hazard estimates. The adoption of a macroseismic scale, i.e. the European Macroseismic Scale, requires to collect in a short time a comprehensive documentation about damages on buildings and structures; this task was performed till now by field surveying methods, but this is clearly inadequate in case of large damaged areas or when problems exist in accessing the site. Furthermore, in some situations (e.g. destructive events in developing countries) a cartographic base is not available or is largely obsolete. Satellite imagery can constitute an effective solution to obtain a map of the damaged area, either supporting the rescue operations and providing an extensive evaluation of earthquake effects on the built estate. This assessment can be reinforced by the availability of pre-event images and the application of change detection techniques. Besides, the new generation of very high resolution imagery, with a ground pixel size on the order of one meter, represents the first possibility to have a detailed description of the damage referred to each building, with very significant saving in time and costs in respect to other field-based traditional techniques. 2 Damage classification and macroseismic survey techniques After a severe earthquake the data about aftershocks are collected by local and temporary instrumental networks. Collection of macroseismic data derives instead principally from two sources: questionnaire surveys and field investigations. Questionnaire surveys are commonly used for assessing macroseismic intensities in the range of 2 to 5, while for 6 (the damage threshold) and above field investigations are necessary. The engineers are mainly interested in severe failures and damage assessment to decide if a building should be used, restored or demolished; the macroseismologist is interested instead in the overall damage distribution, including the borderline between slight damage and no damage, and the spatial patterns of variation in intensity caused by local conditions. These data can be the core of the possibility of establish a link between past and present events. Sometime, an isolated small damage to an old structure could be much more important to this aim rather that an extensive, uniform collapse of RC buildings. Distinctions between different construction types, usage, height and age and quality of construction should always be made wherever possible.

3 Risk Analysis IV 727 Strong earthquakes require extensive and immediate field investigation to record damage patterns. The revised version of the European Macroseismic Scale (EMS-98) based on the MSK scale, provided by Grünthal [8], tries to define accurately some key terms, like building type, damage grade and quantity. The EMS-98 scale needs a lot of detailed information on buildings types and vulnerability classes, on damage grades and its percentage distribution among the total number of structures interested by the earthquakes. A modern macroseismic survey should then include techniques that may give a more precise scientific meaning to the observation carried out, and ideally to do this they should provide the possibility of being repeatable and investigate neighbouring conditions that may have led to a certain damage level. The first proposition is the hardest to satisfy: few days (or even hours) after an earthquake the state of buildings is forever changed by restoration or rescue operations. The simple photographic or movie survey is giving only a limited view of damage, extrapolating a single case from the context and thus providing a subjective interpretation by the surveyor. The Quick Time Virtual Reality (QTVR) technique, based on photographic images stitched together by software providing 360 panoramas subsequently linked in complex scenes, was then recently adopted by the authors: it has the advantage of providing not-interpreted data, allowing each user to perform a virtual navigation inside a damaged neighbourhood (Mucciarelli et al [10]). Nevertheless, immediate direct intervention on the damaged areas are often not possible, and other information unavailable. Scope of this work is to investigate the possibility to undertake some damage evaluation from satellite imagery, even without ground data, in support of macroseismic classification. 3 High resolution satellite imagery after earthquakes Modern Very High Resolution (VHR) sensors aboard remote sensing platforms can provide imagery with a ground resolution up to 70cm, involving in this way the possibility to obtain precise information about objects like buildings, infrastructures, trees, etc. For these reasons, interpretation and classification of VHR imagery could provide useful detailed data about ground situation and intervened modifications, especially for urban scenarios. In the case of natural disasters, like catastrophic earthquakes, VHR imagery could become essential, particularly when a cartographic base of affected areas is absent or obsolete (this is a typical situation for developing countries), or when damaged zones are hardly accessible (in this case the images could also be helpful to identify the quickest way to reach them and go through). Moreover, VHR images are very important in the emergency phase because rescue teams can use them as base maps in handheld GPS-GIS systems. They could be very useful in a pre-event phase in order to update existing databases related to exposure and vulnerability to earthquake of buildings and lifelines, and their possibility to reduce both the cost and the number of field surveys could be of great interest also for the insurance and reinsurance industry.

728 Risk Analysis IV In the following paragraphs different classification approaches are described; they were applied also for medium-high resolution imagery.

2, 152.4 5 m 1D IRS 09.27.1999 (b) 47.0, 167.0 5 m 1D BOUMERDES IRS 08.12.2002 63.5, 141.0 5 m 1D IRS 06.08.2003 (c) 69.8, 129.3 5 m 1D QuickBird 4.22.2002 61.4, 144.2 11.2,176 0.70 m 2A QuickBird 6.

syscorr. (a) (b) (c) (d) Figure 1: Post-event imagery used for the study: from left to right, the increase in resolution for different satellite products (Marmara and Boumerdes earthquakes).

4, about 17,100 victims and 25,000 injured people), and the Boumerdes earthquake, Algeria, occurred in 2003, May 21 (Magnitude Richter 6.8, about 2,300 victims and 11,000 injured people).

4 728 Risk Analysis IV In the following paragraphs different classification approaches are described; they were applied also for medium-high resolution imagery. Table 1: Specifications of the scenes acquired. MARMARA ETM TM sample in fig.1 sun elevation, azimut off nadir, target azimuth resolution product level 30 m (a) 30 m IRS , m 1D IRS (b) 47.0, m 1D BOUMERDES IRS , m 1D IRS (c) 69.8, m 1D QuickBird , , m 2A QuickBird (d) 67.2, , m 2A syscorr. syscorr. (a) (b) (c) (d) Figure 1: Post-event imagery used for the study: from left to right, the increase in resolution for different satellite products (Marmara and Boumerdes earthquakes). The case studies are related to two recent highly destructive earthquakes occurred in the Mediterranean area: the Marmara earthquake, Turkey, occurred in 1999, 17 August (Magnitude Richter 7.4, about 17,100 victims and 25,000 injured people), and the Boumerdes earthquake, Algeria, occurred in 2003, May 21 (Magnitude Richter 6.8, about 2,300 victims and 11,000 injured people). Table 1 shows the images used in the tests till now carried out. The visual effect of the different image resolutions is shown in fig. 1. The study conducted by the authors in relation to Marmara earthquake is presented in Bitelli et al [2] [3], the present paper is mainly referred to the Boumerdes area, where VHR imagery is available. 4 Macroseismic image classification In EMS98 specifics, the object of interest is the single building, which has to be classified in relation to typology and grade of damage: building typologies and material (masonry, reinforced concrete, steel ) are quite impossible to extract

However, a macroseismic classification of VHR imagery suffers some limitation, due to image resolution, not adequate to achieve a classification of soft damage (< 3 on the EMS98 scale) and due to the

5 Risk Analysis IV 729 from VHR images, and also grade of damage is not clearly visible as by a ground survey. The use of satellite imagery brings nevertheless some advantages, in terms of rapidity in saving resources and human life. However, a macroseismic classification of VHR imagery suffers some limitation, due to image resolution, not adequate to achieve a classification of soft damage (< 3 on the EMS98 scale) and due to the point of view: for a damage assessment by satellite or aerial imagery, an off-nadir view is preferable, since is possible to see building façades, but satellite imagery usually is almost nadiral and, if the post-event image is intentionally oblique, considerable problems arise for automatic change detection algorithms. Furthermore a change detection algorithm have to deal not only with earthquake related damage but also with differences on shadow, seasonal changes, presence of new housing units, etc. In Table 2, macroseismic damage assessment for the city of Boumerdes is reported, showing also the ratio between panchromatic bands and the sum of absolute difference in each multispectral band (eqns (1) and (2)). Ratio = Mean Jun ) / Mean( Apr ) (1) AD = i= R, G, B, IR ( PAN PAN abs( Mean( Juni ) Mean( Apr i )) (2) It is also to notice that the images used for the test aren t orthorectified and have the standard 2A geocoding accuracy; no cartographic support is available for this study. The correspondence between object shapes on two images was carried out by visual interpretation and an assisted matching procedure between features: this punctual information of displacement is also used for the automatic change detection as input for post-event registration process, producing a virtually precise registered image, fig. 2. Figure 2: Objects are manually registered between pre- and post-event images: a partial collapse of the structure is clearly visible both observing shadows and comparing object shape in the two images. Table 2 shows the summary of visual image interpretation of the damage carried out on the two images; areas characterized by new constructions and by

6 730 Risk Analysis IV the presence of debris are also considered as changes. The average values reported for the indexes are referred to the eqns (1) and (2). Table 2: Image classification and damage indexes. Damage Class Buildings Ratio AD Class Class Class Class Built new Debris Automatic image classification The first phase before applying an image classification algorithm is to perform an accurate image co-registration (as outlined, orthorectification with absolute georeferencing was not possible in this case due to lack of ground control data). For the Boumerdes Earthquake there are problems in registering VHR images, because of buildings geometry and the presence of shadows caused by different acquisition parameters and geometry. A rubber sheeting model, implemented in Erdas Imagine, was applied using 116 CPs (Control Points) and 2512 CPs, obtained from a manual classification of Boumerdes urban area. In the first case the results weren t so satisfactory, instead in the second case the shift in buildings position between the registered and the master image was reduced to minimum. It s obvious that this is a long work and is generally not applicable, especially when the aim is a quick post-event damage assessment. 5.1 Pixel based classification After registering the images, the work focuses on the identification of damaged areas. Erdas 8.6 Knowledge Engineer module, by Leica, was used to conduct a pixel based classification of pre- and post-event QuickBird pan images in order to extract a class representative of damaged buildings. This classifier uses hypotheses, rules and variables to create classes of interest, and makes a rulebased classification through their implementation in a decision tree. The target is to produce a semi-automatic procedure, applicable not only in the cases studied: for this reason it s necessary to define, where possible, standard values for variables. Areas featuring damaged buildings present high values of brightness due to the spread of debris after the building collapse (Estrada et al [7]). Through a user-specified threshold on the brightness change, the model permits to identify large damaged areas, but lots of bright zones not damaged are also included, which can be called false alarms. In order to improve the result and remove these false alarms the model was implemented in the decision tree and combined with other variables. These come from ISODATA unsupervised classification of pre-event image (extraction of the urban areas from QuickBird multispectral registered images), the evaluation of NDVI index, and finally the calculation of

7 Risk Analysis IV 731 the distances between the bright areas and the zones classified as shadows in the post-event image; it must be however noticed that not all the false alarms are removed. 5.2 Object oriented classification The object oriented classification was performed using ecognition 3.0, by edefiniens, using the QuickBird images above described and IRS images. In this case, the scope of an object oriented classification is to simulate human perception of image features, using a segmentation procedure (Baatz & Shape [1]) to recognize the objects, applying classification rules to describe the perception of changes. In ecognition, it is necessary to perform the segmentation before the classification processes: segments extracted represent objects of interest. The post-event image is thus processed. The particularity of ecognition is the possibility to perform a multiscale segmentation and hierarchy; it is then possible to use a higher scale value to separate big-scale objects, such as urban areas, and a lower scale value to separate the objects of interest (in this case damaged buildings and debris). The hierarchy is therefore composed by at least two levels, the texture (level 3, scale ) and the urban built (level 2, scale 15-30). In the urban level, it is necessary to define how an object is perceived as changed. Human eye is able to recognize reflectance changes without considering shadows; contrariwise, using image differencing techniques, an increasing in reflectance could happens when a building fall across a shadow in the other image. Thus, a sublevel of urban level (level 1) is created, to recognize reflectance only on not saturated and not shadow zones. The final level hierarchy structure is resumed in fig Classification evaluation The accuracy assessment of automatic change detection for VHR imagery suffers from problems derived by geometric issues and methods of interpretation of the results, and it is quite difficult to consider separately each question. Geometric problems arise from image registration, image resolution and off-nadir effects. Evaluation of results is related to how the percentage of hit areas is considered, and how to take into account false alarms. The percentage can be calculated in terms of damaged edified area detected in relation to the total built area, or in terms of buildings exactly detected as damaged, as outlined in fig. 4. Accuracy evaluation shows quite similar results in object-oriented classification (o.o.c., software ecognition) and in pixel-based classification (p.b.c., software ERDAS Imagine). From registered to precise registered image, damaged buildings detection (grade 4-5) increases from 58% to 74% using o.o.c. and from 60% to 82% using p.b.c. It is also to notice that p.b.c. is less accurate in extracting undamaged objects (75% for o.o.c., 55% for p.b.c.) and in general shows less stable results without a strict co-registration. With IRS imagery, results are quite similar as using registered QuickBird imagery, indicatively 51%

of buildings correctly classified as damaged (respectively 83% and 77% of buildings correctly classified as no damaged).

9 IRS Scale: 1 5 30 Color: 1 0.7 0.7 Smoothness 0.9 0.

detected (c), then the image is classified (d) using a damage index. Class hierarchy and segmentation parameter are reported.

$In terms of false alarms, calculated as the fraction of the damaged detected area that doesn t fall in a change area$

8 732 Risk Analysis IV for o.o.c. and 65% for p.b.c. of buildings correctly classified as damaged (respectively 83% and 77% of buildings correctly classified as no damaged). (a) (b) (c) (d) Data Set Level 1 Level 2 Level 3 Object hierarchy QuickBird Scale: Color: Smoothness IRS Scale: Color: Smoothness Figure 3: For each object, in the pre-event (a) and post-event (b) image, portions to exclude as shadow and as saturated are detected (c), then the image is classified (d) using a damage index. Class hierarchy and segmentation parameter are reported. Figure 4: Accuracy assessment explanation. On the left: object oriented classification, on the right: pixel based classification. In terms of false alarms, calculated as the fraction of the damaged detected area that doesn t fall in a change area corresponding to buildings, using registered image the percentage is 77% for p.b.c. and 69% for o.o.c.; this percentage is reduced using precise registered imagery (respectively 56% and 54%). This widespread phenomenon is primarily due to the large areas covered by debris after a strong earthquake and its entity could be clearly reduced by different procedures for building classification or by the availability of a large

Risk Analysis IV 733 scale vector cartographic database. In fig.

area. (a) (b) (c) (d) (e) (f) (g) (h) Figure 5: (a) pre-event

Results from classification Erdas with (c) 116 GCP, (d) 2512 GCP,

assessment on urban areas have been introduced and the results

damage and suffers from geometric image co-registration problems,

9 Risk Analysis IV 733 scale vector cartographic database. In fig. 5, results of classification are visually presented for a sample area. (a) (b) (c) (d) (e) (f) (g) (h) Figure 5: (a) pre-event image, (b) post-event image. Results from classification Erdas with (c) 116 GCP, (d) 2512 GCP, (e) IRS imagery; ecognition with (f) 116 GCP, (g) 2512 GCP, (h) IRS imagery. 6 Conclusion Some problems related to the macroseismic damage assessment on urban areas have been introduced and the results provided by remote sensing techniques, in high and very high resolution, presented. The use of these data is in general not much sensitive to soft damage and suffers from geometric image co-registration problems, permitting an acceptable quantitative assessment of damage for grades higher than 3. The overall results are promising and deserve further investigation. The object-oriented classification technique, in particular, shows

10 734 Risk Analysis IV in this case interesting capabilities (e.g. behaviour regarding co-registration problems) in comparison with the more diffused pixel-based methods, even though they provide comparable results. References [1] Baatz, M. & A. Schape., Multiresolution Segmentation an optimization approach for high quality multi-scale image segmentation, in: Strobl, J. et al. (Hrsg.): Angewandte Geographische Information-sverarbeitung XII. Beiträge zum AGIT-Symposium, Salzburg, Karlsruhe, H. Wichmann Verlag, pp.12 23, [2] Bitelli, G., Camassi, R., Gusella, L., Mognol, A., Applicazione del telerilevamento nello studio del danneggiamento sismico in aree urbane: il caso di Izmit (Turchia). Atti VII Conf. ASITA, Verona, pp , [3] Bitelli, G., Gusella, L., Mognol, A., Using an object oriented approach for earthquake macroseismic damage classification, 3rd Int. ecognition User Meeting, Munich, [4] Bitelli, G., Camassi, R., Gusella, L., Mognol, A., Image change detection on urban area: the earthquake case. ISPRS Archives, Istanbul, [5] CEOS, The Use of Earth Observation Satellite for hazard support: assessment and scenarios ( [6] Eguchi, R.T., Houshmand, B., Huyck, C.K., Mansouri, B., Matsuoka, M., Shinozuka, M., Yamazaki, F., Ulgen, S., Using Advanced Technologies to Conduct Earthquake Reconnaissance After the 1999 Marmara Earthquake. 2nd Workshop on Adv. Techn. in Urban Earthquake Disaster Mitigation, DPRI, Kyoto University, Uji, Kyoto, [7] Estrada, M., Matsuoka, M., Yamazaki, F., Damage Detection from the Space. EDM Techn. Rep. No. 6, Report on the Kocaeli, Turkey Earthquake of August 17, 1999, 4, pp , Earthquake Disaster Mitigation Research Center, RIKEN Institute, Miki, Hyogo Prefecture, Japan. [8] Grünthal, G. (ed.), 1998, European Macroseismic Scale 1998 EMS-98, European Seismological Commission, Subcomm. on Eng. Seismology, Working Group Macroseismic Scales, Helfent-Bertrange, Luxembourg. [9] Gusella, L., Registration of multitemporal Landsat images: a local correlation approach, Reports on Geodesy, Vol.2(65), pp , [10] Mucciarelli, M., Camassi, R., Gallipoli, R., Collection of macroseismic data in a digital age: lessons from the 1999 Kocaeli, Turkey earthquake. Seism. Res. Lett., 73, 3, pp , [11] Weston, A-J.S., Burton, P.W., Lovett, A.A., Imagery and Ground Truth in an Assessment of the Built Environment within an Earthquake Impact Scenario: Colchester, UK. Workshop on the Appl. of Rem. Sens. Techn. for Disaster Response, Univ. of California, Irvine.

Remote sensing satellite imagery and risk management: image based information extraction

Risk Analysis VI 149 Remote sensing satellite imagery and risk management: image based information extraction G. Bitelli & L. Gusella DISTART Department, University of Bologna, Italy Abstract The introduction